/ASR-Benchmarks

An effort to track benchmarking results over widely-used datasets for ASR.

ASR benchmarks

An effort to track benchmarking results over widely-used datasets for ASR (Automatic Speech Recognition). Note that the ASR results are affected by a number of factors, it is thus important to report results along with those factors for fair comparisons. In this way, we can measure the progress in a more scientifically way. Feel free to add and correct!

Nomenclature

Terms Explanations
Unit phone (namely monophone), biphone, triphone, wp (word-piece), character, BPE (byte-pair encoding)
AM Acoustic Model. Note that: we also list the end-to-end (e2e) models (e.g., Attention based Seq2Seq, RNN-T) in this field, although those e2e models contains an implicit/internal LM through the encoder.
AM size (M) The number of parameters in millions in the Acoustic Model. Also we report the total number of parameters in the e2e models in this field.
LM Language Model, explicitly used, word-level (by default). ''---'' denotes not using shallow fusion with explicit/external LMs, particularly for Attention based Seq2Seq, RNN-T.
LM size (M) The number of parameters in millions in the neural Language Model. For n-gram LMs, this field denotes the total number of n-gram features.
Data Aug. whether any forms of data augmentations are used, such as SP (3-fold Speech Perturbation from Kaldi), SA (SpecAugment)
Ext. Data whether any forms of external data (either speech data or text corpus) are used
ATT Attention based Seq2Seq
NAS Neural Architecture Search
WPM Word Piece Model
WER Word Error Rate
CER Character Error Rate
--- not applied
? not known from the original paper

WSJ

This dataset contains about 80 hours of training data, consisting of read sentences from the Wall Street Journal, recorded under clean conditions. Available from the LDC as WSJ0 under the catalog number LDC93S6B.

The evaluation dataset contains the simpler eval92 subset and the harder dev93 subset.

Results are sorted by eval92 WER.

eval92 WER dev93 WER Unit AM AM size (M) LM LM size (M) Data Aug. Ext. Data Paper
2.7 5.3 bi-phone LF-MMI, TDNN-LSTM ? 4-gram ? SP --- LF-MMI TASLP2018
2.77 5.68 mono-phone CTC-CRF, TDNN NAS 11.9 4-gram 2.59 SP --- NAS SLT2021
3.0 6.0 bi-phone EE-LF-MMI, TDNN-LSTM ? 4-gram ? SP --- EE-LF-MMI TASLP2018
3.2 5.7 mono-phone CTC-CRF, VGG BLSTM 16 4-gram 2.59 SP --- CAT IS2020
3.4 5.9 sub-word ATT, LSTM 18 --- 113 --- --- ESPRESSO ASRU2019
3.79 6.23 mono-phone CTC-CRF, BLSTM 13.5 4-gram 2.59 SP --- CTC-CRF ICASSP2019
5.0 8.1 mono-char CTC-CRF, VGG BLSTM 16 4-gram 2.59 SP --- CAT IS2020

Swbd

This dataset contains about 260 hours of English telephone conversations between two strangers on a preassigned topic (LDC97S62). The testing is commonly conducted on eval2000 (a.k.a. hub5'00 evaluation, LDC2002S09 for speech data and LDC2002T43 for transcripts), which consists of two test subsets - Switchboard (SW) and CallHome (CH).

Results in square brackets denote the weighted average over SW and CH based on our calculation when not reported in the original paper.

Results are sorted by Sum WER.

SW CH Sum Unit AM AM size (M) LM LM size (M) Data Aug. Ext. Data Paper
6.4 13.4 9.9 char RNN-T, BLSTM-LSTM, ivector 57 LSTM 84 SP, SA, etc. Fisher transcripts Advancing RNN-T ICASSP2021
6.5 13.9 10.2 phone LF-MMI, TDNN-f ? Transformer 25 SP Fisher transcripts P-Rescoring ICASSP2021
6.8 14.1 [10.5] WPM 1k ATT ? LSTM ? SA Fisher transcripts SpecAug IS2019
7.9 15.7 11.8 char RNN-T BLSTM-LSTM 57 LSTM 5 SP, SA, etc. --- Advancing RNN-T ICASSP2021
8.3 17.1 [12.7] bi-phone LF-MMI, TDNN-LSTM ? LSTM ? SP Fisher transcripts LF-MMI TASLP2018
8.6 17.0 12.8 phone LF-MMI, TDNN-f ? 4-gram ? SP Fisher transcripts P-Rescoring ICASSP2021
8.5 17.4 [13.0] bi-phone EE-LF-MMI, TDNN-LSTM ? LSTM ? SP Fisher transcripts EE-LF-MMI TASLP2018
8.8 17.4 13.1 mono-phone CTC-CRF, VGG BLSTM 39.2 LSTM ? SP Fisher transcripts CAT IS2020
9.0 18.1 [13.6] BPE ATT/CTC ? Transformer ? SP Fisher transcripts ESPnet-Transformer ASRU2019
9.7 18.4 14.1 mono-phone CTC-CRF, chunk-based VGG BLSTM 39.2 4-gram 1.74 SP Fisher transcripts CAT IS2020
9.8 18.8 14.3 mono-phone CTC-CRF, VGG BLSTM 39.2 4-gram 1.74 SP Fisher transcripts CAT IS2020
10.3 19.3 [14.8] mono-phone CTC-CRF, BLSTM 13.5 4-gram 1.74 SP Fisher transcripts CTC-CRF ICASSP2019

FisherSwbd

The Fisher dataset contains about 1600 hours of English conversational telephone speech (First part: LDC2004S13 for speech data, LDC2004T19 for transcripts; second part: LDC2005S13 for speech data, LDC2005T19 for transcripts).

