when i typed start.bat beartype.roar.BeartypeDecorHintPep484Exception: Function torch.onnx.symbolic_helper._onnx_unsupported() return PEP 484 type hint "typing.NoReturn" invalid in this type hint context (i.e., "typing.NoReturn" valid only as non-nested return annotation). Press any key to continue . . how
Sorry you all have this issue which seems to be a newly introduced bug but hopefully this is the solution: Run this command inside your venv: "pip install beartype==0.15.0" Step by Step: 1.Use python browser 2. Activate the environment 3. Run "pip install beartype==0.15.0" 4. Deactivate the environment Find more info on it here: git.ecker.tech/mrq/ai-voice-cloning/issues/388 If the fix doesn't work for you, I'm sure the author of the tool will patch this soon (once he does that, just install Tortoise again or run "git pull" via command window in the folder where to tool is installed).
I'm glad for this tutorial. At first I got some error problems, but when I tried it again it worked out really fine! Thank you very much for this tutorial, this is going to help me so much. It's so sad to see that you only have 600 subscribers, you deserve more. But very happy to see no dislikes on this video! 👍
Ohh my god, took very long time to do it but after all, it was worth it, the problem is many people don't have that much patience to go through with all, but thank you very much your effort is appreciated.❤❤
Could not find module 'C:\ai-voice-cloning\venv\Lib\site-packages\torchaudio\lib\libtorchaudio.pyd' (or one of its dependencies). Try using the full path with constructor syntax. got this error even if the files exist. after the step (start.bat) please help
Wow! I could not tell at first that this voice was AI generate. Your tutorial also has very step by step detailed process, liked and subscribed. Could you please share the voice data that you used in this video so that we can use it without needing to retrain? I have previously trained some voices but they were not as good as yours.
Meta has launched its own TTS model named voicebox,(not realising it fearing its consequences).But they have a github repository named meta tts . You should make a video on this
Thank you for letting me know! I heard about what Meta announced but for me it wasn't that interesting because they won't release the code for us to try out. I hope we'll soon have even better ways to generate TTS locally
@@AiVOICETUTOR Your tutorial is the only one I've been able to follow to successfully install all of this! Thank you so much! I am getting a lot of errors using it though. I tried to train a 1 minute model and it was saying Index Error User [Training] [2023-10-14T00:53:15.571531] warnings.warn("Detected call of `lr_scheduler.step()` before `optimizer.step()`. " [Training] [2023-10-14T00:53:17.514094] C:\AI\ai-voice-cloning\venv\lib\site-packages\torch\utils\checkpoint.py:429: UserWarning: torch.utils.checkpoint: please pass in use_reentrant=True or use_reentrant=False explicitly. The default value of use_reentrant will be updated to be False in the future. To maintain current behavior, pass use_reentrant=True. It is recommended that you use use_reentrant=False. Refer to docs for more details on the differences between the two variants. [Training] [2023-10-14T00:53:17.514094] warnings.warn( [Training] [2023-10-14T00:55:06.284260] Disabled distributed training. [Training] [2023-10-14T00:55:06.284260] Path already exists. Rename it to [./training\JohnKushite\finetune_archived_231014-005259] [Training] [2023-10-14T00:55:06.284260] Loading from ./models/tortoise/dvae.pth [Training] [2023-10-14T00:55:06.284260] Traceback (most recent call last): [Training] [2023-10-14T00:55:06.284260] File "C:\AI\ai-voice-cloning\src\train.py", line 64, in [Training] [2023-10-14T00:55:06.285268] train(config_path, args.launcher) [Training] [2023-10-14T00:55:06.285268] File "C:\AI\ai-voice-cloning\src\train.py", line 31, in train [Training] [2023-10-14T00:55:06.285268] trainer.do_training() [Training] [2023-10-14T00:55:06.285268] File "C:\AI\ai-voice-cloning\modules\dlas\dlas\train.py", line 408, in do_training [Training] [2023-10-14T00:55:06.285268] metric = self.do_step(train_data) [Training] [2023-10-14T00:55:06.285268] File "C:\AI\ai-voice-cloning\modules\dlas\dlas\train.py", line 271, in do_step [Training] [2023-10-14T00:55:06.286264] gradient_norms_dict = self.model.optimize_parameters( [Training] [2023-10-14T00:55:06.286264] File "C:\AI\ai-voice-cloning\modules\dlas\dlas\trainer\ExtensibleTrainer.py", line 321, in optimize_parameters [Training] [2023-10-14T00:55:06.286264] ns = step.do_forward_backward( [Training] [2023-10-14T00:55:06.286264] File "C:\AI\ai-voice-cloning\modules\dlas\dlas\trainer\steps.py", line 242, in do_forward_backward [Training] [2023-10-14T00:55:06.286264] local_state[k] = v[grad_accum_step] [Training] [2023-10-14T00:55:06.287262] IndexError: list index out of range. And then when trying to generate different things I'd also get an error saying Text too long even though it was likei 4 or 5 words. o____O. I managed to successfully train when I first ran the programme but I don't think I had enough EPOCHS as it sounded like a British man instead of an American Voice that I cloned. It sounded like the voice but without the accent which was strange.
Pretty cool and a good tutorial. I did have an issue the the install bombing the first time so I deleted the install and started over and it went well without red error messages. Tested with the default voice and it worked. Now on to training some new voices. Currently in the middle of training with Sylvester Stallone's voice as a learning process.
Awesome! Glad you didn't give up the first time and that you managed to get it to work. Thanks for sharing that and fingers crossed that Sly sounds good!
I was looking for this video for a long time. Thanks for the effort. Great video, will do a re-install today. I want to know if there is a place to download more voices to use. Without training everything by myself. Thanks in advance.
No problem. I’m glad you found the video! I haven’t seen any downloads for voice models but the wiki has links to voice samples that you’ll need to train yourself: git.ecker.tech/mrq/ai-voice-cloning/wiki/Collecting-Samples. For the RVC method that requires another voice as input rather than text (where I also made a video about), there’s a discord with voice models to download: discord.com/invite/aihub
Hi, thank you very much for your thorough tutorial. But I encountered this when installing TorToiSe TTS: "DEPRECATION: voicefixer 0.1.2 has a non-standard dependency specifier streamlit>=1.12.0pyyaml. pip 23.3 will enforce this behaviour change." . Any idea to solve it?
I recently trained a model for RVC to clone a singer (I personally know) and it worked quite well. I've seen people use TTS produced audio and put it through RVC to get a better more realistic voice. Now my question: TTS can be trained in many languages, is RVC language specific or could I train a voice in TTS, let it speak a text and then send it through RVC, when I have trained a model there (with the same training material) ? Have you tried to combine the 2 techniques?
I haven't tried to combine those techniques but I read that it works well (at least in English). AFAIK the issue with TortoiseTTS and other languages is that you need a lot of GPU power and time to train a base model for a different language. So far I have not seen any of them appear online. RVC voice cloning works in any language as long as the input voice and the voice to clone are the same language
Tortoise TTS cloned voices don't sound very accurate on their own. Which is why i come with this recommendation: After generating voices with tortoise, give them a pass with RVC. they'll be pretty much on the level of eleven labs v1 afterwards. it's awesome! the main issue with this is that RVC also requires a long training session for it to do its thing. but no biggie, as you can use the same samples you used to train tortoise. just leave it training overnight and GG
Wow thats very interesting! Thanks for sharing this but could you let me know if I understood it correctly? You mean that you use the output .wav files from tortoise and then run them through RVC GUI? That means you trained a model with the same voice input samples in Tortoise AND in RVC beta?
Am I able to start this while waiting on a newer GPU? My current 3080 ti is out of commission in a build I'm waiting to finish, so I am running on my old 1070. Which apparently NSight Compute no longer supports. Am I screwed until then?
how do you restart the URL for the user interface? I shut down my computer and know the URL does not work. do i need to start the Command prompt all the time and what is the command I need to use to start the URL again?
After installing (and closing) the tool you can always open it again by running "start.bat" in the "ai voice cloning" folder. Then open the URL it shows you (127.0.0.1:7860) in a web browser.
Hi. Thanks for the tutorial. It really worked well. After training with around 8 mins audio and your settings (and baking my PC for ~45 mins ^^ ), I got a pleasing result. As the vid is now 1 month old I wonder though, if you have updated your workflow or new suggestions. Although the training setting was slightly changed by the "validate configuration" process (some value from 8 to 7, does it change things you cannot change yourself inside the training settings?) because of VRAM. But then it only used ~12GB in the process (I ve got 24GB). I've got to dive more into the program, but I wonder how much better RVC might be, as you've said yourself? This is already good. Does the quality of the training data have a huge impact on the output quality or is not big, because of the diffusion and just the abstract characteristics of the voice are the base of what's then put into the new model? So many questions and lota more hehe, Thanks again for the well done tutorial!
Hi, thank you very much for your feedback and the kind words! Glad to hear it worked well for you. AFAIK this is still the best free/local method for AI text-to-speech. Although its not perfect, I was happy with the results I got so I never looked into tweaking the workflow. There's probably some room for improvement. RVC is (supposed to be) used to clone a human voice which therefore IMO overall sounds less artificial than the TTS does sometimes. As far as I can tell, the quality of the input voice does matter, though the tool sometimes seems to be very forgiving if that makes any sense. BTW someone in the comments suggested training the same voice in Tortoise AND RVC. Then after generating the .wav file with TTS, run that audio through RVC GUI for better results.
@@AiVOICETUTOR yeah, the only strange thing though, i cant seem to achieve the same quality from the moment I ran the test voice line now, even using a short sentence only and the same seed. I dont know why this could be.
solution: in the settings "autoregressive model" i switched back to autoregressive (it was on my trained model 501 gpt.pth). then the computerish-undertone seemed to have dissapeared, but still not sure, maybe its also to do with other things, have to test more..
