search

k2-fsa / icefall

https://k2-fsa.github.io/icefall/
Apache License 2.0
802 stars 270 forks source link

[WIP] Add phone based train and decode for gigaspeech #1023

Open yfyeung opened 1 year ago

wangtiance commented 1 year ago

Hi, have you got any results with phone based models? I previously tried with librispeech and the result was worse than BPE. For pruned transducer I only got 4-5 WER for test-clean.

yfyeung commented 1 year ago

Hi, have you got any results with phone based models? I previously tried with librispeech and the result was worse than BPE. For pruned transducer I only got 4-5 WER for test-clean.

I tried the pruned transducer on gigaspeech M, and the result was worse than BPE too.

unit level dev & test lm ngram-lm-scale ppl checkpoint 
phone 76 13.15 & 13.46 3gram_pruned_1e8 0.235 192.176 & 213.068 epoch 30 avg 7 
bpe 500 12.88 & 12.87 -  - -  epoch 30 avg 8
wangtiance commented 1 year ago

Thanks! But your result seems very close. Will try your recipe on librispeech sometime.

yfyeung commented 1 year ago

Maybe sometime later. Not recently.

pingfengluo commented 1 year ago

@yfyeung can you update and merge this PR?