README.md

Performance Record

This is a Chinese speech recognition recipe that trains on all Chinese corpora including:

Dataset	Duration (Hours)
Aidatatang	140
Aishell	151
MagicData	712
Primewords	99
ST-CMDS	110
THCHS-30	26
TAL-ASR	587
AISHELL2	1000

Unified Transformer Result

Data info:

• Dataset: Aidatatang, Aishell, MagicData, Primewords, ST-CMDS, and THCHS-30.
• Feature info: using fbank feature, with cmvn, no speed perturb.
• Training info: lr 0.004, batch size 18, 3 machines, 3*8 = 24 GPUs, acc_grad 1, 220 epochs, dither 0.1
• Decoding info: ctc_weight 0.5, average_num 30
• Git hash: 013794572a55c7d0dbea23a66106ccf3e5d3b8d4

WER

Dataset	chunk size	attention decoder	ctc greedy search	ctc prefix beam search	attention rescoring
Aidatatang	full	4.23	5.82	5.82	4.71
	16	4.59	6.99	6.99	5.29
Aishell	full	4.69	5.80	5.80	4.64
	16	4.97	6.75	6.75	5.37
MagicData	full	2.86	4.01	4.00	3.07
	16	3.10	5.02	5.02	3.68
THCHS-30	full	16.68	15.46	15.46	14.38
	16	17.47	16.81	16.82	15.63

Unified Conformer Result

Data info:

• Dataset: Aidatatang, Aishell, MagicData, Primewords, ST-CMDS, and THCHS-30.
• Feature info: using fbank feature, with cmvn, speed perturb.
• Training info: lr 0.001, batch size 8, 1 machines, 1*8 = 8 GPUs, acc_grad 12, 60 epochs
• Decoding info: ctc_weight 0.5, average_num 10
• Git hash: 5bdf436e671ef4c696d1b039f29cc33109e072fa

WER

Dataset	chunk size	attention decoder	ctc greedy search	ctc prefix beam search	attention rescoring
Aidatatang	full	4.12	4.97	4.97	4.22
	16	4.45	5.73	5.73	4.75
Aishell	full	4.49	5.07	5.05	4.43
	16	4.77	5.77	5.77	4.85
MagicData	full	2.55	3.07	3.05	2.59
	16	2.81	3.88	3.86	3.08
THCHS-30	full	13.55	13.75	13.76	12.72
	16	13.78	15.10	15.08	13.90

Unified Conformer Result

Data info:

• Dataset: Aidatatang, Aishell, MagicData, Primewords, ST-CMDS, THCHS-30, TAL-ASR, and AISHELL2.
• Feature info: using fbank feature, dither=0, cmvn, speed perturb
• Training info: lr 0.001, batch size 22, 4 GPUs, acc_grad 4, 120 epochs, dither 0.1
• Decoding info: ctc_weight 0.5, average_num 10
• Git hash: 66f30c197d00c59fdeda3bc8ada801f867b73f78

WER

Dataset	chunk size	attention decoder	ctc greedy search	ctc prefix beam search	attention rescoring
Aidatatang	full	3.22	4.00	4.01	3.35
	16	3.50	4.63	4.63	3.79
Aishell	full	1.23	2.12	2.13	1.42
	16	1.33	2.72	2.72	1.72
MagicData	full	2.38	3.07	3.05	2.52
	16	2.66	3.80	3.78	2.94
THCHS-30	full	9.93	11.07	11.06	10.16
	16	10.28	11.85	11.85	10.81
AISHELL2	full	5.25	5.81	5.79	5.22
	16	5.48	6.48	6.50	5.61
TAL-ASR	full	9.54	10.35	10.28	9.66
	16	10.04	11.43	11.39	10.55