[Training] [2024-06-15T19:20:29.805467] local_state[k] = v[grad_accum_step] [Training] [2024-06-15T19:20:29.805467] IndexError: list index out of range i dont know if its an issue, but after this the training stops, i can use the voice but its extremely deep and robotic: any idea on what to do? thanks in advance!
This is really cool! I got it to work for a minute but then my computer crashed. I am wondering, do i have to redo everything again to get it to work everytime i shut my computer down or if it crashes?
Did it crash during training or afterwards? If it happens during the training then I would restart the training from scratch. If it happened after the training, you're fine and can select the models to start generating the voice with your text prompts (as shown in the video)
Hey just to let you know that I tried to run three different TTS voice cloning tools that are available through Pinokio. Sadly two of them didn't work for me (Bark Voice Cloning and XTTS) and the other one (VALL-E-X) sounds quite bad compared to TortoiseTTS. So I won't be making a new video but if you want to, check them out yourself and see if they work for you. More info on installing Pinokio here: ruclips.net/video/ln1qEglnpMo/видео.html
I have an issue. It seems that I ain't got the hardware it requires (I lack a non-integrated graphics card). WOuld it still be possible to make it work? I'm not particularly savvy with computers
Got my voice cloned and it works like charm..btw any idea how can I get this voice I just made work with my assistant? I need to pass the text from the assistant somehow to the tts and make it generate the audio files.. I have already figured out how to play the generated files.
You can use the tool via its API. For more info see here: github.com/neonbjb/tortoise-tts#api I''m not sure if the API calls work with the version from my tutorial or if you need to install the one from that git. Hope that helps
The tool supports any language but you need to teach it the language (check the wiki for more on that git.ecker.tech/mrq/ai-voice-cloning/wiki/Training). From what I understand, sooner or later there will be public models for other languages so that you won’t need to train the language (and can use the same process as in this video).
Hello @AiVOICETUTOR Thanks for this clean and detailed video on TorToiSe TTS! I enjoyed the whole process. I was almost done but during the end of "08:53 Train Voice model in TorToiSe TTS" my system got restarted. Since I am a non technical guy and new to this AI tools, I don't know how to start from where it stopped while restarting. How do I get the "TorToiSe TTS" user interface? Do I need to start from the beginning? Please help me.. Thanks in advance!!
Hi, I'm glad you like the video. You don't have to start from scratch. Just open the "start.bat" file through windows explorer and once it's loaded, open the URL that is shown in the command window. Sadly I couldn't find any information on resuming the training of the model. So worst case is that you need to do all the steps regarding the training again. Hope this helps and good luck!
@@AiVOICETUTOR Thanks for the quick response. When I tried to generate it gives the following error: "Something went wrong CUDA out of memory. Tried to allocate 12.00 MiB (GPU 0; 4.00 GiB total capacity; 6.47 GiB already allocated; 0 bytes free; 6.60 GiB reserved in total by PyTorch) If reserved memory is >> allocated memory try setting max_split_size_mb to avoid fragmentation. See documentation for Memory Management and PYTORCH_CUDA_ALLOC_CONF"
Hi I get this error when clicking "Transcribe and Process" during training of clone voice "RuntimeError: cannot reshape tensor of 0 elements into shape [0, -1, 256, 256] because the unspecified dimension size -1 can be any value and is ambiguous #391"
Make sure you don't have any emotions or empty lines at the end of your prompts. More here: git.ecker.tech/mrq/ai-voice-cloning/issues/158 Hope that helps
Sorry having to disappoint you but my AI legal consultant advised me against doing so, as long as the legality of cloning and using other voices is still uncertain. However, I can tell you that I trained the female voice the exact same way as the male voice in the video. And I found the voice by searching for 20+ minute lectures and speeches on RUclips and it didn't take me too long to find her.
The tool supports any language but you need to teach it the language (check the wiki for more on that). From what I understand, sooner or later there will be public models for other languages so that you won’t need to train the language (and can use the same process as in this video). Those models for additional languages take very very long to train and require powerful hardware so it might take a while.
Thank you for the explanations. I'm curious to try this out but I'm tempted to ask if this can clone a Nigerian voice somewhat accurately. I'm asking because the google colab tool that I previously tried with this AI tend to come out in American accent.
Yes you'd need another model file that is especially made for the Nigerian language. The one that comes with the tool is made for English only, so all other languages trained with it will have an accent. I hope it won't be much longer before we can use more languages other than English.
The transcription of my audio samples had some mistakes, but once I corrected the text in the 'whisper.json' and recreated the dataset, the tokens weren't updated to reflect the corrected text in the 'whisper.json'. It seems my only options are to transcribe again, but then it just makes the same mistakes. Any idea how to fix this while keeping my manually edited text?
I couldn't find any info about editing the transcript on the official git. It must be doing some more stuff in the background so you can't just edit the .json? My best advice would be to remove the voice bits that don't transcribe properly from the input data.
Hi, yes this should work on a Mac if you have a M1 or M2 chip. There's some more info on the official git: github.com/neonbjb/tortoise-tts#apple-silicon
i have created more folders under the "voices" folder, but when i click refresh voices, i only get the first folder i created, anyone else with the same problem?
This might be the solution to your issue: "Add pydantic==1.10.11 to the end of the requirements.txt file, re-run setup-cuda.bat then run start.bat. Everything will be ok." More info here: git.ecker.tech/mrq/ai-voice-cloning/issues/300. Hope it helps.
I trained the female voice the exact same way as the male voice in the video. And also generated it the same way. I varied between the ultra fast and fast presets without changing any of the other default settings. Hope that helps
@@AiVOICETUTOR Awesome. Thank you. Can you talk about the settings of batch size and gradient size? The validate training configuration does sometime not work for me as it starts to give error. It is recommended to reduce batch size to resolve this issue, but don't know optimal batch size. I have 4070Ti
The batch size is a tradeoff between speed and accuracy. Smaller values will be slower but more accurate while higher values will be faster and less accurate. Maybe try lowering it to 32 or 64. I wouldn’t change the gradient accumulation size if it works for you with the lower batch size but here’s what Bard says: Gradient accumulation size is a parameter that controls how many batches of data are processed before the model's parameters are updated. A larger gradient accumulation size means that more data will be processed before the parameters are updated. Gradient accumulation can be used to train on larger batch sizes than your machine would normally be able to fit into memory. This is because the model's parameters are only updated after a certain number of batches have been processed. The optimal gradient accumulation size will depend on the specific model and dataset being used. A good starting point is to use a value of 1. If the model is not converging, you can try increasing the gradient accumulation size. If the model is becoming unstable, you can try decreasing the gradient accumulation size.
I followed your tutorial step by step, yet it doesn't seem to work for me. Once I go into the command prompt, all I get is error messages. I managed to get the Cuda setup bat to run by running it through a double click. Trying so in the command prompt always gives an error. After it was done, I tried start.bat, which says that I should enter a parameter. I couldn't find anything relating to that, so, just to try, I entered a path to a folder, which was then opened in a new window. Only by starting start.bat, which simply refers to activate.bat in another folder, as an administrator, could I get it to not show an error, but it basically is stuck on that ever since, not doing any downloads or the like. So, while this is a good video, it looks like this is a bust for me.
Failed here at the Training Model, in the console it says no NVIDIA driver found albeit I installed it (got AMD Radeon GPU), there is any workaround for it? Lets say CPU only?
@@AiVOICETUTOR yea I think it's the only next move to make. But once I turn off my computer I'll have to redo everything again in the CMD hey? There's exe installed that I can run that will make it work is there? Because I ran into an issue the first time with my computer crashing. When I went back to the web interface it no longer worked because the commandline had been shut down and I wasn't sure if there was a short cut to access what had already been installed through this.
Once installed, you can launch the webUI by running „start_http.bat“ which should open your web browser and you can go straight to the voice prompts. You won’t have to do anything in a command window again
I managed to reach the "start.bat" step but it give me this error: No module named 'psutil'. Also it states more things on line 11 in main.py and line 21 in utils.py How can I solve?
Does not work for me. I followed every step. When I execute start.bat the following error message will appear: (venv) C:\ai-voice-cloning>start.bat (venv) C:\ai-voice-cloning>call .\venv\Scripts\activate.bat Traceback (most recent call last): File "C:\ai-voice-cloning\src\main.py", line 11, in from utils import * File "C:\ai-voice-cloning\src\utils.py", line 40, in from tortoise.api import TextToSpeech as TorToise_TTS, MODELS, get_model_path, pad_or_truncate ModuleNotFoundError: No module named 'tortoise.api'
@@AiVOICETUTOR I kind of get the same issue, and tried resolving it with the thread. Firstly I get this: (venv) C:\ai-voice-cloning>call .\venv\Scripts\activate.bat Traceback (most recent call last): File "C:\ai-voice-cloning\src\main.py", line 11, in from utils import * File "C:\ai-voice-cloning\src\utils.py", line 40, in from tortoise.api import TextToSpeech as TorToise_TTS, MODELS, get_model_path, pad_or_truncate ModuleNotFoundError: No module named 'tortoise' When I change it to "from api import TextToSpeech as TorToise_TTS, MODELS, get_model_path, pad_or_truncate" I instead get this: (venv) C:\ai-voice-cloning>call .\venv\Scripts\activate.bat Traceback (most recent call last): File "C:\ai-voice-cloning\src\main.py", line 11, in from utils import * File "C:\ai-voice-cloning\src\utils.py", line 40, in from api import TextToSpeech as TorToise_TTS, MODELS, get_model_path, pad_or_truncate ModuleNotFoundError: No module named 'api'
when i typed start(.)bat it says FileNotFoundError: Could not find module 'D:\ai-voice-cloning\venv\Lib\site-packages\torchaudio\lib\libtorchaudio.pyd' (or one of its dependencies). Try using the full path with constructor syntax. any solution? EDIT: you have to use python v 3.9 or 3.10, the newest version doesnt work
also before that i got this Does .\venv\Lib\site-packages\bitsandbytes\cuda_setup\. specify a file name or directory name on the target (F = file, D = directory)?
