TEVR: Improving Speech Recognition by Token Entropy Variance Reduction

Benchmark Model Rank Results
speech-recognition-on-common-voice-germanwav2vec 2.0 XLS-R 1B + TEVR (5-gram)#1Test WER: 3.64%Test CER: 1.54%
speech-recognition-on-common-voice-germanwav2vec 2.0 XLS-R 1B + TEVR (4-gram)#2Test WER: 3.70%
speech-recognition-on-common-voice-germanwav2vec 2.0 XLS-R 1B (5-gram)#5Test WER: 4.38%Test CER: 1.62%
speech-recognition-on-common-voice-germanwav2vec 2.0 XLS-R 1B + TEVR (no LM)#13Test WER: 10.10%
speech-recognition-on-common-voice-germanwav2vec 2.0 XLS-R (no LM)#14Test WER: 12.06%