🐸 TTS roadmap

[erogol]

These are the main dev plans for 🐸 TTS.

If you want to contribute to 🐸 TTS and don't know where to start you can pick one here and start with our Contribution Guideline. We're also always here to help.

Feel free to pick one or suggest a new one.

Contributions are always welcome 💪 .

v0.1.0 Milestones

• Better model config handling [Discussion] Ideas for better model config management #21
• TTS recipes for public datasets.
• TTS trainer API to unify all the model training scripts.
• TTS, Vocoder and SpeakerEncoder model abstractions and APIs.
• Documentation for
  • Implementing a new model using 🐸 TTS.
  • Training a model on a new dataset from gecko.
  • Using Synthesizer interface on CLI or Server.
  • Extracting Spectrograms for Vocoder training.
  • Contributing a new pre-trained 🐸 TTS model.
  • Explanation for Model config parameters/

v0.2.0 Milestones

• Grapheme 2 Phoneme in-house conversion. (Thx to gruut 👍 )
• Implement VITS model.

v0.3.0 Milestones

• Implement generic ForwardTTS API.
• Implement Fast Speech model.
• Implement Fast Pitch model.

v0.4.0 Milestones

• Trainer API v2 - join the discussion
• Multi-speaker VCTK recipes for all the TTS.tts models.

v0.5.0 Milestones

• Support for multi-lingual models
• YourTTS release 🚀

v0.6.0 Milestones

• Add ESpeak support
• New Tokenizer and Phonemizer APIs New Tokenizer API #937
• New Model API Update models (Rebased) #1078
• Splitting the trainer as a separate repo 👟Trainer
• Update VITS model API
• Gradient accumulation. Accumulate grads ( Larger batch size for low gpu memory) #560 (in 👟)

v0.7.0 Milestones

• Implement Capacitron 👑 @a-froghyar 👑 @WeberJulian
• Release pretrained Capacitron

v0.8.0 Milestones

• Separate numpy transforms
• Better data sampling for VITS
• New Thorsten DE models 👑 @thorstenMueller

🏃‍♀️ Milestones along the way

• Implement End-to-end training API for ForwardTTS models a vocoder. ForwardTTSE2E implementations and related API changes  #1510
• Implement a Python voice synthesis API.
• Inject phonemes to the input text at inference. SSML support #1452
• AdaSpeech1/2 https://arxiv.org/pdf/2104.09715 and https://arxiv.org/abs/2103.00993
• Let the user pass a custom text cleaner function.
• Refactor the text cleaners for a more flexible and transparent API.
• Implement HifiGAN2 (not the vocoder)
• Implement emotion and style adaptation.
• Implement FastSpeech2 (https://arxiv.org/abs/2006.04558).
• AutoTTS 🤖 (👑 @loganhart420)
• Watermarking TTS outputs to sign against DeepFakes.
• Implement SSML v0.0.1
• ONNX and TorchScript model exports.
• TensorFlow run-time for training models.

🤖 New TTS models

• AlignTTS (@erogol)
• HiFiGAN (Hifigan Vocoder #16 👑 @rishikksh20 and @erogol)
• UnivNet Vocoder ( 👑 @rishikksh20)
• VITS paper
• FastPitch source
• Alignment Network paper
• End2End TTS combining aligner + tts + vocoder.
• Multi-Lingual TTS (Add multilingual support #11 👑 @WeberJulian )
• ParallelTacotron paper (open for contribution)
• Efficient TTS paper (open for contribution)
• Gaussian length regulator from https://arxiv.org/pdf/2010.04301.pdf (open for contribution)
• LightSpeech from https://arxiv.org/pdf/2102.04040.pdf (open for contribution)
• AdaSpeech1/2 https://arxiv.org/pdf/2104.09715 and https://arxiv.org/abs/2103.00993

[lucascassiano]

great project! Excited to see this growing!

[AndrewBarfield]

I'm learning the code/API and performing experiments. I hope to contribute soon.

I'm also wondering if I can donate (money) to Coqui?

[kdavis-coqui]

I'm learning the code/API and performing experiments. I hope to contribute soon.

I'm also wondering if I can donate (money) to Coqui?

Wow! Thanks! Humbling.

We were setting up GitHub sponsors, but the tax implications were onerous.

We're currently exploring Patreon. So stay tuned!

[agrinh]

@erogol Thanks for sharing the plans!

Do you have any thoughts (or need help to) simplifying the dependencies a bit? I'm thinking that if TTS is used as a lib installed over pip it might be nice to remove visualisation dependencies only used in notebooks, removing test/dev dependencies and moving e.g. tensorflow into extras to reduce the footprint. Personally would love to use this as a dependency rather than maintaining my own fork.

[erogol]

@agrinh Why do you need to keep your own fork exactly? It'd be better to expand the conversation on gitter if you like.

[agrinh]

@agrinh Why do you need to keep your own fork exactly? It'd be better to expand the conversation on gitter if you like.

Wow, thanks for the super fast reply. Sure, we can move the discussion to gitter.

[Sadam1195]

Please add DC-TTS to the the list of models.

