search

wilrop / DuStt-Demo

A demo for DuSst - Dutch Speech to Text
3 stars 1 forks source link

Are the trained DeepSpeech models available somewhere? #4

Open dietercastel opened 3 years ago

dietercastel commented 3 years ago

Hi Willem,

I just stumbled upon the Demo of DuStt. Very interesting! But since I couldn't find the trained models (and suspect you had them locally as per:) https://github.com/wilrop/DuStt-Demo/blob/a0956acbfe70c51d160a4b25b649d2b3ab30879c/stt4dutch/inference/functions.py#L5 I was wondering whether you still have them somewhere and if so if I could try them out some time?

If licenses permit it of course, I think they would also be welcomed in this DeepSpeech topic: https://discourse.mozilla.org/t/links-to-pretrained-models/62688

Finally I was also wondering whether there has been any follow up research of this project?

Looking forward to reading from you! kr, Dieter

wilrop commented 3 years ago

Hi Dieter,

My apologies for not replying faster, for some reason I don't get notifications reliably on GitHub.

I indeed had the models locally. I am not entirely sure if I am able to share the models as the data was licensed to me with an academic license. I also believe that the VUB actually owns the actual models as I made them during my bachelor thesis (this particular point I find quite weird but it is what I was told). You can however easily create your own models! You can use the code in Import-CGN https://github.com/wilrop/Import-CGN to preprocess the Corpus Gesproken Nederlands and use these audio files to train models on the latest deepspeech version. I have had the best performance by using a pre-trained English model and finetuning that on my Dutch audio files. Good luck!

Willem

Edit: I just saw I didn't answer the last part of your question. So far, there has not been any follow-up research. However, that might change in the future depending on the specific circumstances.

justinhunt commented 1 year ago

Do you think that you could just ask for permission from VUB to submit your trained models to? https://github.com/coqui-ai/STT-models

There is only one other Dutch model and its tiny. Most of us, actually dont have the skills to build our own language models. I think there is no issue with releasing speech models trained on licensed data, as long as you didn't already agree not to.

wilrop commented 1 year ago

Unfortunately, I really cannot do that. In addition, the performance of my language model will most likely be genuinely horrible compared to what modern models are able to do.