Could not find module 'C:\ai-voice-cloning\venv\Lib\site-packages\torchaudio\lib\libtorchaudio.pyd' (or one of its dependencies). Try using the full path with constructor syntax. got this error even if the files exist. after the step (start.bat) please help
I get this error when transcribing: "cannot reshape tensor of 0 elements into shape [-1, 0] because the unspecified dimension size -1 can be any value and is ambiguous"
Make sure you don't have any emotions or empty lines at the end of your prompts. More here: git.ecker.tech/mrq/ai-voice-cloning/issues/158 Hope that helps
The tool supports any language but you need to teach it the language (check the wiki for more on that). From what I understand, sooner or later there will be public models for other languages so that you won’t need to train the language (and can use the same process as in this video).
Thank you for the awesome tutorial. It's weird though, I followed everything until it came to the training. Since my source audio isn't 10min, but a little shorter, I thought I'll set the epochs to 400. Everything else was the same. I clicked train and got this error: 'utf-8' codec can't decode byte 0x81 in position 2: invalid start byte Then I changed it back to 500 epochs, tried again and it worked. lol Either way it's working now if I don't change that number.
Hmm that UTF-8 thing is an odd bug. I'm glad you got it to work and thanks for sharing your solution! Maybe check if you have any special characters in your folder path or file names
I ran in to this when trying to train. [Training] [2023-11-04T18:17:17.305090] from axial_positional_embedding import AxialPositionalEmbedding [Training] [2023-11-04T18:17:17.308976] ModuleNotFoundError: No module named 'axial_positional_embedding' is there someway to fix it ?
Hey I got a question. I tried the TTS and it works really fine, but it takes decades to generate one sentence. Is there a way to speed up the generating process?
No, you can read here why it's not that easy: github.com/neonbjb/tortoise-tts/issues/567. I made a video about how to clone your own voice (which works in any language): ruclips.net/video/5i_Pyw0gH-M/видео.html&lc=UgxvH3x_DKWkw8gLvkZ4AaABAg
sorry i just have to ask. so if i do want to reopen the application how do i do it? and if i want to download the generated audio how can i do this too?
After installing (and closing) the tool you can always open it again by running "start.bat" in the "ai voice cloning" folder. Then open the URL it shows you (127.0.0.1:7860) in a web browser. The generated audio can be downloaded by clicking the arrow next to the audio file player in the Web UI
solution please .........(venv) D:\ai-voice-cloning>call .\venv\Scripts\activate.bat Whisper detected Error: No module named 'vall_e' Error: No module named 'bark'
do you know any way to make TTS speaks like original input audio? I want to make AI dubbing on my movie and can't figure it to speak like original language audio file :(
As far as I know, this is still the best free TTS voice method. It's not perfect yet but it's only gonna get better from here on. Hopefully there will be new models out soon that will improve the quality
Question about the part where you train your model. My auto regressive model only seems to go up to 201. I use it but I don't get good results. The higher that number the more accurate to the training data? Is it possible to dedicate more GPU resources to train it better? I have a 4090 currently for this. If I keep hitting "train" over and over again, each time it finishes I can get a more and more accurate model? Sorry i'm new to this. edit: after I get done training it, I hit train again, because I assume that will make the model better, but nothing happens, no output or anything.
Makes sense that you can't just train it again but thanks for confirming it. If you set the 500 epochs and you only get a model saying 201 then the training must have stopped too soon (the highest regressive number should be the same as the epochs you set). Do you know if the training progress said it's doing 500 epochs? And maybe check disk space too. BTW Too many epochs can make it sound worse so its a good idea to try all the regressive models
@@AiVOICETUTOR interesting, i'll try that and keep more of an eye on it. Disk space might be an issue i'll check that as well. I'll keep this thread updated and try the different regressive models. Another question I have, lets say one regressive model sounds better but still not quite there, and the one i'm currently using and that one are the ONLY regressive models I have, what are my other choices? For reference, the voice i'm training I have about an hour of clean, clear audio and each clip of that hour is spliced into 5 second chunks.
Not sure if that's what you meant but the "Save Frequency (In Epochs)" determines how many models you'll end up with. So if you lower that number you'll end up with more models to try. The tool definitely works better with some voices while other voice don't seem to work well (yet).
during training the process isnt progressing for quite a while and the cmd says "RuntimeError: FIND was unable to find an engine to execute this computation" is it normal or the process is actually failed?
Hello so I have been following all the steps and even checking to make sure everything is installed. However, when trying to run the command setup-cuda.bat I encounter an error that says "setup-cuda.bat is not recognized as an internal or external command operable program or batch file." I even double checked the Cuda files were installed. Any ideas what I should do?
Hello this sounds like a general issue with how you run the bat file. Did you try to double click or open the .bat file (like any executable) or are you trying to run it from a command window?
Someone please help. I followed everything in the video exactly as it's shown but when I get to training my voice when I click train after adding the folder and a good chunk of wav files for it to train from I get an error stating [WinError 2] The system cannot find the file specified and it stops training.
@@AiVOICETUTOR I don't have any network of cloud folders everything is stored on my computer's C: drive. The file path is C:\ai-voice-cloning\training\voicetest/train.yaml I even did an entire fresh install cloning the github again and repeating all the steps and I have the same issue
When viewed in File Explorer the file path is C:\ai-voice-cloning\training\voicetest/train.yaml But when I click the drop down for training configuration it auto detects ./training/voicetest/train.yaml There shouldn't be any special characters or spaces and everything is stored locally on my PC with full access to file editting reading and writing. Really no idea what's going on :(
Your path looks fine. I guess your audio file names are fine too. Unless someone else has any idea, I'd suggest to post your issue to the developer at git.ecker.tech/mrq/ai-voice-cloning/issues
Hello help me error (When start Training get an Error: (result, consumed) = self._buffer_decode(data, self.errors, final) UnicodeDecodeError: 'utf-8' codec can't decode byte 0x84 in position 0: invalid start byte) If you use the learning rate scheduler (calling scheduler.step()) before the optimizer’s update (calling optimizer.step()), this will skip the first value of the learning rate schedule. If you are unable to reproduce results after upgrading to PyTorch 1.1.0, please check if you are calling scheduler.step() at the wrong time.
Sadly no matter what number for the echos I have. (I started with 500) I always get the error "utf-8...". I've changed the Echos number to 400, 499, 600. Nothing helps.
So I have a new issue. A reinstall fixed my earlier issues. I trained my voice and set the autoregressive model but even after refreshing voice lists I can't select it, it's just not there? Checked the CMD window and it appears to load the autoregressive model fine but no voice option
That is very strange if you didn't move any files around. Even though it's unlikely to be the cause, could you check and make sure that you have the tool installed on a local drive (not network) and in a folder that doesn't contain any spaces or special characters
@@AiVOICETUTOR I was an idiot. I hadn't put my audio files in a new folder so they were just sat in 'Voices' lol My new issue is that the generation always gives the voice a very posh sounding, southern English accent. There's definitely elements of the voice I cloned there but it just sounds weird.
Thanks for sharing the solution to your issue! For your new issue, I suggest you make sure your input voice data is clean (so that no other voices or sounds are in it). You could also try increasing or decreasing the length of the input voice or increase or decrease the number of epochs. Also I have noticed that it just doesn't work well with some voices sadly (especially if they have a strong accent)
@@AiVOICETUTOR thanks, I'll try that. I also discovered a TTS to RVC pipeline which generates the voice in Tortoise from a python file then runs it through a trained RVC model. I've set it up and have left a voice training in RVC. Looking forward to trying it. Also, my earlier issue (IndexError: List out of Range). It turns out that when setting up the training config you need to make sure that the batch size is evenly divisible by the gradient accumulator otherwise you get that error. Changing mine to 96 worked.
Cool! Again thanks for sharing the fix and the Tortoise to RVC pipeline. It’s something I’ve been wanting to try for a while myself now as it should give near perfect TTS output, right? Hope it works well for you and would love to hear what you think about the quality of the output.
Sorry I missed your reply. Not sure if you've know about Pinokio (ruclips.net/video/ln1qEglnpMo/видео.html). Sadly it doesn't support TortoiseTTS yet but I hope it will be added soon.
Sadly the .pth files trained in RVC are not compatible with this tool. Check out this colab but I think it doesn’t have a proper UI: colab.research.google.com/drive/1wVVqUPqwiDBUVeWWOUNglpGhU3hg_cbR
Hi, I'm getting this error when I type start.bat: Traceback (most recent call last): File "C:\TTS\ai-voice-cloning\src\main.py", line 11, in from utils import * File "C:\TTS\ai-voice-cloning\src\utils.py", line 29, in import torchaudio ModuleNotFoundError: No module named 'torchaudio' Press any key to continue . . .
@@AiVOICETUTOR Thank you. How do I start a new installation. I initially installed Python 3.11 but I've uninstalled it and installed Python 3.10 just like your tutorial. What do I delete or uninstall to start a fresh installation.
When I put in start.bat, I get this error FileNotFoundError: Could not find module 'E:\TorToiSe TTS\ai-voice-cloning\venv\Lib\site-packages\torchaudio\lib\libtorchaudio.pyd' (or one of its dependencies). Try using the full path with constructor syntax. I have no idea what to do from here.
