GigaSpeech

A Large, modern and evolving dataset for automatic speech recognition. More details about GigaSpeech can be found: https://github.com/SpeechColab/GigaSpeech

Performance Record

Conformer Result

• Feature info: using fbank feature, dither 1.0
• Training info: train_conformer.yaml, kernel size 31, lr 0.002, batch size 48, 8 gpu, acc_grad 4, 15 epochs, use amp
• Decoding info: ctc_weight 0.5, average_num 3
• Git hash: 11aaea7195ffe1014a4bbd16a1fb47ec7dbee2ac
• Model link: http://mobvoi-speech-public.ufile.ucloud.cn/public/wenet/gigaspeech/20210520_conformer_exp.tar.gz

gigaspeech scoring

the raw scoring version wer