Quality suffers on earnings22 dataset

[soupslurpr]

whisper-tiny.en gets 18 WER without dynamic audio context on https://huggingface.co/datasets/distil-whisper/earnings22 (chunked, test) using evaluation.ipynb while acft-whisper-tiny.en with dynamic audio context gets 318 WER. This indicates that the acft fine tuned model with dynamic audio context may not work well in real-world conditions which include diverse accents and varying speech conditions.

[soupslurpr]

Not sure why but changing ADD_AUDIO_CTX to 64 makes acft-whisper-tiny.en achieve 19 WER on earnings22.

[abb128]

For production deployment you should use additional context of at least 32, I did some tests here with different values showing that even with librispeech the WER improves with slight added context. It's possible that earnings22 particularly demonstrates weakness when there is not enough extra silence