AISHELL-1: An open-source Mandarin speech corpus and a speech recognition baseline

By Hui Bu, Jiayu Du, Xingyu Na, Ben Wu, Hao Zheng. 2017

1,067 citations recorded by the source corpus.

An open-source Mandarin speech corpus called AISHELL-1 is released. It is by far the largest corpus which is suitable for conducting the speech recognition research and building speech recognition systems for Mandarin. The recording procedure, including audio capturing devices and environments are presented in details. The preparation of the related resources, including transcriptions and lexicon are described. The corpus is released with a Kaldi recipe. Experimental results implies that the quality of audio recordings and transcriptions are promising.

Subfields: Artificial Intelligence, Speech & Audio, Signal Processing

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

1. Jinyu Li
2. Dong Yu
3. Shinji Watanabe
4. Xu Tan
5. Tao Qin
6. Tie-Yan Liu
7. Yoshua Bengio
8. Hung-yi Lee

In Speech & Audio

1. Shinji Watanabe
2. Jinyu Li
3. Hung-yi Lee
4. Dong Yu
5. Yoshua Bengio
6. Xu Tan
7. Takaaki Hori
8. Tao Qin

In Signal Processing

1. Dong Yu
2. Jinyu Li
3. Shinji Watanabe
4. Yoshua Bengio
5. Hung-yi Lee
6. Li Deng
7. Xu Tan
8. Tie-Yan Liu

Citations in the corpus

• Improving Transformer-based Speech Recognition Using Unsupervised Pre-training (2019, methodology): “The corpora we collected include: HKUST Mandarin Telephone Speech Corpus (HKUST/MTS) [18], AISHELL-1 [21], aidatatang 200zh, MAGICDATA Mandarin Chinese Read Speech Corpus, Free ST Chinese Mandarin Corpus (STCMDS) and Primewords Chinese Corpus Set 1.”
• FunCodec: A Fundamental, Reproducible and Integrable Open-Source Toolkit for Neural Speech Codec (2023, methodology): “To assess the generalization ability of our generalized model, we evaluate it on the test sets of multiple open-source corpora, including Librispeech [28], aishell-1 [29], aishell-2 [30], Wenet [31] and Gigaspeech [32].”
• A Survey on Neural Speech Synthesis (2021, neutral): “1 English CSMSC [15] 12 1 48 Mandarin HKUST [205] 200 2100 8 Mandarin AISHELL-1 [28] 170 400 16 Mandarin AISHELL-2 [69] 1000 1991 44.”
• A Comparative Study on Transformer vs RNN in Speech Applications (2019, background): “AISHELL [28] zh 170 read dev / test AURORA4 [29] (*) en 15 noisy read (dev 0330) A / B / C / D CSJ [30] ja 581 spontaneous eval1 / eval2 / eval3 CHiME4 [31] (*) en 108 noisy far-field multi-ch read dt05 simu / dt05 real / et05 simu / et05 real CHiME5 [32] en 40 noisy far-field multi-ch conversationa”
• UWSpeech: Speech to Speech Translation for Unwritten Languages (2020, methodology): “For the Chinese dataset, we use AIShell (Bu et al. 2017) which contains about 140K training examples (178 hours) with 400 different speakers.”
• WENETSPEECH: A 10000+ Hours Multi-Domain Mandarin Corpus for Speech Recognition (2021, methodology): “AIShell-1 [36]) with Kaldi, ESPNet and WeNet toolkits respectively.”
• Didispeech: A Large Scale Mandarin Speech Corpus (2020, background): “For ASR, several open-source read speech corpora, such as HKUST and AISHELL-1, have also been released.”
• Simplified Self-Attention for Transformer-Based end-to-end Speech Recognition (2020, methodology): “In this paper, we validated the proposed SSAN-based transformer on three Mandarin speech recognition datasets: public AISHELL-1 corpus [24], internal 1000 and 20,000 hours corpus same as used in [25].”
• Non-Autoregressive Transformer Automatic Speech Recognition (2019, positive): “In Aishell experiments encoder includes 12 transformer blocks with convolu- tional layers at the beginning for downsampling.”
• Interspeech 2021 Deep Noise Suppression Challenge (2021, methodology): “Mandarin data consists of OpenSLR18 5 THCHS-30 [15] and OpenSLR33 6 AISHELL [16] datasets, both with the Apache 2.0 license.”
• Leveraging Pseudo-labeled Data to Improve Direct Speech-to-Speech Translation (2022, methodology): “For Chinese ASR, we use the attention-based Conformer[24] model(2) from Wenet[25] trained on the AISHELL[26] corpus.”
• AISHELL-4: An Open Source Dataset for Speech Enhancement, Separation, Recognition and Speaker Diarization in Conference Scenario (2021, methodology): “For Mandarin speech corpus, the commonly used open source datasets for speech recognition include AISHELL1 [20], ASHELL-2 [21], aidatatang(2), etc.”
• Universal Phone Recognition with a Multilingual Allophone System (2020, background): “English voxforge, Tedlium [17], Switchboard [18] 1148k Japanese Japanese CSJ [19] 440k Mandarin Hkust [20], openSLR [21, 22] 377k Tagalog IARPA-babel106b-v0.”
• WeNet: Production Oriented Streaming and Non-Streaming End-to-End Speech Recognition Toolkit (2021, methodology): “Since the average utterance duration in AISHELL-1 is much longer than that in the other two test sets, AISHELL-1 task may need stronger global information modeling ability.”
• Multi-Lingual Multi-Speaker Text-to-Speech Synthesis for Voice Cloning with Online Speaker Enrollment (2020, methodology): “We further randomly select 800 speakers’ data for training, others for testing; (3) CUSENT (CU) [21] is a Cantonese corpus with 20 hours of speech from 68/12 training/test speakers; (4) Aishell (AI) [22] is a Mandarin corpus consisting of 150 hours of speech from 340 speakers.”
• SpeechBrain: A General-Purpose Speech Toolkit (2021, methodology): “CTC [24] Transducers [25] CTC+Attention [26] Shallow fusion [27] LibriSpeech [28] Common Voice [29] AISHELL [30] TIMIT [31] separation, pyannote [17] and sidekit [18] for speaker diarization, and s3prl [19] for self-supervised speech representations.”
• Code-Switched Speech Synthesis Using Bilingual Phonetic Posteriorgram with Only Monolingual Corpora (2020, methodology): “The English and Mandarin ASR models are trained on TIMIT [33] and AI-SHELL1 [34] corpus respectively.”
• AISHELL-2: Transforming Mandarin ASR Research Into Industrial Scale (2018, methodology): “In AISHELL-1, word segmentation was implemented via forward maximum matching algo-rithm, based on a variation of open-source mandarin dictionary CC-CEDIT 1 .”
• Speech Recognition via Large-Scale Weak Supervision (positive): “In Chinese, for example, the HKUST Mandarin Telephone Speech corpus has 1206 transcribed HKUST ten-minute telephone conversations between speakers of Mandarin across China including conversations between friends and between strangers (Liu et al., 2006).”