Zipformer: A faster and better encoder for automatic speech recognition

Benchmark Model Rank Results
speech-recognition-on-amisoundsgoodai/Zipformer-transducer-XL-290M#24Word Error Rate (WER): 13.96
speech-recognition-on-ami-cleanedsoundsgoodai/Zipformer-transducer-XL-290M#28Word Error Rate (WER): 12.8
speech-recognition-on-earnings-22soundsgoodai/Zipformer-transducer-XL-290M#7Word Error Rate (WER): 9.02
speech-recognition-on-gigaspeech-cleanedsoundsgoodai/Zipformer-transducer-XL-290M#43Word Error Rate (WER): 9.21
speech-recognition-on-gigaspeech-testsoundsgoodai/Zipformer-transducer-XL-290M#41Word Error Rate (WER): 9.31
speech-recognition-on-librispeech-test-cleansoundsgoodai/Zipformer-transducer-XL-290M#48Word Error Rate (WER): 1.67
speech-recognition-on-librispeech-test-cleanZipformer+pruned transducer (no external language model)#70Word Error Rate (WER): 2.00
speech-recognition-on-librispeech-test-otherZipformer+pruned transducer (no external language model)#70Word Error Rate (WER): 4.38
speech-recognition-on-librispeech-test-othersoundsgoodai/Zipformer-transducer-XL-290M#73Word Error Rate (WER): 4.54
speech-recognition-on-spgispeechsoundsgoodai/Zipformer-transducer-XL-290M#7Word Error Rate (WER): 1.91
speech-recognition-on-voxpopulisoundsgoodai/Zipformer-transducer-XL-290M#27Word Error Rate (WER): 6.56
speech-recognition-on-voxpopuli-cleanedsoundsgoodai/Zipformer-transducer-XL-290M#35Word Error Rate (WER): 4.08
speech-recognition-on-wenetspeechZipformer+pruned transducer (no external language model)#3Character Error Rate (CER): 7.29