PaperlayerView exact revision on arXiv

Source-linked AI summary

AISHELL-1: An Open-Source Mandarin Speech Corpus and A Speech Recognition Baseline

Hui Bu, Jiayu Du, Xingyu Na, Bengu Wu, Hao Zheng

arXiv:1709.05522v1cs.CL

TL;DR

Mandarin ASR research lacked a large, publicly accessible corpus because industrial datasets were generally unavailable and THCHS30 was small. The paper releases AISHELL-1 with documented recordings, text resources, and a Kaldi baseline, and reports promising audio and transcription quality under baseline evaluation.

  • Problem

    Chinese ASR lacked a large open-source corpus: THCHS30 contained 50 speakers and around 30 hours, while industrial datasets were often inaccessible to academics.

  • Method

    The paper constructs and releases AISHELL-1, documenting recording procedures, transcriptions, a lexicon, corpus resources, and a Kaldi Mandarin ASR recipe.

  • Results

    Experimental evaluations with the released recipe indicate promising audio and transcription quality for constructing Mandarin speech recognition systems.

  • Takeaways & Limitations

    AISHELL-1 provides publicly available Mandarin speech data and a baseline for speech recognition research and system construction.

Abstract

from arXiv · show

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.

1. INTRODUCTION

AISHELL-1 addresses the scarcity of publicly available, large-scale Mandarin ASR data by releasing a 400-speaker corpus with over 170 hours of speech, an Apache 2.0 license, and a Kaldi baseline.

  • Industrial datasets are often inaccessible to academic researchers, widening the gap between research and industry.
  • Chinese ASR has had only THCHS30 as an open-source corpus, with 50 speakers and around 30 hours of Mandarin speech.
  • AISHELL-1 contains 400 speakers and over 170 hours of Mandarin speech data, and is publicly available under Apache 2.0.
  • The release includes recording procedures, transcriptions, a lexicon, corpus resources, and a “drop-in and run” Kaldi Mandarin ASR baseline.

2. CORPUS PROFILE

AISHELL-1 is drawn from a larger multi-channel Mandarin corpus and standardizes high-fidelity microphone recordings into 16 kHz, 16-bit WAV data while documenting speaker metadata.

  • AISHELL-1 is a subset of AISHELL-ASR0009, a 500-hour multi-channel Mandarin corpus designed for speech and speaker processing tasks.
  • Recordings use high-fidelity microphones, Android phones, and iPhones, with the phone devices operating at 16 kHz and 16-bit.
  • AISHELL-1 selects high-fidelity microphone audio and resamples it to 16 kHz, 16-bit WAV, a mainstream commercial-product setup.
  • The corpus records speaker gender, accent, age, and birthplace metadata for 400 participants, with 47% male and 53% female speakers.

3. TRANSCRIPTION AND LEXICON

The corpus prepares Mandarin text and pronunciation resources through domain selection, text filtering, transcription checks, normalization, and a Chinese lexicon.

  • AISHELL-1 covers five text domains from the larger corpus, including Finance, Science and Technology, Sports, Entertainments, and News.
  • Raw texts are manually filtered to remove sensitive, private, pornographic, and violent content, symbols, and sentences longer than 25 words.
  • Quality checking removes utterances with inconsistent transcriptions and obvious mispronunciations.
  • Text normalization is applied to English words, numbers, names, places, organizations, streets, shops, and brands.
  • The provided Chinese lexicon covers most commonly used Chinese words and characters, with pronunciations represented in initial-final syllables.

4. DATA STRUCTURE

AISHELL-1 is divided into speaker-disjoint training, development, and test sets with specified utterance and speaker counts.

  • The training set contains 120,098 utterances from 340 speakers.
  • The development set contains 14,326 utterances from 40 speakers.
  • The test set contains 7,176 utterances from 20 speakers.
  • The training, development, and test sets do not overlap in speakers, and each speaker contributes about 26 minutes of speech.

5. SPEECH RECOGNITION BASELINE

The paper releases a Kaldi speech-recognition baseline for AISHELL-1, combining conventional and neural acoustic models with a trigram language model. Evaluation examines transcription quality, device mismatch, and performance on an unrelated corpus.

  • 5. SPEECH RECOGNITION BASELINE: The released Kaldi recipe demonstrates Mandarin speech recognition using GMM-HMM, TDNN-HMM, and LFMMI acoustic models.The baseline uses tone-dependent GMM-HMM training, DNN-based TDNN models, and LFMMI training for comparison.
  • 5. SPEECH RECOGNITION BASELINE: The acoustic models use MFCC and pitch features, with DNN systems additionally incorporating high-resolution features and i-vectors.The GMM-HMM pipeline applies MLLT and speaker adaptive training, while DNN training uses feature splicing and augmentation.
  • 5. SPEECH RECOGNITION BASELINE: A trigram language model is trained on 1.3 million transcript words using interpolated Kneser-Ney smoothing.The resulting model contains 137,076 unigrams, 438,252 bigrams, and 100,860 trigrams.
  • 5. SPEECH RECOGNITION BASELINE: The results are reported as character error rate, and LFMMI significantly outperforms TDNN-HMM, indicating high transcription quality.The evaluation also compares performance on mobile recordings and the THCHS30 testing set.
  • 5. SPEECH RECOGNITION BASELINE: Device mismatch causes significant performance loss on mobile recordings, but stronger acoustic models improve performance on less-fidelity devices.The mobile evaluation uses parallel Android and iOS recordings from the AISHELL-ASR0009 corpus.
  • 5. SPEECH RECOGNITION BASELINE: On the unrelated THCHS30 domain, stronger acoustic models perform better, suggesting phonetic coverage while an adapted language model can reduce the remaining gap.Table 7 is used to assess overall corpus quality on data from a different language-model domain.

6. CONCLUSIONS

AISHELL-1 releases an open-source Mandarin speech corpus with a Kaldi recipe and accompanying resources. Experimental results indicate promising audio and transcription quality for constructing Mandarin speech-recognition systems.

  • 6. CONCLUSIONS: AISHELL-1 is released as an open-source Mandarin corpus with a Kaldi recipe and related experimental results.The corpus is described as the largest academically free dataset for Mandarin speech-recognition tasks.
  • 6. CONCLUSIONS: The reported audio and transcription qualities are promising for constructing Mandarin speech-recognition systems.
Loading 1709.05522v1…