STT 1.4.0 - CommonVoice 12

[wasertech]

What about we keep commonvoice-fr/DeeepSpeech/ as it is and simply add commonvoice-fr/STT/ instead?

I've started from my branch stt140-cv9 which produced my latest models so I know it works. I updated it to use CV12 but that's basically it.

There are still some points I want to address before I train my next models so that's why it's still a draft but let me know if you already want to merge this new directory and let me update it later.

It's now much easier to compare the differences between DeepSpeech and STT.

❯ diff --recursive DeepSpeech STT

Checkout deepspeech-stt.diff

[wasertech]

I would still like to:

• Fix best_dev_checkpoint path #167
  Fix checkpoints.
• Expérimenter l'utilisation de DSAlign avec le modèle actuel #82
  Align new data for training.
• Base on upstream docker #131
  We can use coqui's image as our base one.
• And more
  Noise overlays, paraphrase et translate tons of sentences for the scorer, replace KenLM with a transformer and so on.

But we can do those in other PRs.

[lissyx]

fixme? :)

[wasertech]

This is your patch ahahha :) I have not touched it.
I'm not even sure why it's there to begin with :')

[wasertech]

Oh no it was added with #121
f62b21b

[wasertech]

I updated my patch to reflect commit 13d9316

[wasertech]

To reflect them all...

[drzraf]

What happened with this? That would be really nice to get a new model based on Coqui and the latest CV release (it's the way to go isn't?)

[lissyx]

Coqui is defunct, and I guess @wasertech stopped working on his model as well? I agree it would be fine but ...

[wasertech]

Well, I just moved on to other (more maintained) architectures, as I didn't wanted to maintain it alone. I tried Wav2Vec 2.0 for a while and currently, I am using Faster-Whisper for real-time transcription because those architecture are better suited for my use. I have a relatively big language model (~7B parameters) behind my transcription services so I need a large enough model to use for real-time open domain conversations. DeepSpeech/STT are just not cutting it for this use case unfortunately.

The code in my branch should still (mostly) work with the latest release of STT (1.4.0), so you (or anyone interested) can still use it to train or fine-tune models, but yeah it's unlikely for me to produce a new STT model given the availability of better suited architectures for my application.