[Paper Review] AISHELL-1: An Open-Source Mandarin Speech Corpus and A Speech Recognition Baseline

Hui Bu, Jiayu Du+3 more|arXiv (Cornell University)|Sep 16, 2017

Speech Recognition and Synthesis6 references50 citations

TL;DR

AISHELL-1 is the largest open-source Mandarin ASR corpus (400 speakers, 170+ hours) with a Kaldi baseline and drop-in setup, plus a detailed recording, transcription, and lexicon pipeline.

ABSTRACT

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.

Motivation & Objective

• Provide an open, large-scale Mandarin speech corpus suitable for ASR research and product-level development.
• Detail recording procedures, metadata, transcription and lexicon preparation to ensure data quality and reproducibility.
• Offer a drop-in Kaldi recipe to establish a reliable Mandarin ASR baseline.
• Demonstrate corpus quality through baseline experiments across multiple acoustic models and channels.

Proposed method

• Release of AISHELL-1 as part of a larger AISHELL-ASR0009 corpus subset with 400 speakers and 170+ hours.
• Multi-device recording setup including high-fidelity microphone, Android, and iPhone channels; data re-sampled to 16 kHz, 16-bit WAV.
• Manual transcription, text normalization, and a Chinese lexicon with initial-final pronunciations.
• Kaldi-based ASR baseline starting from GMM-HMM to TDNN-HMM and LFMMI with MFCC and pitch features plus i-vector adaptation.
• Audio augmentation and i-vector-based adaptation used to improve DNN robustness.
• Language model built from 1.3 million words with trigram smoothing and OOV handling.

Fig. 1 : Recording setup

Experimental results

Research questions

• RQ1 Can AISHELL-1 support robust Mandarin ASR research and practical system development across multiple devices?
• RQ2 How does model architecture (GMM-HMM, TDNN-HMM, LFMMI) affect Mandarin ASR performance on AISHELL-1?
• RQ3 What is the impact of device mismatch on ASR performance using AISHELL-1?
• RQ4 Does the AISHELL-1 baseline generalize to unrelated domains (e.g., THCHS30) and different language-model domains?

Key findings

• AISHELL-1 comprises 120,098 training utterances from 340 speakers, 14,326 dev from 40, and 7,176 test from 20 (with around 26 minutes per speaker).
• Baseline results show LFMMI outperforms TDNN-HMM and MLLT+SAT, with CERs of 6.44% dev and 7.62% test for LFMMI.
• TDNN-HMM achieves 7.23% dev and 8.42% test CER, while MLLT+SAT achieves 10.43% dev and 12.23% test CER.
• Device-mismatch tests (iOS vs Android) show LFMMI with 10.90% dev, 10.09% test CER on Android vs 12.64% dev, 11.88% test on iOS for MLLT+SAT, and up to 12.42%/10.81% for TDNN-HMM.
• AISHELL-1 model applied to THCHS30 shows higher CERs for all models, e.g., LFMMI 25.00% on THCHS30, indicating domain adaptation benefits.
• The results suggest the corpus is phonetically rich and suitable for building Mandarin ASR baselines with a Kaldi recipe.