Training dataset & LID head

#3
by Champion - opened

Hi, impressive work!
I’m baffled by how much better this model is than the original Nvidia-trained one.
Thanks for sharing it so openly.
Could you share some stats about the training data you used for this post-training? A per-language table with training hours would be enough.

Also, I like the addition of the VAD head, but what’s really missing, IMO, is a language identification head!
The model has to know which language it’s decoding each word in, but there’s no easy way to tell after the fact, and I think having it embedded in an ASR model makes a lot of sense! (ATM I'm using lingua, but it's not so precise when dealing with short mixed sentences)

Sign up or log in to comment