FisherSwbd includes both Fisher and Switchboard datasets, which is around 2000 hours in total. Evaluation is commonly conducted over eval2000 and RT03 (LDC2007S10) datasets.

Results are sorted by Sum WER.

SW CH Sum RT03 Unit AM AM size (M) LM LM size (M) Data Aug. Ext. Data Paper
7.5 14.3 [10.9] 10.7 bi-phone LF-MMI, TDNN-LSTM ? LSTM ? SP --- LF-MMI TASLP2018
7.6 14.5 [11.1] 11.0 bi-phone EE-LF-MMI, TDNN-LSTM ? LSTM ? SP --- EE-LF-MMI TASLP2018
7.3 15.0 11.2 ? mono-phone CTC-CRF, VGG BLSTM 39.2 LSTM ? SP --- CAT IS2020
8.3 15.5 [11.9] ? char ATT ? --- ? SP --- Tencent-IS2018
8.1 17.5 [12.8] ? char RNN-T ? 4-gram ? SP --- Baidu-ASRU2017

Librispeech

The LibriSpeech corpus is derived from audiobooks that are part of the LibriVox project, and contains 1000 hours of speech sampled at 16 kHz. The corpus is freely available for download, along with separately prepared language-model training data and pre-built language models.

There are four test sets: dev-clean, dev-other, test-clean and test-other. For the sake of display, the results are sorted by test-clean WER.

dev clean WER dev other WER test clean WER test other WER Unit AM AM size (M) LM LM size (M) Data Aug. Ext. Data Paper
1.55 4.22 1.75 4.46 triphone multistream CNN ? self-attentive simple recurrent unit (SRU) 139 SA --- ASAPP-ASR
--- --- 1.9 3.9 wp Conformer 119 LSTM ? SA --- Conformer
--- --- 1.9 4.1 wp ContextNet (L) 112.7 LSTM ? SA --- ContextNet
3.87 10.28 4.09 10.65 phone CTC-CRF, BLSTM 13 4-gram 1.45 --- --- CTC-CRF ICASSP2019
--- --- 4.28 --- tri-phone LF-MMI, TDNN ? 4-gram ? SP --- LF-MMI Interspeech

AISHELL-1

AISHELL-ASR0009-OS1, is a 178- hour open source mandarin speech corpus. It is a part of AISHELL-ASR0009, which contains utterances from 11 domains, including smart home, autonomous driving, and industrial production. The whole recording was made in quiet indoor environment, using 3 different devices at the same time: high fidelity microphone (44.1kHz, 16-bit,); Android-system mobile phone (16kHz, 16-bit), iOS-system mobile phone (16kHz, 16-bit). Audios in high fidelity were re-sampled to 16kHz to build AISHELL- ASR0009-OS1. 400 speakers from different accent areas in China were invited to participate in the recording. The corpus is divided into training, development and testing sets.

test CER Unit AM AM size (M) LM LM size (M) Data Aug. Ext. Data Paper
4.5 char Conformer based CTC/attention ? LSTM ? SA+SP --- WNARS
4.72 char Conformer based CTC/attention ? attention rescoring ? SA+SP --- U2
5.2 char Comformer ? --- ? SA --- intermediate CTC loss
6.34 phone VGGBLSTM 16M 4-gram 0.7 SP --- CAT IS2020

CHiME-4

The 4th CHiME challenge sets a target for distant-talking automatic speech recognition using a read speech corpus. Two types of data are employed: 'Real data' - speech data that is recorded in real noisy environments (on a bus, cafe, pedestrian area, and street junction) uttered by actual talkers. 'Simulated data' - noisy utterances that have been generated by artificially mixing clean speech data with noisy backgrounds.

