[WIP] Fix MMI recipe

[pkufool]

trafficstars

This PR try to fix https://github.com/k2-fsa/icefall/issues/685 and to tune a better result for MMI models.

• [x] Make current model converge.
• [ ] Tune a better result.
• [ ] Apply Zipformer to MMI recipe.

[pkufool]

The results I have for libri-100.

model	ctc-decoding	HLG 1best	HLG + 4-gram rescoring	HLG + 4-gram rescoring + attention decoder
conformer-mmi (with attention decoder)	95.68 & 122.38	6.15 & 17.58	5.6 & 16.02 (lm-scale=1.3)	5.37 & 16.28 (lm-scale=2.0; att-scale=1.0)	--epoch 29 --avg 12

The ctc-decoding results is weird, I haven't figured out the issue.

Training command:

./conformer_mmi/train-with-attention.py \
  --exp-dir conformer_mmi/exp_100_att \
  --lang-dir data/lang_bpe_500 \
  --full-libri 0 \
  --max-duration 200 \
  --world-size 2 \
  --start-epoch 0 \
  --num-epochs 30 \
  --master-port 35673

[dan-legit]

In my opinion, bad ctc-decoding result is reasonable for MMI model. I think your decoding results with HLG look good now.

Btw, does & sign in your table mean both MMI only and MMI with attention decoder? I'm wondering because I just only tried MMI model, not MMI with attention decoder.

[pkufool]

Btw, does & sign in your table mean both MMI only and MMI with attention decoder? I'm wondering because I just only tried MMI model, not MMI with attention decoder.

No, & means test and test-other. I am running MMI without attention decoder, will update the results here soon.