Why you should move from DeepSpeech to coqui.ai

Archives DeepSpeech
#1

Don’t get me wrong, I love DeepSpeech and we use it in production for clients. But the codebase isn’t really maintained any longer and most of the developers formed a new company with the same license and great service. So, nothing any longer to see here and move on to:

They have a good newsletter too, and for support, head over to their gitter:

There was talk about some grants for DeepSpeech a couple months ago but nothing materialized, so I guess DeepSpeech is officially dead … and I have no connection to coqui except for using their great product :slight_smile:

4 Likes
Training DeepSpeech on gpu failed
Creating DeepSpeech Model for Hindi
How to choose number of hidden layers?
Using CTC decoder with custom logits on Android
Make your own simple model
I got stuck while running training
MozillaTTS: RuntimeError: size mismatch for embedding.weight
Always get stuck after 1 Training Epoch when I do transfer learning
Deepspeech training problem: DeepSpeech.py execution gives Illegal instruction (core dumped)
Error when fine tuning DeepSpeech with custom dataset
Getting error while working with API
Errors with using transcribe.py
Any way to change the weight matrix?
How to create a good model with small dataset?
DeepSpeech update, grant and playbook
Dependency conflict ds_ctcdecoder and tensorflow
How to train ASR model for Isolated words (not sentences)
Web_microphone_websocket demo work fine and not very well on android_mic_streaming demo
#2

I agree. I would love to see an official collaboration between Coqui and Mozilla, it would help both projects a lot and decrease the confusion. @heyhillary are there any news on this situation? Will deepspeech be further supported? If not, giving the name “deepspeech” to coqui would be something worth considering. Don’t make the same mistake that open office has done :wink:

@othiele Now that the new Common Voice dataset is available, what would you suggests? In my understanding version 0.9.3 of Deepspeech are identical to 0.9.3 of Coqui STT and both models are compatible. But this will likely change in the future. So should we train modells with the Coqui repo, or does is matter at all right now?

1 Like
#3

@stergro, currently there is not much going on in the codebase, therefore I don’t think the Mozilla repo is good for production use. Coqui on the other hand just got funding and will develop STT with some fulltime devs. Therefore I won’t use any future versions of the Moz code , if there are any. If I remember correctly, the name “DeepSpeech” is problematic, therefore coqui got STT and TTS for pypi.

New models are better trained with coqui as you can use all the current features plus everything they provide on top. Backwards compatibility will be gone at some point though. And don’t forget wav2vec, as I see it there are more CommonVoice (CV) models trained with with wav2vec than Moz/coqui currently :slight_smile: So CV is really important to the community even if they don’t train with Deepspeech. So long, and thanks for all the fish

2 Likes