Performance Record

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

• Aidatatang (140 hours)
• Aishell (151 hours)
• MagicData (712 hours)
• Primewords (99 hours)
• ST-CMDS (110 hours)
• THCHS-30 (26 hours)
• optional AISHELL2 (~1000 hours) if available
• optional TAL ASR (~600 hours) if available

Unified Transformer Result

data info:

• Aidatatang (140 hours)

• Aishell (151 hours)

• MagicData (712 hours)

• Primewords (99 hours)

• ST-CMDS (110 hours)

• THCHS-30 (26 hours)

• Feature info: using fbank feature, with cmvn, no speed perturb.

• Training info: lr 0.004, batch size 18, 3 machines, 3*8 = 24 gpu, acc_grad 1, 220 epochs, dither 0.1

• Decoding info: ctc_weight 0.5, average_num 30

• Git hash: 013794572a55c7d0dbea23a66106ccf3e5d3b8d4

• Model link: http://mobvoi-speech-public.ufile.ucloud.cn/public/wenet/multi_cn/20210315_unified_transformer_exp.tar.gz

aishell results

aidatatang results

thcs30 results

magic results

Unified Conformer Result

data info:

• Aidatatang (140 hours)

• Aishell (151 hours)

• MagicData (712 hours)

• Primewords (99 hours)

• ST-CMDS (110 hours)

• THCHS-30 (26 hours)

• AISHELL2 (~1000 hours)

• TAL ASR (~600 hours)

• Feature info: using fbank feature, dither=0, cmvn, speed perturb

• Training info: lr 0.001, batch size 12, 16 gpu, acc_grad 4, 48 epochs, dither 0.1

• Decoding info: ctc_weight 0.5, average_num 10

• Git hash: 18a75244f39b3403ff7b39f791e7af4fb93d4d03

• Model link:

aishell results

aidatatang results

thcs30 results

magic results

aishell-2 results

tal results