I deleted everything and tried again, but now I'm getting this when I put in start.bat Traceback (most recent call last): File ".\src\main.py", line 11, in from utils import * File "E:\TorToiSe TTS\ai-voice-cloning\src\utils.py", line 28, in import torch File "C:\Users\purpl\AppData\Local\Programs\Python\Python38\lib\site-packages\torch\__init__.py", line 451, in raise ImportError(textwrap.dedent(''' ImportError: Failed to load PyTorch C extensions: It appears that PyTorch has loaded the `torch/_C` folder of the PyTorch repository rather than the C extensions which are expected in the `torch._C` namespace. This can occur when using the `install` workflow. e.g. $ python setup.py install && python -c "import torch" This error can generally be solved using the `develop` workflow $ python setup.py develop && python -c "import torch" # This should succeed or by running Python from a different directory.
Sorry it's not working for you. Unfortunately I can't find anything about your issue. You could try it again in a few days, assuming the dev will fix it or you could post the error you're getting at git.ecker.tech/mrq/ai-voice-cloning/issues
Can you make sure that the tool is actually using your GPU and not the CPU? You can check task manager while the training is running. The tool is using CUDA and the GPU I'm using has 16,384 cores while the 2060 has 1920 cores so it should be slower but i'm not sure if it should be that slow.
Are you getting the error when trying to train the voice or when trying to generate the voice? Lowering the batch size for training shouldn't affect the quality. Maybe set it to 32 or 64 and give it another try
WHEN I RAN IT DID THIS: Traceback (most recent call last): File "C:\ai-voice-cloning\src\main.py", line 11, in from utils import * File "C:\ai-voice-cloning\src\utils.py", line 21, in import psutil ModuleNotFoundError: No module named 'psutil' Press any key to continue . . .
Sorry it took me so long to get back to you but i’ve been wiped out by a flu. Sadly I can’t find any more info on your issue. You could try to reinstall it from scratch and if you still have the issue, maybe let the author know at git.ecker.tech/mrq/ai-voice-cloning/issues
The tool supports any language but you need to teach it the language (check the wiki for more on that). From what I understand, sooner or later there will be public models for other languages so that you won’t need to train the language (and can use the same process as in this video). Those models for additional languages take very very long to train and require powerful hardware so it might take a while.
is there anyway you can let us download this trained woman's voice so we can use it? Or at least tell us who she is so we can do the training ourselves
Sorry having to disappoint you but my AI legal consultant advised me against doing so, as long as the legality of cloning and using other voices is still uncertain. However, I can tell you that I trained the female voice the exact same way as the male voice in the video. And I found the voice by searching for 20+ minute lectures and speeches on RUclips and it didn't take me too long to find her.
Someone knows how to run it with cpu only I tried then i got a error what said: Attempting to deserialize object on a CUDA device but torch.cuda.is_available() is False. If you are running on a CPU-only machine, please use torch.load with map_location='cpu' to map your storages to the CPU.
The tool supports any language but you need to teach it the language (check the wiki for more on that git.ecker.tech/mrq/ai-voice-cloning/wiki/Training). From what I understand, sooner or later there will be public models for other languages so that you won’t need to train the language (and can use the same process as in this video).
hey i have problem when i type start.bat, i see this error: ImportError: DLL load failed while importing torch_directml_native: The specified procedure could not be found. Press any key to continue . . . can u help me?
Sorry you run into this issue. It has been reported a couple times now (git.ecker.tech/mrq/ai-voice-cloning/issues?type=all&state=open&labels=&milestone=0&project=0&assignee=0&poster=0&q=torch_directml_native) but seems the developer hasn't fixed it yet.
I was able to fix the installation problem! I can produce sound, it's great, but I can't make my voice the best! gives this error: "dataset.py", line 202, in get_wav_text_pair [Training] [2023-11-12T09:53:38.790387] wav = load_audio(audiopath, self.sample_rate) [Training] [2023-11-12T09:53:38.790387] File "E:\TTS\ai-voice-cloning\modules\dlas\dlas\data\audio\unsupervised_audio_dataset.py", line 23, in load_audio [Training] [2023-11-12T09:53:38.790387] from pyfastmp3decoder.mp3decoder import load_mp3 [Training] [2023-11-12T09:53:38.791387] ModuleNotFoundError: No module named 'pyfastmp3decoder'" Is there a way I can use ready-made models or combine them?@@AiVOICETUTOR
I was able to solve the second problem, but there is a third problem that has no error In fact, the speed is low! I have RTX 3060, 64 RAM , ryzen 7 But it takes 15 minutes for 30 seconds in low quality Is this normal?
Glad you solved the other issues. Yeah I think the speed is normal as mentioned by others in the comments here. The speed depends on the amount of CUDA cores that your GPU has
when i typed start.bat
beartype.roar.BeartypeDecorHintPep484Exception: Function torch.onnx.symbolic_helper._onnx_unsupported() return PEP 484 type hint "typing.NoReturn" invalid in this type hint context (i.e., "typing.NoReturn" valid only as non-nested return annotation).
Press any key to continue . .
how
That's the same thing I keep getting, I have no clue what any of it means.
did u guys got it? i have the same issue too
Sorry you all have this issue which seems to be a newly introduced bug but hopefully this is the solution:
Run this command inside your venv: "pip install beartype==0.15.0"
Step by Step:
1.Use python browser
2. Activate the environment
3. Run "pip install beartype==0.15.0"
4. Deactivate the environment
Find more info on it here: git.ecker.tech/mrq/ai-voice-cloning/issues/388
If the fix doesn't work for you, I'm sure the author of the tool will patch this soon (once he does that, just install Tortoise again or run "git pull" via command window in the folder where to tool is installed).
Got it working now, thank you!
This is one of the most well structured tutorials I've seen, and of course the voice sounds excellent. Cheers!
Thank you so much! Reading that made my day!
I'm glad for this tutorial. At first I got some error problems, but when I tried it again it worked out really fine! Thank you very much for this tutorial, this is going to help me so much. It's so sad to see that you only have 600 subscribers, you deserve more. But very happy to see no dislikes on this video! 👍
Thank you so much for your kind words! Much appreciated! I’m glad you didn’t give up and that you were successful when trying again 👍
Ohh my god, took very long time to do it but after all, it was worth it, the problem is many people don't have that much patience to go through with all, but thank you very much your effort is appreciated.❤❤
Thank you, that's great to hear ❤️ I'm glad you didn't give up and got it working in the end! Congrats!
Could not find module 'C:\ai-voice-cloning\venv\Lib\site-packages\torchaudio\lib\libtorchaudio.pyd' (or one of its dependencies). Try using the full path with constructor syntax.
got this error even if the files exist. after the step (start.bat) please help
Wow! I could not tell at first that this voice was AI generate. Your tutorial also has very step by step detailed process, liked and subscribed.
Could you please share the voice data that you used in this video so that we can use it without needing to retrain? I have previously trained some voices but they were not as good as yours.
the laugh at the end 😂
Meta has launched its own TTS model named voicebox,(not realising it fearing its consequences).But they have a github repository named meta tts . You should make a video on this
Thank you for letting me know! I heard about what Meta announced but for me it wasn't that interesting because they won't release the code for us to try out. I hope we'll soon have even better ways to generate TTS locally
@@AiVOICETUTOR Your tutorial is the only one I've been able to follow to successfully install all of this! Thank you so much! I am getting a lot of errors using it though. I tried to train a 1 minute model and it was saying Index Error
User
[Training] [2023-10-14T00:53:15.571531] warnings.warn("Detected call of `lr_scheduler.step()` before `optimizer.step()`. "
[Training] [2023-10-14T00:53:17.514094] C:\AI\ai-voice-cloning\venv\lib\site-packages\torch\utils\checkpoint.py:429: UserWarning: torch.utils.checkpoint: please pass in use_reentrant=True or use_reentrant=False explicitly. The default value of use_reentrant will be updated to be False in the future. To maintain current behavior, pass use_reentrant=True. It is recommended that you use use_reentrant=False. Refer to docs for more details on the differences between the two variants.
[Training] [2023-10-14T00:53:17.514094] warnings.warn(
[Training] [2023-10-14T00:55:06.284260] Disabled distributed training.
[Training] [2023-10-14T00:55:06.284260] Path already exists. Rename it to [./training\JohnKushite\finetune_archived_231014-005259]
[Training] [2023-10-14T00:55:06.284260] Loading from ./models/tortoise/dvae.pth
[Training] [2023-10-14T00:55:06.284260] Traceback (most recent call last):
[Training] [2023-10-14T00:55:06.284260] File "C:\AI\ai-voice-cloning\src\train.py", line 64, in
[Training] [2023-10-14T00:55:06.285268] train(config_path, args.launcher)
[Training] [2023-10-14T00:55:06.285268] File "C:\AI\ai-voice-cloning\src\train.py", line 31, in train
[Training] [2023-10-14T00:55:06.285268] trainer.do_training()
[Training] [2023-10-14T00:55:06.285268] File "C:\AI\ai-voice-cloning\modules\dlas\dlas\train.py", line 408, in do_training
[Training] [2023-10-14T00:55:06.285268] metric = self.do_step(train_data)
[Training] [2023-10-14T00:55:06.285268] File "C:\AI\ai-voice-cloning\modules\dlas\dlas\train.py", line 271, in do_step
[Training] [2023-10-14T00:55:06.286264] gradient_norms_dict = self.model.optimize_parameters(
[Training] [2023-10-14T00:55:06.286264] File "C:\AI\ai-voice-cloning\modules\dlas\dlas\trainer\ExtensibleTrainer.py", line 321, in optimize_parameters
[Training] [2023-10-14T00:55:06.286264] ns = step.do_forward_backward(
[Training] [2023-10-14T00:55:06.286264] File "C:\AI\ai-voice-cloning\modules\dlas\dlas\trainer\steps.py", line 242, in do_forward_backward
[Training] [2023-10-14T00:55:06.286264] local_state[k] = v[grad_accum_step]
[Training] [2023-10-14T00:55:06.287262] IndexError: list index out of range.