There are four test sets. For the sake of display, the results are sorted by eval real WER.

dev simu WER dev real WER eval simu WER eval real WER Unit AM AM size (M) LM LM size (M) Data Aug. Ext. Data Paper
1.15 1.50 1.45 1.99 phone wide-residual BLSTM ? LSTM ? --- --- Complex Spectral Mapping
1.78 1.69 2.12 2.24 phone 6 DCNN ensemble ? LSTM ? --- --- USTC-iFlytek CHiME4 system
2.10 1.90 2.66 2.74 phone TDNN with LF-MMI ? LSTM ? --- --- Kaldi-CHiME4

References

Short-hands Full references
CTC-CRF ICASSP2019 H. Xiang, Z. Ou. CRF-based Single-stage Acoustic Modeling with CTC Topology. ICASSP 2019.
CAT IS2020 K. An, H. Xiang, Z. Ou. CAT: A CTC-CRF based ASR Toolkit Bridging the Hybrid and the End-to-end Approaches towards Data Efficiency and Low Latency. INTERSPEECH 2020.
NAS SLT2021 H. Zheng, K. AN, Z. Ou. Efficient Neural Architecture Search for End-to-end Speech Recognition via Straight-Through Gradients. SLT 2021.
Conformer Anmol Gulati, James Qin, Chung-Cheng Chiu, Niki Parmar, Yu Zhang, Jiahui Yu, Wei Han, Shibo Wang, Zhengdong Zhang, Yonghui Wu, Ruoming Pang. Conformer: Convolution-augmented Transformer for Speech Recognition. INTERSPEECH 2020.
ContextNet Wei Han∗ , Zhengdong Zhang∗ , Yu Zhang, Jiahui Yu, Chung-Cheng Chiu, James Qin, Anmol Gulati, Ruoming Pang, Yonghui Wu. ContextNet: Improving Convolutional Neural Networks for Automatic Speech Recognition with Global Context. INTERSPEECH 2020.
ASAPP-ASR Jing Pan, Joshua Shapiro, Jeremy Wohlwend, Kyu J. Han, Tao Lei, Tao Ma. ASAPP-ASR: Multistream CNN and Self-Attentive SRU for SOTA Speech Recognition. INTERSPEECH 2020.
U2 Binbin Zhang , Di Wu , Zhuoyuan Yao , Xiong Wang, Fan Yu, Chao Yang, Liyong Guo, Yaguang Hu, Lei Xie , Xin Lei. Unified Streaming and Non-streaming Two-pass End-to-end Model for Speech Recognition, arXiv:2012.05481.
Kaldi-CHiME4 Szu-Jui Chen, Aswin Shanmugam Subramanian, Hainan Xu, Shinji Watanabe. Building state-of-the-art distant speech recognition using the CHiME-4 challenge with a setup of speech enhancement baseline. INTERSPEECH 2018.
USTC-iFlytek CHiME4 system Jun Du , Yan-Hui Tu , Lei Sun , Feng Ma , Hai-Kun Wang , Jia Pan , Cong Liu , Jing-Dong Chen , Chin-Hui Lee. The USTC-iFlytek System for CHiME-4 Challenge.
Complex Spectral Mapping Zhong-Qiu Wang, Peidong Wang, DeLiang Wang. Complex Spectral Mapping for Single- and Multi-Channel Speech Enhancement and Robust ASR. TASLP 2020.
intermediate CTC loss Jaesong Lee , Shinji Watanabe. Intermediate Loss Regularization for CTC-based Speech Recognition. ICASSP 2021
WNARS Zhichao Wang, Wenwen Yang, Pan Zhou, Wei Chen. WNARS: WFST based Non-autoregressive Streaming End-to-End Speech Recognition.
LF-MMI H. Hadian, H. Sameti, D. Povey, and S. Khudanpur, “Flat-start single-stage discriminatively trained HMM-based models for ASR,” TASLP 2018.
LF-MMI Interspeech D. Povey, et al. Purely Sequence-Trained Neural Networks for ASR Based on Lattice-Free MMI. Interspeech 2016.
ESPRESSO Yiming Wang, Tongfei Chen, Hainan Xu, Shuoyang Ding, Hang Lv, Yiwen Shao, Nanyun Peng, Lei Xie, Shinji Watanabe, and Sanjeev Khudanpur. Espresso: A fast end- to-end neural speech recognition toolkit. ASRU 2019.
Advancing RNN-T George Saon, Zoltan Tueske, Daniel Bolanos, Brian Kingsbury. Advancing RNN Transducer Technology for Speech Recognition. ICASSP 2021.
P-Rescroing Ke Li, Daniel Povey, Sanjeev Khudanpur. A Parallelizable Lattice Rescoring Strategy with Neural Language Models. ICASSP 2021.
SpecAug D. S. Park, W. Chan, Y. Zhang, et al., SpecAugment: A simple data augmentation method for automatic speech recognition. Interspeech 2019.
ESPnet-Transformer S. Karita, N. Chen, and et al. A comparative study on transformer vs RNN in speech applications,” ASRU 2019.
Baidu-ASRU2017 E. Battenberg, J. Chen, R. Child, A. Coates, Y. Li, H. Liu, S. Satheesh, A. Sriram, and Z. Zhu. Exploring neural transducers for end-to-end speech recognition. ASRU 2017.
Tencent-IS2018 C. Weng, J. Cui, G. Wang, J. Wang, C. Yu, D. Su, and D. Yu. Improving attention based sequence-to-sequence models for end-to-end English conversational speech recognition. Interspeech 2018