DC-TTS implementation available with MIT Licence code available here
EFFICIENTLY TRAINABLE TEXT-TO-SPEECH SYSTEM BASED ON DEEP CONVOLUTIONAL NETWORKS WITH GUIDED ATTENTION paper
@erogol

[will-rice]

What were you thinking about the "TensorFlow run-time for training models"? Like giving the user the option of using TensorFlow or PyTorch? I wouldn't mind taking a stab at the TensorFlow part.

[erogol]

@will-rice the plan is to mirror what we have in torch to TF as much as possible. It'd be great if you initiate the work

[lucashueda]

Are you guys planning to develop some expressive TTS architectures? I'm currently studying this topic and planning to implement some of them based on Coqui, part of them just controlling latent space using GST Kwon et al 2020 or RE Sorin et al 2020, and others that actually changes the architecture by adding VAE, normalizing flows and gradient reversal

[a-froghyar]

@lucashueda Capacitron VAE: #510

[lucashueda]

@lucashueda Capacitron VAE: #510

Oh nice, hope to see Capacitron integrated soon. So maybe, in the future I'll be able to contribute with some others expressive architectures

[BillyBobQuebec]

@erogol Look forward to new End-to-End models being implemented, specfically Efficient-TTS! if the paper is accurate, it should blow most 2 stage configurations out of the water, considering it seems to have higher MOS than tacotron2+hifigan, while also seeming to have significantly faster speed than glowtts+fastest vocoder! I have not seen a single repo replicating the EFTS-Wav architecture described in the paper released 10 months ago, it would be amazing to see it in Coqui first!

[erogol]

@BillyBobQuebec I don't think I will implement these models anytime soon. But as they stand, contributions are welcome

[WeberJulian]

@BillyBobQuebec but you can try VITS which is close to what you're describing :)

[BillyBobQuebec]

@BillyBobQuebec but you can try VITS which is close to what you're describing :)

Agreed, I am currently trying VITS actually, I have some issues training with the coqui implementation unfortunately, I've posted the issue about the bug today and hope I can get it resolved.

[hemath1001]

Hi there! Thanks for your great work! I'm looking forward to training YourTTS on other languages. Will training and fine-tuning code of YourTTS be published soon? I would be very grateful if you could tell me an approximate time~ Have a nice day :-D

[nfaraji2002]

Hi
thanks for delightful codes!
I want to use this version of TTS on raspberry pi 4, but I think this version does not support real time processing.
Are there TF utilities provided as in Mozilla TTS to convert trained models to tf-lite?
Can the strategy of quantization work here for real-time processing?
I need some roadmaps in this regard.

Thanks
Neda

[jhj0517]

Thank you for your great work for TTS.

Is there any progress on Let the user pass a custom text cleaner function. ?
If it's possible, I want to pass my own Korean cleaners.

[erogol]

You can currently do it by creating your own tokenizer or overloading the class.

[stale]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions. You might also look our discussion channels.

[MaxIakovliev]

Marvelous project.
Any ways to donate to core contributors?
I would prefer to use paypal.

[stale]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions. You might also look our discussion channels.

[erogol]

@MaxIakovliev you can use https://coqui.ai/ :)

[erogol]

This roadmap issue is quite outdated. I'll keep it open to keep the references to some of the issues and models we like to tackle but won't be updating until one day officially becomes 48 hours.

[jmlcoliveira]

Any update regarding SSML implementation?

[erogol]

We are not working on SSML currently, it is back in the list without a precise timeline.

[offside609]

Please do!!

[stale]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions. You might also look our discussion channels.

[violet17]

Will you support bark-small? Thanks.

[csukuangfj]

Any plan to a port of coqui-ai engine for android? TTS on android is very robotic (espeak, rhvoice, festival lite).

@paolo-caroni

Please take a look at
#3194

You can use sherpa-onnx to run VITS models from Coqui on Android and also embedded devices, e.g., raspberry pi.

We have pre-built Android APKs for the VITS English models from Coqui.
https://k2-fsa.github.io/sherpa/onnx/tts/apk.html

[DmitryVN]

Fix it plz #3039 #3282
The problem persists and because of this, normal correct use is not possible. Also at the moment it kind of breaks off the phrase at the end of each sentence and it turns out a jerky reading.

[MarkChrisE2091]

Any new update?

[csukuangfj]

Any plan to a port of coqui-ai engine for android? TTS on android is very robotic (espeak, rhvoice, festival lite).

@paolo-caroni

We have supported it in k2-fsa/sherpa-onnx#508

The following is a YouTube video
https://www.youtube.com/watch?v=33QYuVzDORA

You can use all coqui-ai/TTS models and piper models listed in
https://github.com/k2-fsa/sherpa-onnx/releases/tag/tts-models
with k2-fsa/sherpa-onnx#508

[imevro]

hi guys, why?

upd: found https://twitter.com/_josh_meyer_/status/1742522906041635166

[NicoleKai]

Their ability to exist and be profitable was dependent on how much better their tech was compared to everyone else. It may not feel like it, but we are in the middle of an AI singularity. Coqui's business model might have stood a chance if they started with this tech 5 years earlier, but it was probably too little too late. Eleven labs is probably eating their lunch :/