And then when trying to generate different things I'd also get an error saying Text too long even though it was likei 4 or 5 words. o____O. I managed to successfully train when I first ran the programme but I don't think I had enough EPOCHS as it sounded like a British man instead of an American Voice that I cloned. It sounded like the voice but without the accent which was strange.
Pretty cool and a good tutorial. I did have an issue the the install bombing the first time so I deleted the install and started over and it went well without red error messages. Tested with the default voice and it worked. Now on to training some new voices. Currently in the middle of training with Sylvester Stallone's voice as a learning process.
Awesome! Glad you didn't give up the first time and that you managed to get it to work. Thanks for sharing that and fingers crossed that Sly sounds good!
This method is very clear. thank you very much.
You're welcome! Im glad you like it.
imo the tts voice was great. very neutral and easy to follow. Your voice is great too don't get me wrong :)
Thank you! I'm glad to hear that!
That's impressive! Thanks for sharing. Terry
Thank you for watching!
Thank you. worked as described.
You're welcome and thanks for your feedback!
After training: "max(): Expected reduction dim to be specified for input.numel() == 0. Specify the reduction dim with the 'dim' argument." :(!
Please guide me on how to use Google Colab to train voice mode
Sorry I haven't used Colab yet
PytorchStreamReader failed reading zip archive: failed finding central directory, i am getting this error, how to solve this ? please help me
I was looking for this video for a long time.
Thanks for the effort. Great video, will do a re-install today.
I want to know if there is a place to download more voices to use. Without training everything by myself.
Thanks in advance.
No problem. I’m glad you found the video! I haven’t seen any downloads for voice models but the wiki has links to voice samples that you’ll need to train yourself: git.ecker.tech/mrq/ai-voice-cloning/wiki/Collecting-Samples. For the RVC method that requires another voice as input rather than text (where I also made a video about), there’s a discord with voice models to download: discord.com/invite/aihub
Hi, thank you very much for your thorough tutorial. But I encountered this when installing TorToiSe TTS: "DEPRECATION: voicefixer 0.1.2 has a non-standard dependency specifier streamlit>=1.12.0pyyaml. pip 23.3 will enforce this behaviour change." . Any idea to solve it?
Hi, does this error make the installation stop? If not, then you can just ignore it. Sadly I couldn't find any information about it
@@AiVOICETUTOR it does make the installation stop but I used another method to install it and it worked. Thanks a lot.
@@AiVOICETUTOR by the way, may I know what graphic card do use for Tortoise TTS?
Cool I'm glad you figured it out. Would you mind sharing the method? BTW I'm using a RTX 4090
I recently trained a model for RVC to clone a singer (I personally know) and it worked quite well. I've seen people use TTS produced audio and put it through RVC to get a better more realistic voice. Now my question: TTS can be trained in many languages, is RVC language specific or could I train a voice in TTS, let it speak a text and then send it through RVC, when I have trained a model there (with the same training material) ? Have you tried to combine the 2 techniques?
I haven't tried to combine those techniques but I read that it works well (at least in English). AFAIK the issue with TortoiseTTS and other languages is that you need a lot of GPU power and time to train a base model for a different language. So far I have not seen any of them appear online. RVC voice cloning works in any language as long as the input voice and the voice to clone are the same language
Tem suporte de texto/voz para Português? Onde consigo baixar a biblioteca caso for necessário.
There’s even an easier way of installing the easy AI Voice, and not only that, but they now have their own website
thats it ! love it !
Awesome!
Tortoise TTS cloned voices don't sound very accurate on their own. Which is why i come with this recommendation: After generating voices with tortoise, give them a pass with RVC. they'll be pretty much on the level of eleven labs v1 afterwards. it's awesome! the main issue with this is that RVC also requires a long training session for it to do its thing. but no biggie, as you can use the same samples you used to train tortoise. just leave it training overnight and GG
Wow thats very interesting! Thanks for sharing this but could you let me know if I understood it correctly? You mean that you use the output .wav files from tortoise and then run them through RVC GUI? That means you trained a model with the same voice input samples in Tortoise AND in RVC beta?
@@AiVOICETUTOR yep, that's exactly it.
Awesome I’ll have to try that!
cool! I assumed the training tool used RVC
@@GraveUypo
Your voice cloning setup appears to be running smoothly. Could you share your system configuration , is your laptop or a desktop PC
Sure, I'm using a desktop PC with i9 13900KF and RTX 4090. I'll try to add the full config to all video descriptions in the coming days
Am I able to start this while waiting on a newer GPU? My current 3080 ti is out of commission in a build I'm waiting to finish, so I am running on my old 1070. Which apparently NSight Compute no longer supports. Am I screwed until then?
how do you restart the URL for the user interface? I shut down my computer and know the URL does not work. do i need to start the Command prompt all the time and what is the command I need to use to start the URL again?
After installing (and closing) the tool you can always open it again by running "start.bat" in the "ai voice cloning" folder. Then open the URL it shows you (127.0.0.1:7860) in a web browser.
@@AiVOICETUTOR it worked thank you :)
Clear and good instructions. 😮 does someone know if this is also possible on M1 or M2 macs?
Thanks. The developer is looking into Mac support but at the moment its not out yet (more here: github.com/neonbjb/tortoise-tts/discussions/115)
Hi. Thanks for the tutorial. It really worked well. After training with around 8 mins audio and your settings (and baking my PC for ~45 mins ^^ ), I got a pleasing result. As the vid is now 1 month old I wonder though, if you have updated your workflow or new suggestions. Although the training setting was slightly changed by the "validate configuration" process (some value from 8 to 7, does it change things you cannot change yourself inside the training settings?) because of VRAM. But then it only used ~12GB in the process (I ve got 24GB). I've got to dive more into the program, but I wonder how much better RVC might be, as you've said yourself? This is already good. Does the quality of the training data have a huge impact on the output quality or is not big, because of the diffusion and just the abstract characteristics of the voice are the base of what's then put into the new model? So many questions and lota more hehe, Thanks again for the well done tutorial!
Hi, thank you very much for your feedback and the kind words! Glad to hear it worked well for you. AFAIK this is still the best free/local method for AI text-to-speech. Although its not perfect, I was happy with the results I got so I never looked into tweaking the workflow. There's probably some room for improvement. RVC is (supposed to be) used to clone a human voice which therefore IMO overall sounds less artificial than the TTS does sometimes. As far as I can tell, the quality of the input voice does matter, though the tool sometimes seems to be very forgiving if that makes any sense. BTW someone in the comments suggested training the same voice in Tortoise AND RVC. Then after generating the .wav file with TTS, run that audio through RVC GUI for better results.
@@AiVOICETUTOR yeah, the only strange thing though, i cant seem to achieve the same quality from the moment I ran the test voice line now, even using a short sentence only and the same seed. I dont know why this could be.
solution: in the settings "autoregressive model" i switched back to autoregressive (it was on my trained model 501 gpt.pth). then the computerish-undertone seemed to have dissapeared, but still not sure, maybe its also to do with other things, have to test more..
Interesting! Thanks for sharing! I'll have to play around with this and do some testing myself
[Training] [2024-06-15T19:20:29.805467] local_state[k] = v[grad_accum_step]
[Training] [2024-06-15T19:20:29.805467] IndexError: list index out of range
i dont know if its an issue, but after this the training stops, i can use the voice but its extremely deep and robotic:
any idea on what to do?
thanks in advance!
This is really cool! I got it to work for a minute but then my computer crashed. I am wondering, do i have to redo everything again to get it to work everytime i shut my computer down or if it crashes?
Did it crash during training or afterwards? If it happens during the training then I would restart the training from scratch. If it happened after the training, you're fine and can select the models to start generating the voice with your text prompts (as shown in the video)
Great content! Please do more text-to-speach with different languages and for Mac users without Nvidia GPU ⭐
Thanks I hope to have something ready for that by the end of next week
Hey just to let you know that I tried to run three different TTS voice cloning tools that are available through Pinokio. Sadly two of them didn't work for me (Bark Voice Cloning and XTTS) and the other one (VALL-E-X) sounds quite bad compared to TortoiseTTS. So I won't be making a new video but if you want to, check them out yourself and see if they work for you. More info on installing Pinokio here: ruclips.net/video/ln1qEglnpMo/видео.html
@@AiVOICETUTOR Thank you for letting me know! Is it safe to say that Elevenlabs still is the best?
Sorry of the late reply. Yeah to my knowledge they are still the best
I have an issue. It seems that I ain't got the hardware it requires (I lack a non-integrated graphics card). WOuld it still be possible to make it work? I'm not particularly savvy with computers
Sadly you can't run it without a dedicated GPU but there is a way to run it via Colab: git.ecker.tech/mrq/ai-voice-cloning/wiki/Installation#colab
Got my voice cloned and it works like charm..btw any idea how can I get this voice I just made work with my assistant? I need to pass the text from the assistant somehow to the tts and make it generate the audio files.. I have already figured out how to play the generated files.
You can use the tool via its API. For more info see here: github.com/neonbjb/tortoise-tts#api I''m not sure if the API calls work with the version from my tutorial or if you need to install the one from that git. Hope that helps
that so good thanks
Thank you too! Glad you like it
Thanks for your video. What languages does it support?
The tool supports any language but you need to teach it the language (check the wiki for more on that git.ecker.tech/mrq/ai-voice-cloning/wiki/Training). From what I understand, sooner or later there will be public models for other languages so that you won’t need to train the language (and can use the same process as in this video).
