Add AliMeeting multi-condition training recipe

[desh2608]

This is similar to the AMI ASR recipe, i.e., we train a single model by combining training data from different mic settings: IHM, IHM + reverb, SDM, and GSS-enhanced audio, and use this model to evaluate different conditions.

This recipe also includes text normalization similar to the M2MeT challenge baseline. Here are the results using --epoch 15 --avg 8 and modified beam search:

Evaluation set	eval CER	test CER
IHM	9.58	11.53
SDM	23.37	25.85
MDM (GSS-enhanced)	11.82	14.22

As a comparison, the eval set results for single-speaker ASR from the baseline paper are:

In the above table, the Eval-Ali-near is equivalent to IHM, and Eval-Ali-far is equivalent to SDM.

[csukuangfj]

Could you upload some test_waves to this huggingface repo?
It makes testing the model easier in sherpa.

Please have a look at
https://huggingface.co/csukuangfj/icefall-asr-librispeech-pruned-transducer-stateless8-2022-11-14/tree/main/test_wavs

[desh2608]

Done.

[desh2608]

If there are no further changes needed, perhaps this can be merged? (There is a failing test, but it seems unrelated to the changes in this PR.)

[csukuangfj]

Thanks!

[csukuangfj]

@desh2608 You can try your pre-trained model at https://huggingface.co/spaces/k2-fsa/automatic-speech-recognition

Thanks again for the contribution.

[csukuangfj]

By the way, I have added the model to
https://huggingface.co/spaces/k2-fsa/automatic-speech-recognition

[desh2608]

By the way, I have added the model to https://huggingface.co/spaces/k2-fsa/automatic-speech-recognition

Thanks! I'll check it out.

[csukuangfj]

@desh2608
Could you please merge
https://huggingface.co/desh2608/icefall-asr-alimeeting-pruned-transducer-stateless7/discussions/1

(Your test waves are 32-bit encoded using floats. I converted it to 16-bit encoded using int16 in the above PR).

[desh2608]

@desh2608 Could you please merge https://huggingface.co/desh2608/icefall-asr-alimeeting-pruned-transducer-stateless7/discussions/1

(Your test waves are 32-bit encoded using floats. I converted it to 16-bit encoded using int16 in the above PR).

Thanks. Done.