v0.21.0

[erogol]

What's Changed

• Remove duplicate/unused code by @eginhard in Remove duplicate/unused code #3243
• Making the Model Manager's Progress bar statically accessible via the class. by @FlorianEagox in Making the Model Manager's Progress bar statically accessible via the class. #3297
• More informative error for wrong --language argument by @eginhard in More informative error for wrong --language argument #3294
• Don't pass quotes to espeak by @eginhard in Don't pass quotes to espeak #3286
• Fix tts_with_vc by @eginhard in Fix tts_with_vc #3275
• Misjudgment of is_multi_lingual When Loading Multilingual Model via model_path by @TITC in Misjudgment of is_multi_lingual When Loading Multilingual Model via model_path #3273
• Introducing Development Dockerfile by @Kaszanas in Introducing Development Dockerfile #3263
• update deepspeed version by @WeberJulian in update deepspeed version #3281

New Contributors

• @FlorianEagox made their first contribution in Making the Model Manager's Progress bar statically accessible via the class. #3297
• @TITC made their first contribution in Misjudgment of is_multi_lingual When Loading Multilingual Model via model_path #3273
• @Kaszanas made their first contribution in Introducing Development Dockerfile #3263

Full Changelog: v0.20.6...v0.21.0

---

This discussion was created from the release v0.21.0.

[erew123]

Since updating to the new tts_models/multilingual/multi-dataset/xtts_v2 model (today) and also updating from TTS 0.20.6 to 0.21.0, while using a sample UK English voice, it now comes out sounding like a quite strange American and no longer like the sample wav file, which it did on the prior version of xtts_v2 (no matter how many samples I create).

Or perhaps I should say, it starts out kind of ok for the first 10 seconds and then starts to trend towards sounding American, which it didnt before.

I appreciate its all a work, in progress and not sure if this is the right place to feed back about the models.

EDIT - Because its not showing clearly on this message chain. There are some here are some solutions for "It keeps downloading" and "I want to use the old model". #3301 (reply in thread)

Thanks

[erogol]

We are discussing supporting earlier versions. Until we come to a conclusion...

You can always download the previous model and then manually load it.

You can browse the branches and find the version you need -> https://huggingface.co/coqui/XTTS-v2

[erogol]

You can also provide your voices so that we can add those to the training, then you never lose them again plus you would get the best possible output. If you are interested ping me on Discord with a DM.

[carinae]

Thank you erew123

The upgrade of tts remove the dowloading of the new model, then I replaced the in new model path, model.pth and vocab.json by the old one.
It's working again. (with the issue of the pronunciation of the point, but can be managed by replacing by ',')

I shared the voice on reddit (it's from a free Lexibook based on J.Verne Book) , so the voice I'm using most

french voice used

Example -> answer to question about constitution of atom

Thank you for you job Erogol, coqui-ai is the best open source result for the French accent (with model 2.0.2 :) )

[dancodedev]

I think the problem with the points only happens when you use tts through cli. I also had that problem, making the inference like this code. I don't have the problem with the points.

from TTS.api import TTS
tts = TTS("tts_models/multilingual/multi-dataset/xtts_v2", gpu=True)

tts.tts_to_file(
    text="It took me quite a long time to develop a voice, and now that I have it I'm not going to be silent.",
    file_path="output.wav",
    speaker_wav="/path/to/target/speaker.wav",
    language="en"
)

[Tessory]

How in that kind code use a custom model path?
where here to add path:
tts = TTS("tts_models/multilingual/multi-dataset/xtts_v2", gpu=True)
?

[erew123]

You can always download the previous model and then manually load it.

Do you mean using the --model_path and pointing it your downloaded files?

Id just like to say that the previous version 2.0.2 sounded pretty good, close to the voice that I was sampling. And this latest build 2.0.3, sounds at best 60% as good with very strange accents and it degrades after 15-20 seconds into generation.

I'm absolutely not having a go! Don't get me wrong. I really love what you guys are doing and think its fantastic! But I do want to give some honest feedback :)

Thanks!

[erogol]

I don't know how you use TTS. But you can check the docs. It shows you to load models manually in different scenarios.

Sure thanks for the feedback. Noted. You can always keep using v2.0.2 for your voice until you are satisfied with any new models.

[erew123]

@erogol Thanks!

Just to confirm to others, you can use a custom model path as follows:

tts --model_path c:/xttsv2_2.0.3/ --config_path c:/xttsv2_2.0.3/config.json --text "This is a test" --speaker_wav c:\AI\Female.wav --language_idx en --use_cuda true

I copied the entire 2.0.3 model files to a folder I called xttsv2_2.0.3 (but it could be any of the models you want to use)

From there you need to specify the --model_path though you dont add model.pth. So I added --model_path c:/xttsv2_2.0.3/

You also need to specify --config_path though you DO list the config file name --config_path c:/xttsv2_2.0.3/config.json

Obviously, you can do the same for any other model versions, on any path of your choosing.

[allenhs]

I have the same issue with the 2.0.3 model. All the voices I created (American voices, started to sound British or have weird accents and like others have said they would degrade as the audio would go on. They were near perfect on 2.0.2. I'm happy for now that I can downgrade to 2.0.2 and keep using that.

[ElhamAhmedian]

Keep getting:
"tts_models/multilingual/multi-dataset/xtts_v2 is already downloaded however it has been changed. Redownloading it..."
No matter what I do.

[erew123]

Your TTS isnt updated would be my guess. In your Python environment, at a terminal you can do a pip list and see what version of TTS you are using. It should be 0.21.1. If its not then the pip install --upgrade tts

[carinae]

Upgrading TTS worked for me, the model keep the same (previous 2.0.2) and not redownloading it...

Attempting uninstall: tts
Found existing installation: TTS 0.20.6
Uninstalling TTS-0.20.6:
Successfully uninstalled TTS-0.20.6
Successfully installed tts-0.21.1

As erew123 told you, you maybe need to activate the environement where you run coqui-ai...

[ElhamAhmedian]

[XTTS] Loading XTTS...
 > tts_models/multilingual/multi-dataset/xtts_v2 is already downloaded.
 > tts_models/multilingual/multi-dataset/xtts_v2 is already downloaded however it has been changed. Redownloading it...
 > Downloading model to C:\Users\elham\AppData\Local\tts\tts_models--multilingual--multi-dataset--xtts_v2
 10%|██████████▋                                                                                                   | 182M/1.87G [02:21<15:29, 1.81MiB/s

It is the latest version of TTS, and I am not using the command line, its directly from python.

[erew123]

If your python environment is on that version of the TTS engine (not the model that its downloading, the engine e.g. TTS.exe etc) then I don't know what else it could be. I only observed the behaviour of re-downloading the model, when the TTS engine hadn't been updated to the current build. If you have confirmed with pip list that you have the current TTS 0.21.1 then this will be something else... maybe the model hash file is funny. I'm not a developer here so maybe they can help you

This is what id expect you would see on the pip list at your command prompt.

[marioguima]

See it "Model xxts_v2 always downloads (solution)"
#3304

[marioguima]

TTS is truly amazing, and I know there are many updates to come.
However, I also needed to go back to version 2.0.2 due to the mixed accent in the pronunciation of the Brazilian Portuguese language.

[Qualzz]

I confirm using 2.0.2 lead to waaaay better results, the new model kinda sounds like someone non native tries to speak the language.
For exemple with french, it sounds like a bad russian imitation.
Thank you for your work nonetheless and I understand it's hard to evaluate a model especially when working on so much different languages !