Hello @AiVOICETUTOR
Thanks for this clean and detailed video on TorToiSe TTS!
I enjoyed the whole process. I was almost done but during the end of "08:53 Train Voice model in TorToiSe TTS" my system got restarted. Since I am a non technical guy and new to this AI tools, I don't know how to start from where it stopped while restarting. How do I get the "TorToiSe TTS" user interface? Do I need to start from the beginning?
Please help me..
Thanks in advance!!
Hi, I'm glad you like the video. You don't have to start from scratch. Just open the "start.bat" file through windows explorer and once it's loaded, open the URL that is shown in the command window. Sadly I couldn't find any information on resuming the training of the model. So worst case is that you need to do all the steps regarding the training again. Hope this helps and good luck!
@@AiVOICETUTOR Thanks for the quick response.
When I tried to generate it gives the following error: "Something went wrong
CUDA out of memory. Tried to allocate 12.00 MiB (GPU 0; 4.00 GiB total capacity; 6.47 GiB already allocated; 0 bytes free; 6.60 GiB reserved in total by PyTorch) If reserved memory is >> allocated memory try setting max_split_size_mb to avoid fragmentation. See documentation for Memory Management and PYTORCH_CUDA_ALLOC_CONF"
You can try to set the autoregressive_batch_size to 1 (more here: github.com/neonbjb/tortoise-tts/issues/37)
@@AiVOICETUTOR Thanks!
I will try.
Hi I get this error when clicking "Transcribe and Process" during training of clone voice
"RuntimeError: cannot reshape tensor of 0 elements into shape [0, -1, 256, 256] because the unspecified dimension size -1 can be any value and is ambiguous #391"
Make sure you don't have any emotions or empty lines at the end of your prompts. More here: git.ecker.tech/mrq/ai-voice-cloning/issues/158 Hope that helps
we need graphics card for this right?,,,, is there colab version available?
Does this work for other languages?
Yes you need a GPU. There’s a collab and it’s English only for now
I LIKE, I Smash d LIKE Button!!
Awesome! Thanks for your support!
Great video, can you share the trained model used to create this video
Sorry having to disappoint you but my AI legal consultant advised me against doing so, as long as the legality of cloning and using other voices is still uncertain. However, I can tell you that I trained the female voice the exact same way as the male voice in the video. And I found the voice by searching for 20+ minute lectures and speeches on RUclips and it didn't take me too long to find her.
thanks for your video. is it possible to train "french" voice ? thanks
The tool supports any language but you need to teach it the language (check the wiki for more on that). From what I understand, sooner or later there will be public models for other languages so that you won’t need to train the language (and can use the same process as in this video). Those models for additional languages take very very long to train and require powerful hardware so it might take a while.
Thank you for the explanations.
I'm curious to try this out but I'm tempted to ask if this can clone a Nigerian voice somewhat accurately. I'm asking because the google colab tool that I previously tried with this AI tend to come out in American accent.
Yes you'd need another model file that is especially made for the Nigerian language. The one that comes with the tool is made for English only, so all other languages trained with it will have an accent. I hope it won't be much longer before we can use more languages other than English.
@@AiVOICETUTOR, Can we train our own model?
can we use it for different languages? if yes, how?
Can i make tts to vegeta voice in this... ❤
The transcription of my audio samples had some mistakes, but once I corrected the text in the 'whisper.json' and recreated the dataset, the tokens weren't updated to reflect the corrected text in the 'whisper.json'. It seems my only options are to transcribe again, but then it just makes the same mistakes. Any idea how to fix this while keeping my manually edited text?
I couldn't find any info about editing the transcript on the official git. It must be doing some more stuff in the background so you can't just edit the .json? My best advice would be to remove the voice bits that don't transcribe properly from the input data.
Hi, do you know if I can do all those steps on MAC system? Thanks
Hi, yes this should work on a Mac if you have a M1 or M2 chip. There's some more info on the official git: github.com/neonbjb/tortoise-tts#apple-silicon
i have created more folders under the "voices" folder, but when i click refresh voices, i only get the first folder i created, anyone else with the same problem?
Keep in mind that the smaller RVC voice models are meant to clone your voice (ruclips.net/video/5i_Pyw0gH-M/видео.html) and won't work with TTS.
After run start.bat, I got this error: "Task exception was never retrieved future:
This might be the solution to your issue: "Add pydantic==1.10.11 to the end of the requirements.txt file, re-run setup-cuda.bat then run start.bat. Everything will be ok." More info here: git.ecker.tech/mrq/ai-voice-cloning/issues/300. Hope it helps.
Can you share exact settings you used to train female voice. And to generate voice.
I trained the female voice the exact same way as the male voice in the video. And also generated it the same way. I varied between the ultra fast and fast presets without changing any of the other default settings. Hope that helps
@@AiVOICETUTOR Awesome. Thank you. Can you talk about the settings of batch size and gradient size? The validate training configuration does sometime not work for me as it starts to give error. It is recommended to reduce batch size to resolve this issue, but don't know optimal batch size. I have 4070Ti
The batch size is a tradeoff between speed and accuracy. Smaller values will be slower but more accurate while higher values will be faster and less accurate. Maybe try lowering it to 32 or 64.
I wouldn’t change the gradient accumulation size if it works for you with the lower batch size but here’s what Bard says: Gradient accumulation size is a parameter that controls how many batches of data are processed before the model's parameters are updated. A larger gradient accumulation size means that more data will be processed before the parameters are updated.
Gradient accumulation can be used to train on larger batch sizes than your machine would normally be able to fit into memory. This is because the model's parameters are only updated after a certain number of batches have been processed.
The optimal gradient accumulation size will depend on the specific model and dataset being used. A good starting point is to use a value of 1. If the model is not converging, you can try increasing the gradient accumulation size. If the model is becoming unstable, you can try decreasing the gradient accumulation size.
I followed your tutorial step by step, yet it doesn't seem to work for me. Once I go into the command prompt, all I get is error messages. I managed to get the Cuda setup bat to run by running it through a double click. Trying so in the command prompt always gives an error. After it was done, I tried start.bat, which says that I should enter a parameter. I couldn't find anything relating to that, so, just to try, I entered a path to a folder, which was then opened in a new window. Only by starting start.bat, which simply refers to activate.bat in another folder, as an administrator, could I get it to not show an error, but it basically is stuck on that ever since, not doing any downloads or the like. So, while this is a good video, it looks like this is a bust for me.
Sorry that it didn't work for you. What does the first error message you're getting say?
How can you setup to do real time voice changing? Like when talking in a game or something.
I know it’s possible but I haven’t looked into it myself. It’s on my list though and I’ll make a video once I figure it out
Failed here at the Training Model, in the console it says no NVIDIA driver found albeit I installed it (got AMD Radeon GPU), there is any workaround for it? Lets say CPU only?
Sorry but afaik, you still need an NVIDIA card to run this. Installing the driver alone won't do the trick
Oh man I got everything working up until trying to generate a voice promt. I get cuda out of memory 😭😭 I'm not sure how to proceed.
Odd that you could do everything before that without having any memory issues. Have you restarted your PC and tried generating the voice again?
@@AiVOICETUTOR yea I think it's the only next move to make. But once I turn off my computer I'll have to redo everything again in the CMD hey? There's exe installed that I can run that will make it work is there? Because I ran into an issue the first time with my computer crashing. When I went back to the web interface it no longer worked because the commandline had been shut down and I wasn't sure if there was a short cut to access what had already been installed through this.
Once installed, you can launch the webUI by running „start_http.bat“ which should open your web browser and you can go straight to the voice prompts. You won’t have to do anything in a command window again
I managed to reach the "start.bat" step but it give me this error: No module named 'psutil'.
Also it states more things on line 11 in main.py and line 21 in utils.py
How can I solve?
I think you need to downgrade your python: more info here: git.ecker.tech/mrq/ai-voice-cloning/issues/175
Does not work for me. I followed every step. When I execute start.bat the following error message will appear:
(venv) C:\ai-voice-cloning>start.bat
(venv) C:\ai-voice-cloning>call .\venv\Scripts\activate.bat
Traceback (most recent call last):
File "C:\ai-voice-cloning\src\main.py", line 11, in
from utils import *
File "C:\ai-voice-cloning\src\utils.py", line 40, in
from tortoise.api import TextToSpeech as TorToise_TTS, MODELS, get_model_path, pad_or_truncate
ModuleNotFoundError: No module named 'tortoise.api'
Check this out: github.com/neonbjb/tortoise-tts/issues/196. Maybe you only need to run this command: "python setup.py install"?
@@AiVOICETUTOR I kind of get the same issue, and tried resolving it with the thread. Firstly I get this:
(venv) C:\ai-voice-cloning>call .\venv\Scripts\activate.bat
Traceback (most recent call last):
File "C:\ai-voice-cloning\src\main.py", line 11, in
from utils import *
File "C:\ai-voice-cloning\src\utils.py", line 40, in
from tortoise.api import TextToSpeech as TorToise_TTS, MODELS, get_model_path, pad_or_truncate
ModuleNotFoundError: No module named 'tortoise'
When I change it to "from api import TextToSpeech as TorToise_TTS, MODELS, get_model_path, pad_or_truncate" I instead get this:
(venv) C:\ai-voice-cloning>call .\venv\Scripts\activate.bat
Traceback (most recent call last):
File "C:\ai-voice-cloning\src\main.py", line 11, in
from utils import *
File "C:\ai-voice-cloning\src\utils.py", line 40, in
from api import TextToSpeech as TorToise_TTS, MODELS, get_model_path, pad_or_truncate
ModuleNotFoundError: No module named 'api'
when i typed start(.)bat it says FileNotFoundError: Could not find module 'D:\ai-voice-cloning\venv\Lib\site-packages\torchaudio\lib\libtorchaudio.pyd' (or one of its dependencies). Try using the full path with constructor syntax.
any solution?
