Modelpublic

shaojieli/icefall-asr-commonvoice-fr-pruned-transducer-stateless7-streaming-2023-04-02

sourceHugging Faceapache-2.0updated 3y agoView on Hugging Face

3likes

Model Card

training on full commonvoice

The WERs are:

decoding method	chunk size	test	comment	decoding mode
greedy search	640ms	10.90	--epoch 30 --avg 9	simulated streaming
modified beam search	640ms	10.55	--epoch 30 --avg 9	simulated streaming
fast beam search	640ms	10.75	--epoch 30 --avg 9	simulated streaming

training on full librispeech then finetune on full commonvoice

The WERs are:

decoding method	chunk size	test	comment	decoding mode
greedy search	640ms	10.57	--epoch 29 --avg 9	simulated streaming
modified beam search	640ms	10.19	--epoch 29 --avg 9	simulated streaming
fast beam search	640ms	10.25	--epoch 29 --avg 9	simulated streaming

training on full librispeech and gigaspeech then finetune on full commonvoice

The WERs are:

decoding method	chunk size	test	comment	decoding mode
greedy search	640ms	9.95	--epoch 30 --avg 9	simulated streaming
modified beam search	640ms	9.57	--epoch 30 --avg 9	simulated streaming
fast beam search	640ms	9.67	--epoch 30 --avg 9	simulated streaming