Training ASR models in multiple languages

[monatis]

TensorFlowASR makes it quite easy to train and deploy almost SOTA ASR models, but it provides a pretrained model only in English. On the other hand, FAIR has recently published an open and free dataset in 8 languages (see the paper). It is in the public domain and of a large size, and has the same quality as LibriSpeech. So, my suggestion is form a volunteer working group to collaborate on training ASR models in multiple languages and share them publically.

Maintainers of the repo can pin the issue and label it with help-wanted for visibility if this idea makes sense.

[nglehuy]

This is the great idea 😆

[monatis]

Hi @usimarit! Thanks for such a great project and your support to boost visibility of this issue 😻 So I will start by writing a helper script to automatically download MLS dataset in a given language and prepare the transcription and alphabet files and PR to add it to the repo. Then I can train a Conformer model in German by using this script as a first step 🚀

[nglehuy]

Hi @monatis, just for your information, we should train using subwords instead of characters for performance boost

[monatis]

Hi @usimarit, yeah I know that training with subwords yields a better performance, but I'm automatically generating an alphabet file in #92 for those who want to use characters anyway.

[nglehuy]

Hi everyone, if you guys want to share your pretrained models, just upload the .h5 or .pb files along with config.yml to any drive (google, dropbox, etc.), then write a section, add the link, add your contact to a subsection in that section, in the README.md in the example directory that belongs to each model like in image (the one I made):

And finally open a pull request to merge to the repo 😄

[monatis]

@usimarit Grat. I've been quite busy for some time, but I'll be more active in this repo on following days and contribute pretrained models in other languages. Thanks

[JStumpp]

@monatis, did you already train a Conformer model in German?

[monatis]

Hi @JStumpp I started to train it and hope to release it next week.

[christina284]

Do you know in which subset of Librispeech is the english pretrained model trained on?