different results with fine-tuned model

[ramil-gadirov]

Hi,

I fune-tuned a new az language based on stt_en_fastconformer_hybrid_large_pc model (100h data, wer 7.4%, 350k steps), when transcribing a normal speech (microphone), audio books or movie, etc, a result is very good, but when trying a phone speech (good quality, no noise) getting very poor results from it (all audio 16khz/mono/16bit). Is there any params to change for training or maybe another nero model is more accurate for phone records?

[nithinraok]

latest model we have https://huggingface.co/nvidia/parakeet-tdt-0.6b-v3