List your company
ResearchFormerly WizWand

Speech Recognition on Wall Street Journal open vocabulary (dev93)

4.4WERDeep Speech 2
3.9846.7929.612.408Dec 17, 2018Jan 31, 2019Mar 18, 2019May 3, 2019Jun 18, 2019Aug 3, 2019Sep 18, 2019
Updated 8mo ago

Evaluation results

MethodLinks
Deep Speech 22019.094.4—
Deep Speech 2labeled speech=12K h2019.044.42—
Supervised transfer-learningphoneme-based=true2019.044.99—
DeepSpeech 2Training hours=12k, Acoustic Model (AM)=true, LM=common crawl2018.125—
wav2vec largeLanguage Model=CHAR CONVLM, Pre-training Data=Librispeech, Pre-training Hours=960h2019.045.12.11
wav2vec largeLanguage Model=WORD CONVLM, Pre-training Data=Librispeech, Pre-training Hours=960h2019.045.162.13
wav2vecLanguage Model=CHAR CONVLM, Pre-training Data=Librispeech, Pre-training Hours=960h2019.045.312.14
wav2vecLanguage Model=WORD CONVLM, Pre-training Data=Librispeech, Pre-training Hours=960h2019.045.392.22
Lattice-free MMIphoneme-based=true2019.045.66—
ESPRESSOArchitecture=LSTM, Look-ahead Word LM=true, Improved Coverage=true, EOS Threshold=false, Beam Size=502019.095.9—
ESPRESSOArchitecture=LSTM, Look-ahead Word LM=true, Improved Coverage=true, EOS Threshold=true, Beam Size=502019.095.9—
Log-mel filterbanks (Baseline)Language Model=WORD CONVLM2019.046.272.57
Likhomanenko et al.2019.096.4—
CNN-DNN-BLSTM-HMMSpeaker adaptation=true, Acoustic states=3k2018.126.6—
Log-mel filterbanks (Baseline)Language Model=CHAR CONVLM2019.046.672.77
Learnable front-endFilters=80, Language Model=ConvLM2018.126.8—
Trainable frontend2019.046.8—
Zeghidour et al.2019.096.8—
Learnable front-endFilters=40, Language Model=ConvLM2018.126.9—
wav2vec largeLanguage Model=4-GRAM LM, Pre-training Data=Librispeech, Pre-training Hours=960h2019.046.962.73
wav2vecLanguage Model=4-GRAM LM, Pre-training Data=Librispeech, Pre-training Hours=960h2019.047.42.85
ESPRESSOArchitecture=LSTM, Look-ahead Word LM=true, Improved Coverage=false, EOS Threshold=false, Beam Size=502019.097.4—
Mel-filterbanksLanguage Model=ConvLM2018.127.5—
wav2vecLanguage Model=4-GRAM LM, Pre-training Data=Libri + WSJ, Pre-training Hours=1,041 h2019.047.592.91
Log-mel filterbanks (Baseline)Language Model=4-GRAM LM2019.048.573.32
wav2vecLanguage Model=4-GRAM LM, Pre-training Data=Librispeech, Pre-training Hours=80h2019.049.113.71
Mel-filterbanksLanguage Model=4-gram2018.129.5—
ESPRESSOArchitecture=LSTM, Look-ahead Word LM=false, Improved Coverage=false, EOS Threshold=false, Beam Size=502019.0914.8—