Bug

[ejkitchen]

When running on audio files with voices, your model is impressive, great work. But using the medium English model on audio with no voices and low ambient background noise, I sometimes get "Thank you for watching!". I listened to the audio, and there is absolutely nothing in there that could even be construed as that. The audio file was only 2 min long. Just an FYI

[jongwook]

This is an example of hallucination mentioned in the model card. This can be mitigated by tuning the VAD hyperparameters (--compression_ratio_threshold, --logprob_threshold, --no_speech_threshold), and we are looking for ways to further prevent hallucination. Meanwhile, combining with an external VAD model seems to be a good option as well #29.