Input Length Matters: Improving RNN-T and MWER Training for Long-form Telephony Speech Recognition
This addresses the problem of inaccurate long-form speech recognition for telephony applications, but it is incremental as it builds on existing methods with a focus on training data length.
The paper tackled the poor performance of end-to-end speech recognition models on long-form telephony audio by studying the effect of training utterance length on word error rate (WER) for RNN-T models, finding that increasing training length reduces WER by 15.7% for log loss and 8.8% for MWER loss.
End-to-end models have achieved state-of-the-art results on several automatic speech recognition tasks. However, they perform poorly when evaluated on long-form data, e.g., minutes long conversational telephony audio. One reason the model fails on long-form speech is that it has only seen short utterances during training. In this paper we study the effect of training utterance length on the word error rate (WER) for RNN-transducer (RNN-T) model. We compare two widely used training objectives, log loss (or RNN-T loss) and minimum word error rate (MWER) loss. We conduct experiments on telephony datasets in four languages. Our experiments show that for both losses, the WER on long-form speech reduces substantially as the training utterance length increases. The average relative WER gain is 15.7% for log loss and 8.8% for MWER loss. When training on short utterances, MWER loss leads to a lower WER than the log loss. Such difference between the two losses diminishes when the input length increases.