Feature request - Finetuning or Pretraining for Urdu

[hunzlausman]

Please tell me if I can fine-tune this model on urdu dataset or train from scratch with same architecture for Urdu??
In short please elaborate the architecture of this model and how it is so fast.

[snakers4]

This year we will probably share a fine-tuning recipe

[hunzlausman]

Thank you so much for the reply Alexander.Can you please tell me the architecture it is based on or how you achieved this perfection in speed and accuracy both at the same time?And also the expected time in days or months for a new release you're talking about as I'm greatly impressed by and looking for this over last few months so can't wait more.If you elaborate the architecture, I can contribute alot in your project.

…

On Wed, 17 Jan 2024 at 20:34, Alexander Veysov ***@***.***> wrote: This year we will probably share a fine-tuning recipe — Reply to this email directly, view it on GitHub <#414 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/A2UAMPXLS7OOK5XK5L4UC23YO7VORAVCNFSM6AAAAABB6XMMSSVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMYTQOJWGA2TSNRSHA> . You are receiving this because you authored the thread.Message ID: ***@***.***>

[filtercodes]

Any news on this??? For me this model turned out very much useless without fine-tuning, because I get too much false positives and that makes the system unstable. Will stick to other options until the fine-tuning recipe is available.

[hunzlausman]

No updates yet. Exactly, any model like this is worthless without having control of finetuning or modifications.

…

On Sun, 17 Mar 2024, 19:38 filtercodes, ***@***.***> wrote: Any news on this??? For me this model turned out very much useless without fine-tuning, because I get too much false positives and that makes the system unstable. Will stick to other options until the fine-tuning recipe is available. — Reply to this email directly, view it on GitHub <#414 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/A2UAMPX47N3KBF67T72ABMTYYWTG3AVCNFSM6AAAAABB6XMMSSVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDAMBSGQ4TGOBZHA> . You are receiving this because you authored the thread.Message ID: ***@***.***>

[snakers4]

As a first step - we released the dataset - https://github.com/snakers4/silero-vad/tree/master/datasets

[markjosims]

This is excting news, I'm looking forward to the finetuning scripts being released

[snakers4]

The new VAD version was released just now - #2 (comment)

It supports more than 6,000 languages now

Fine-tuning code will be released soon

[hunzlausman]

Please we need for speech to text not vad.

[filtercodes]

The new VAD version was released just now - #2 (comment)

It supports more than 6,000 languages now

Fine-tuning code will be released soon

Looking forward to try it out! Is the state in/out basically combined c and h? I see it's tensor: float32[2,?,128], while c and h were float32[2,batch,64]) each.

[filtercodes]

Yes, it doesn't work with the python script. I'm having this error:

[varrerohit]

Do we have any update on the finetuning version?