Update Zipformer-xl 700M Results on multi-hans-zh

[yuekaizhang]

Update training results using 14k hours opensource Chinese data with a 750M zipformer.
The model has got current open model's SOTA for wenetspeech test_meeting set with 5.85% WER.

Model	yuekai/icefall-asr-multi-zh-hans-zipformer-large	yuekai/icefall-asr-multi-zh-hans-zipformer-xl
Config	Transducer Greedy Offline	Transducer Greedy Offline (blank_penalty 0.7)
aishell-1 test	1.38	1.31
aishell-2 test	3.23	3.27
aishell-4 test	15.36	14.64
WenetSpeech test_meeting	6.26	5.85
WenetSpeech tes_net	7.07	6.89

( The model reused whisper 80 dims fbank features from previous experiments. )

[marcoyang1998]

Hi Yuekai,

Nice results! Do you think blank penalty will also help with other decoding methods, e.g. modified_beam_search?

[yuekaizhang]

Hi Yuekai,

Nice results! Do you think blank penalty will also help with other decoding methods, e.g. modified_beam_search?

@marcoyang1998
Modified-beam-search has not been decoded yet, and perhaps even if it helps, the optimal penalty score might still change.

BTW, I did not use blank penalty for the 140M zipformer-large model. The reason I used it for the 700M model is that I found that as the number of training epochs increased, the deletion errors of the 700M model became larger, while the substitution errors consistently decreased (a total of 20 epochs were trained, and this phenomenon started to appear around the 10th epoch).

For greedy search, I also have some logs on tuning the blank penalty, which can be found here: https://huggingface.co/yuekai/icefall-asr-multi-zh-hans-zipformer-xl/tree/main/exp-xl/greedy_search/tmp. I found that a value of 0.7 worked best.