Training Fastconformer-CTC on another language.

[bharathraj-v]

Hi,
I'm new to NeMo and looking to train the Fastconformer-CTC model on Telugu datasets. Telugu is an Indian language which is highly inflectional and somewhat agglutinative. I have around 800 hours of data, what's the best approach to achieve good results here? Do I fine-tune a pre-trained model or train one from scratch?

Also, what's the config that's best suited for this case? And, do I go for subword tokenization for a language such as this or does the standard bpe tokenizer provided in the scripts suffice? Any guidance regarding this would be of great help

Thanks!

[titu1994]

@nithinraok wrote this post for a Telegu speech model here - https://blogs.nvidia.com/blog/speech-ai-telugu-language-breakthrough/

He could provide you with specific advice

[titu1994]

Overall yes, use subwords (spe unigram or just spe bpe). Config - Fast Conformer large - try to finetune and not train from scratch, it will always lead to better results. The ASR CTC finetune tutorial has most of the tricks for fine-tuning that @nithinraok applied

[bharathraj-v]

Thanks!

[nithinraok]

Yes as @titu1994 mentioned, create a spe bpe based tokenizer from telugu words and finetune a XL/L En ASR model.
L: https://huggingface.co/nvidia/stt_en_fastconformer_ctc_large
XL: https://huggingface.co/nvidia/stt_en_fastconformer_ctc_xlarge

[bharathraj-v]

Okay, thanks! So for Telugu's spe bpe tokenizer, what's the ideal vocab size I should go for? @titu1994 @nithinraok

[nithinraok]

You may experiment with various sizes (256,512, 1024). 256 should be good enough based on our prior experiments.