SOTAVerified

SpeechStew: Simply Mix All Available Speech Recognition Data to Train One Large Neural Network

2021-04-05Unverified0· sign in to hype

William Chan, Daniel Park, Chris Lee, Yu Zhang, Quoc Le, Mohammad Norouzi

Unverified — Be the first to reproduce this paper.

Reproduce

Abstract

We present SpeechStew, a speech recognition model that is trained on a combination of various publicly available speech recognition datasets: AMI, Broadcast News, Common Voice, LibriSpeech, Switchboard/Fisher, Tedlium, and Wall Street Journal. SpeechStew simply mixes all of these datasets together, without any special re-weighting or re-balancing of the datasets. SpeechStew achieves SoTA or near SoTA results across a variety of tasks, without the use of an external language model. Our results include 9.0\% WER on AMI-IHM, 4.7\% WER on Switchboard, 8.3\% WER on CallHome, and 1.3\% on WSJ, which significantly outperforms prior work with strong external language models. We also demonstrate that SpeechStew learns powerful transfer learning representations. We fine-tune SpeechStew on a noisy low resource speech dataset, CHiME-6. We achieve 38.9\% WER without a language model, which compares to 38.6\% WER to a strong HMM baseline with a language model.

Tasks

Benchmark Results

DatasetModelMetricClaimedVerifiedStatus
AMI IMHSpeechStew (100M)Word Error Rate (WER)9Unverified
AMI SDM1SpeechStew (100M)Word Error Rate (WER)21.7Unverified
CHiME-6 dev_gss12SpeechStew (1B)Word Error Rate (WER)31.9Unverified
CHiME-6 evalSpeechStew (1B)Word Error Rate (WER)38.9Unverified
Common VoiceSpeechStew (1B)Test WER10.8Unverified
LibriSpeech test-cleanSpeechStew (1B)Word Error Rate (WER)1.7Unverified
LibriSpeech test-cleanSpeechStew (100M)Word Error Rate (WER)2Unverified
LibriSpeech test-otherSpeechStew (1B)Word Error Rate (WER)3.3Unverified
LibriSpeech test-otherSpeechStew (100M)Word Error Rate (WER)4Unverified
Switchboard CallHomeSpeechStew (100M)Word Error Rate (WER)8.3Unverified
Switchboard SWBDSpeechStew (100M)Word Error Rate (WER)4.7Unverified
TedliumSpeechStew (100M)Word Error Rate (WER)5.3Unverified
WSJ eval92Speechstew 100MWord Error Rate (WER)1.3Unverified

Reproductions