SadeghK/stt_nemo_fastconformer_transducer_hybrid_cache_aware_streaming

STT for Persian ASR, with cache enabled

Here is the WER convergence for both CTC and Transducer.

Model

1040ms lookahead settings

Training

Training without tarred and bucketing. Training in different stages, first with some higher LR values, and later decrease in plateues.

Performance

WER: 2.16% CTC WER: 3.22%