EDIT:
you have to use python v 3.9 or 3.10, the newest version doesnt work
also before that i got this
Does .\venv\Lib\site-packages\bitsandbytes\cuda_setup\. specify a file name
or directory name on the target
(F = file, D = directory)?
@@bengreviews8286 delete folder and restart a clean install
@@world_report I switched to python v 3.10 it's worked!
Sweet. Glad you sorted it out!
@@AiVOICETUTOR please make this comment pinned cuz I waste a lot of time using latest version
Could not find module 'C:\ai-voice-cloning\venv\Lib\site-packages\torchaudio\lib\libtorchaudio.pyd' (or one of its dependencies). Try using the full path with constructor syntax.
got this error even if the files exist. after the step (start.bat) please help
This seems to be a strange bug but someone else that commented here fixed it by switching to Python 3.10
If I closed my project and again want to open it tomorrow, what command should I give to activate and use it again?
Run "start.bat" in the "ai voice cloning" folder. Then open the URL it shows you (127.0.0.1:7860) in a web browser
wow that so good thanks
Thank you very much! Glad you like it
I get this error when transcribing: "cannot reshape tensor of 0 elements into shape [-1, 0] because the unspecified dimension size -1 can be any value and is ambiguous"
me as well. Did you find any solution
@@Chriscs7 no unfortunately not
Make sure you don't have any emotions or empty lines at the end of your prompts. More here: git.ecker.tech/mrq/ai-voice-cloning/issues/158 Hope that helps
Can we export the trained voice model into a format that a free speech synthesizer can use? Maybe in Hungarian?
The tool supports any language but you need to teach it the language (check the wiki for more on that). From what I understand, sooner or later there will be public models for other languages so that you won’t need to train the language (and can use the same process as in this video).
Thank you for the awesome tutorial. It's weird though, I followed everything until it came to the training. Since my source audio isn't 10min, but a little shorter, I thought I'll set the epochs to 400. Everything else was the same. I clicked train and got this error: 'utf-8' codec can't decode byte 0x81 in position 2: invalid start byte
Then I changed it back to 500 epochs, tried again and it worked. lol Either way it's working now if I don't change that number.
Hmm that UTF-8 thing is an odd bug. I'm glad you got it to work and thanks for sharing your solution! Maybe check if you have any special characters in your folder path or file names
I ran in to this when trying to train.
[Training] [2023-11-04T18:17:17.305090] from axial_positional_embedding import AxialPositionalEmbedding
[Training] [2023-11-04T18:17:17.308976] ModuleNotFoundError: No module named 'axial_positional_embedding'
is there someway to fix it ?
Seems you’re not the only one with this issue: git.ecker.tech/mrq/ai-voice-cloning/issues/422 I hope the developer will fix it soon!
Thanks for the reply. I hope so, would love to try out this tts. Great video keep up the good work. @@AiVOICETUTOR
Thank you 🙏
Hey I got a question. I tried the TTS and it works really fine, but it takes decades to generate one sentence. Is there a way to speed up the generating process?
Hey AFAIK it's all about the CUDA cores. More cores means more speed. Don't think there's another way to speed it up other than upgrading the GPU
I keep getting the "Empty dataset" error when clicking on the "validate training option" 10:04
Any thoughts on what could be wrong? Using Python 3.10
You could try making sure that you're not running it off a network drive and that there are no spaces or special characters in the folder path
Hi this is amazing video. Does it work for hindi audio also?
No, you can read here why it's not that easy: github.com/neonbjb/tortoise-tts/issues/567. I made a video about how to clone your own voice (which works in any language): ruclips.net/video/5i_Pyw0gH-M/видео.html&lc=UgxvH3x_DKWkw8gLvkZ4AaABAg
Thank you for the quick reply. Understand
This can work offline once everything has been downloaded or it needs constant web connection?
Yes it can run 100% offline once everything is set up
sorry i just have to ask. so if i do want to reopen the application how do i do it? and if i want to download the generated audio how can i do this too?
After installing (and closing) the tool you can always open it again by running "start.bat" in the "ai voice cloning" folder. Then open the URL it shows you (127.0.0.1:7860) in a web browser. The generated audio can be downloaded by clicking the arrow next to the audio file player in the Web UI
Hey, I can't click that url in 6:23. Can you help me please?
Hey you might need to hold CTRL while clicking or alternatively you can just copy the url and paste it in your browser
this work with greek language?
solution please .........(venv) D:\ai-voice-cloning>call .\venv\Scripts\activate.bat
Whisper detected
Error: No module named 'vall_e'
Error: No module named 'bark'
Make sure that you are using Python 3.9 or 3.10. Python 3.11 won't work AFAIK
do you know any way to make TTS speaks like original input audio? I want to make AI dubbing on my movie and can't figure it to speak like original language audio file :(
As far as I know, this is still the best free TTS voice method. It's not perfect yet but it's only gonna get better from here on. Hopefully there will be new models out soon that will improve the quality
Is there anyway to get this high quality tts in the Google Collab?
Yeah from what I can tell there shouldn’t be any difference to running this tool locally compared to colab
Question about the part where you train your model. My auto regressive model only seems to go up to 201. I use it but I don't get good results. The higher that number the more accurate to the training data? Is it possible to dedicate more GPU resources to train it better? I have a 4090 currently for this. If I keep hitting "train" over and over again, each time it finishes I can get a more and more accurate model? Sorry i'm new to this.
edit: after I get done training it, I hit train again, because I assume that will make the model better, but nothing happens, no output or anything.
Makes sense that you can't just train it again but thanks for confirming it. If you set the 500 epochs and you only get a model saying 201 then the training must have stopped too soon (the highest regressive number should be the same as the epochs you set). Do you know if the training progress said it's doing 500 epochs? And maybe check disk space too. BTW Too many epochs can make it sound worse so its a good idea to try all the regressive models
@@AiVOICETUTOR interesting, i'll try that and keep more of an eye on it. Disk space might be an issue i'll check that as well. I'll keep this thread updated and try the different regressive models.
Another question I have, lets say one regressive model sounds better but still not quite there, and the one i'm currently using and that one are the ONLY regressive models I have, what are my other choices? For reference, the voice i'm training I have about an hour of clean, clear audio and each clip of that hour is spliced into 5 second chunks.
Not sure if that's what you meant but the "Save Frequency (In Epochs)" determines how many models you'll end up with. So if you lower that number you'll end up with more models to try. The tool definitely works better with some voices while other voice don't seem to work well (yet).
during training the process isnt progressing for quite a while and the cmd says "RuntimeError: FIND was unable to find an engine to execute this computation" is it normal or the process is actually failed?
Seems this is an issue with the CUDA installation.Maybe try to reinstall the Toolkit. Hope it helps
How much words or characters can speech in prompts
I have tried it with a lot of text and never reached any limit
Hello so I have been following all the steps and even checking to make sure everything is installed. However, when trying to run the command setup-cuda.bat I encounter an error that says "setup-cuda.bat is not recognized as an internal or external command operable program or batch file." I even double checked the Cuda files were installed. Any ideas what I should do?
Hello this sounds like a general issue with how you run the bat file. Did you try to double click or open the .bat file (like any executable) or are you trying to run it from a command window?
@@AiVOICETUTOR I was trying to run it from the command window.
What happens when you double click it (or right click, and then open it)?
Someone please help. I followed everything in the video exactly as it's shown but when I get to training my voice when I click train after adding the folder and a good chunk of wav files for it to train from I get an error stating [WinError 2] The system cannot find the file specified and it stops training.
Make sure that your folder isn’t in the cloud or on a network drive and that there are no spaces or special characters in the folder path
@@AiVOICETUTOR I don't have any network of cloud folders everything is stored on my computer's C: drive. The file path is C:\ai-voice-cloning\training\voicetest/train.yaml
I even did an entire fresh install cloning the github again and repeating all the steps and I have the same issue
When viewed in File Explorer the file path is C:\ai-voice-cloning\training\voicetest/train.yaml
But when I click the drop down for training configuration it auto detects ./training/voicetest/train.yaml
There shouldn't be any special characters or spaces and everything is stored locally on my PC with full access to file editting reading and writing. Really no idea what's going on :(
Your path looks fine. I guess your audio file names are fine too. Unless someone else has any idea, I'd suggest to post your issue to the developer at git.ecker.tech/mrq/ai-voice-cloning/issues
I'll try that. Thanks for replying and trying to help@@AiVOICETUTOR
Hello help me error
(When start Training get an Error:
(result, consumed) = self._buffer_decode(data, self.errors, final)
UnicodeDecodeError: 'utf-8' codec can't decode byte 0x84 in position 0: invalid start byte)
If you use the learning rate scheduler (calling scheduler.step()) before the optimizer’s update (calling optimizer.step()), this will skip the first value of the learning rate schedule. If you are unable to reproduce results after upgrading to PyTorch 1.1.0, please check if you are calling scheduler.step() at the wrong time.
Hello! By any chance, do you you have any special characters in the filenames?
Sadly no matter what number for the echos I have. (I started with 500) I always get the error "utf-8...". I've changed the Echos number to 400, 499, 600. Nothing helps.
Sorry to read that you're also affected by this bug. Maybe check if you have any special characters in your folder path or file names
So I have a new issue. A reinstall fixed my earlier issues. I trained my voice and set the autoregressive model but even after refreshing voice lists I can't select it, it's just not there?
