New model contribution following #380

[erogol]

Hi @Sadam1195 I just wanted to continue our #380 talk here.

Do you have a particular language that you like to attack?

[Sadam1195]

Hi, I would like to start with Urdu, Hindi, Sindhi, Arabic, Persian, French, Portuguese, Spanish, Romanian, Catalan, Romansh. I can work on any of them but if it was up to me I would love to begin with Urdu or Sindhi.

[astricks]

I tried many techniques - 2 of them worked, one better than the other.

For a lower quality dataset:

1. I downloaded over 25 hours of a particular podcast and wrote a python script to segment it on silences - the result was 11,500 audio clips.
2. I then used google speech-to-text on GCP to transcribe each audio clip. It cost me $300 to get it all transcribed. It isn't 100% perfect, but the model aligns and it actually works.

For a higher quality dataset:

1. I basically created a dataset of thousands of lines and paid someone $2000 to record each line in their home studio.
2. To create a dataset of thousands of lines, I tried many sources - the only one that really worked for me was downloading subtitles for TV shows and using python scripts to tie them together to get lines of varying length.

Let me know if I can elaborate on any specific step.

[Sadam1195]

Hmm, interesting. Though GCP doesn't really works good, it removes any stutters like hmm, umm, eh and also it very robust to noisy speech and often in reality you don't get good quality audios so wouldn't work with the datasets resources I am looking for.

Second options seems a bit enterprise to me and not many can afford that luxury.

I will share my steps when I create my dataset.

Thanks for the details.

[SaadBazaz]

@Sadam1195 Were you able to make it for Urdu and Sindhi? Did you check out CommonVoice's Urdu dataset? (https://commonvoice.mozilla.org/en/datasets)

[SaadBazaz]

@astricks Do you have the Hindi model? Have you shared it online anywhere?

[Sadam1195]

@Sadam1195 Were you able to make it for Urdu and Sindhi? Did you check out CommonVoice's Urdu dataset? (https://commonvoice.mozilla.org/en/datasets)

Hi Saad, I did worked on Urdu and Sindhi models. Unfortunately, I couldn't get permission to make it open source as it was paid work and I had a NDA signed. I trained Tacotron2 , FastSpeech and multiple vocoders to see which worked best. Combination with few of them worked very nicely but as I said I couldn't make it public, I was disheartened to see my efforts go futile as my sole motivation for committing long term to the project was to make it open source as a result I abandoned this community of my future engagements.
@SaadBazaz

As far as I know of @astricks he did shared his results and results were decent I'd say but he was working on commercial product from start so he did not made his model public. Maybe he can further clear it up here.