nvidia
/

nemotron-3.5-asr-streaming-0.6b

Model card Files Files and versions

Question about multilingual Nemotron 3.5 ASR training setup (tokenizer & training config)

#26

by HoussamLCHMT - opened 16 days ago

I'm trying to reproduce the multilingual Nemotron 3.5 ASR architecture and have a couple of questions regarding the training setup.

For the multilingual model, was a single shared tokenizer trained across all languages, or were multiple tokenizers aggregated/combined in some way?
Is the following FastConformer streaming configuration representative of the one used for Nemotron 3.5 Streaming, or were there significant modifications?

https://github.com/NVIDIA-NeMo/NeMo/blob/main/examples/asr/conf/fastconformer/cache_aware_streaming/fastconformer_transducer_bpe_streaming.yaml

I'm interested in training a similar multilingual model from scratch, so any guidance or pointers would be greatly appreciated.

NVIDIA org 15 days ago

Just one tokenizer.
We merged all langs tokenizers and removed duplicate tokens
here a tutorial https://github.com/nvidia-riva/tutorials/blob/main/asr-extend-tokenizer-to-newlang-ft-acoustic-model.ipynb

Upload images, audio, and videos by dragging in the text input, pasting, or clicking here.

Tap or paste here to upload images

· Sign up or log in to comment