Checked the CMD window and it appears to load the autoregressive model fine but no voice option
That is very strange if you didn't move any files around. Even though it's unlikely to be the cause, could you check and make sure that you have the tool installed on a local drive (not network) and in a folder that doesn't contain any spaces or special characters
@@AiVOICETUTOR I was an idiot. I hadn't put my audio files in a new folder so they were just sat in 'Voices' lol
My new issue is that the generation always gives the voice a very posh sounding, southern English accent. There's definitely elements of the voice I cloned there but it just sounds weird.
Thanks for sharing the solution to your issue! For your new issue, I suggest you make sure your input voice data is clean (so that no other voices or sounds are in it). You could also try increasing or decreasing the length of the input voice or increase or decrease the number of epochs. Also I have noticed that it just doesn't work well with some voices sadly (especially if they have a strong accent)
@@AiVOICETUTOR thanks, I'll try that. I also discovered a TTS to RVC pipeline which generates the voice in Tortoise from a python file then runs it through a trained RVC model. I've set it up and have left a voice training in RVC. Looking forward to trying it.
Also, my earlier issue (IndexError: List out of Range). It turns out that when setting up the training config you need to make sure that the batch size is evenly divisible by the gradient accumulator otherwise you get that error. Changing mine to 96 worked.
Cool! Again thanks for sharing the fix and the Tortoise to RVC pipeline. It’s something I’ve been wanting to try for a while myself now as it should give near perfect TTS output, right? Hope it works well for you and would love to hear what you think about the quality of the output.
can you create a macro to automate setup in windows 10/ 11?
Sorry I missed your reply. Not sure if you've know about Pinokio (ruclips.net/video/ln1qEglnpMo/видео.html). Sadly it doesn't support TortoiseTTS yet but I hope it will be added soon.
Honestly how much will you rate it to the generated voice !!!Is it really that good means match with the input voice
I have trained my voice using rvc and i have the .pth file .can i use that file without training here in this model
How to use it in Google collab ! Any link?
Sadly the .pth files trained in RVC are not compatible with this tool. Check out this colab but I think it doesn’t have a proper UI: colab.research.google.com/drive/1wVVqUPqwiDBUVeWWOUNglpGhU3hg_cbR
@@AiVOICETUTOR thanks mate
Will it still work without the Cuda?
Yeah it works with AMD cards too. Check the wiki for more info: git.ecker.tech/mrq/ai-voice-cloning/wiki/Installation
If I'm just watching this now, should I use Python 3.11?
Best use 3.9 or 3.10 just to be safe. 3.11 was already out when I made the video but IIRC people were having issues with 3.11
Hi, I'm getting this error when I type start.bat: Traceback (most recent call last):
File "C:\TTS\ai-voice-cloning\src\main.py", line 11, in
from utils import *
File "C:\TTS\ai-voice-cloning\src\utils.py", line 29, in
import torchaudio
ModuleNotFoundError: No module named 'torchaudio'
Press any key to continue . . .
Hi check this out: github.com/neonbjb/tortoise-tts/issues/420. Hope it helps
@@AiVOICETUTOR Thank you. How do I start a new installation. I initially installed Python 3.11 but I've uninstalled it and installed Python 3.10 just like your tutorial. What do I delete or uninstall to start a fresh installation.
You only need to delete the folder where you did the "git clone". And then start from there in another folder
After start.bat I receive a "No module named 'tortoise'" error, any ideas?
Check this out: github.com/neonbjb/tortoise-tts/issues/196. Maybe you only need to run this command: "python setup.py install"?
When I put in start.bat, I get this error
FileNotFoundError: Could not find module 'E:\TorToiSe TTS\ai-voice-cloning\venv\Lib\site-packages\torchaudio\lib\libtorchaudio.pyd' (or one of its dependencies). Try using the full path with constructor syntax.
I have no idea what to do from here.
I deleted everything and tried again, but now I'm getting this when I put in start.bat
Traceback (most recent call last):
File ".\src\main.py", line 11, in
from utils import *
File "E:\TorToiSe TTS\ai-voice-cloning\src\utils.py", line 28, in
import torch
File "C:\Users\purpl\AppData\Local\Programs\Python\Python38\lib\site-packages\torch\__init__.py", line 451, in
raise ImportError(textwrap.dedent('''
ImportError: Failed to load PyTorch C extensions:
It appears that PyTorch has loaded the `torch/_C` folder
of the PyTorch repository rather than the C extensions which
are expected in the `torch._C` namespace. This can occur when
using the `install` workflow. e.g.
$ python setup.py install && python -c "import torch"
This error can generally be solved using the `develop` workflow
$ python setup.py develop && python -c "import torch" # This should succeed
or by running Python from a different directory.
Sorry it's not working for you. Unfortunately I can't find anything about your issue. You could try it again in a few days, assuming the dev will fix it or you could post the error you're getting at git.ecker.tech/mrq/ai-voice-cloning/issues
I have rtx 2060 but it would finish in 36 hours? Is that normal?
Can you make sure that the tool is actually using your GPU and not the CPU? You can check task manager while the training is running. The tool is using CUDA and the GPU I'm using has 16,384 cores while the 2060 has 1920 cores so it should be slower but i'm not sure if it should be that slow.
Lowering batch files could affect voice quality !!!!
As I am getting error
Torch.cuda.outofmemory....
Anu solution !!
Are you getting the error when trying to train the voice or when trying to generate the voice? Lowering the batch size for training shouldn't affect the quality. Maybe set it to 32 or 64 and give it another try
WHEN I RAN IT DID THIS:
Traceback (most recent call last):
File "C:\ai-voice-cloning\src\main.py", line 11, in
from utils import *
File "C:\ai-voice-cloning\src\utils.py", line 21, in
import psutil
ModuleNotFoundError: No module named 'psutil'
Press any key to continue . . .
Are you using a Mac by any chance?
nope
i'm using windows
Sorry it took me so long to get back to you but i’ve been wiped out by a flu. Sadly I can’t find any more info on your issue. You could try to reinstall it from scratch and if you still have the issue, maybe let the author know at git.ecker.tech/mrq/ai-voice-cloning/issues
jesus, same here bud, hope you get better soon tho
Thanks man! Feels like everyone is having it atm. You get well soon too!
compared with RVC which one is better?
RVC I’d say
Supporst other languages other than english?
The tool supports any language but you need to teach it the language (check the wiki for more on that). From what I understand, sooner or later there will be public models for other languages so that you won’t need to train the language (and can use the same process as in this video). Those models for additional languages take very very long to train and require powerful hardware so it might take a while.
is there anyway you can let us download this trained woman's voice so we can use it? Or at least tell us who she is so we can do the training ourselves
Sorry having to disappoint you but my AI legal consultant advised me against doing so, as long as the legality of cloning and using other voices is still uncertain. However, I can tell you that I trained the female voice the exact same way as the male voice in the video. And I found the voice by searching for 20+ minute lectures and speeches on RUclips and it didn't take me too long to find her.
AOC shouldn't be that hard to find 😁
Someone knows how to run it with cpu only
I tried then i got a error what said:
Attempting to deserialize object on a CUDA device but torch.cuda.is_available() is False. If you are running on a CPU-only machine, please use torch.load with map_location='cpu' to map your storages to the CPU.
Sorry for the late reply! AFAIK It requires a graphics card and won't work with CPU only.
Where are the rest of the voices, I only have "random" voices ...
The .pth files are not compatible with the ones from RVC but you can download additional TTS voices here: huggingface.co/ecker/tortoise-tts-models
Put them in the “finetunes” folder like here: huggingface.co/ecker/tortoise-tts-models/tree/main
Do this also work with other languages?
The tool supports any language but you need to teach it the language (check the wiki for more on that git.ecker.tech/mrq/ai-voice-cloning/wiki/Training). From what I understand, sooner or later there will be public models for other languages so that you won’t need to train the language (and can use the same process as in this video).
Can we have conversation of two different voices?
You can create the tts of each voice separately and then use an editing tool to turn it into a conversation
hey i have problem
when i type start.bat, i see this error:
ImportError: DLL load failed while importing torch_directml_native: The specified procedure could not be found.
Press any key to continue . . .
can u help me?
Sorry you run into this issue. It has been reported a couple times now (git.ecker.tech/mrq/ai-voice-cloning/issues?type=all&state=open&labels=&milestone=0&project=0&assignee=0&poster=0&q=torch_directml_native) but seems the developer hasn't fixed it yet.
I was able to fix the installation problem! I can produce sound, it's great, but I can't make my voice the best! gives this error:
"dataset.py", line 202, in get_wav_text_pair
[Training] [2023-11-12T09:53:38.790387] wav = load_audio(audiopath, self.sample_rate)
[Training] [2023-11-12T09:53:38.790387] File "E:\TTS\ai-voice-cloning\modules\dlas\dlas\data\audio\unsupervised_audio_dataset.py", line 23, in load_audio
[Training] [2023-11-12T09:53:38.790387] from pyfastmp3decoder.mp3decoder import load_mp3
[Training] [2023-11-12T09:53:38.791387] ModuleNotFoundError: No module named 'pyfastmp3decoder'"
Is there a way I can use ready-made models or combine them?@@AiVOICETUTOR
I was able to solve the second problem, but there is a third problem that has no error
In fact, the speed is low!
I have RTX 3060, 64 RAM , ryzen 7
But it takes 15 minutes for 30 seconds in low quality
Is this normal?
Glad you solved the other issues. Yeah I think the speed is normal as mentioned by others in the comments here. The speed depends on the amount of CUDA cores that your GPU has
Please help: "something went wrong connection errored out"
Sorry but I can't find anything about this issue. I hope it's a bug that will be fixed in a future release
Save Frequency (in epochs) 100 Validation Frequency (in epochs) 100 = 1day training model xD