[Paper Review] THCHS-30 : A Free Chinese Speech Corpus

Dong Wang, Xuewei Zhang|arXiv (Cornell University)|Dec 7, 2015

Speech Recognition and SynthesisComputer Science12 references190 citations

TL;DR

This paper introduces THCHS-30, a free 30-hour Chinese speech corpus designed to lower barriers for entry in speech recognition research. It presents a DNN-based ASR baseline and demonstrates noise robustness using a deep autoencoder (DAE) for real-time noise cancellation, significantly improving word error rates on noisy conditions—e.g., reducing CER from 96.44% to 75.01% on white noise.

ABSTRACT

Speech data is crucially important for speech recognition research. There are quite some speech databases that can be purchased at prices that are reasonable for most research institutes. However, for young people who just start research activities or those who just gain initial interest in this direction, the cost for data is still an annoying barrier. We support the `free data' movement in speech recognition: research institutes (particularly supported by public funds) publish their data freely so that new researchers can obtain sufficient data to kick of their career. In this paper, we follow this trend and release a free Chinese speech database THCHS-30 that can be used to build a full- edged Chinese speech recognition system. We report the baseline system established with this database, including the performance under highly noisy conditions.

Motivation & Objective

• To address the high cost of commercial speech databases that hinder entry for early-career and developing-country researchers.
• To support the 'free data' movement by releasing a publicly available, high-quality Chinese speech corpus.
• To enable researchers to build a full-fledged Chinese speech recognition system from scratch using only free resources.
• To provide baseline performance metrics on both clean and noisy speech to serve as a reference point for future research.
• To demonstrate the effectiveness of DAE-based noise cancellation in improving ASR robustness without retraining the acoustic model.

Proposed method

• The THCHS-30 corpus was collected from 30 native Mandarin speakers reading 1000 sentences each, covering diverse phonetic and lexical content to maximize phone coverage.
• Audio was recorded at 16 kHz with 16-bit PCM, and transcriptions were manually verified to ensure high accuracy.
• A DNN-based acoustic model was trained using standard features (Fbank), feature splicing, LDA, and global normalization.
• A deep autoencoder (DAE) was applied as a front-end noise canceller, trained to reconstruct clean Fbank features from noisy inputs corrupted at random SNR levels centered at 0 dB.
• The DAE was trained end-to-end on synthetic noisy data generated by mixing clean speech with real noise samples (e.g., car, cafeteria, white noise).
• The DAE output was used as input to the DNN model, replacing standard Fbank features in the pipeline, enabling real-time noise suppression without retraining the main model.

Experimental results

Research questions

• RQ1 Can a free, publicly available Chinese speech corpus support the development of a practical, end-to-end speech recognition system?
• RQ2 How does a DNN-based ASR system perform on the THCHS-30 corpus under clean and noisy conditions?
• RQ3 To what extent can a DAE-based noise cancellation method improve ASR performance on unseen noise types without retraining the acoustic model?
• RQ4 Does the inclusion of highly unusual or complex sentences in the corpus significantly degrade recognition performance, and can this be mitigated?
• RQ5 Can a lightweight, noise-specific DAE be effectively used to enhance robustness in real-world noisy environments?

Key findings

• The THCHS-30 corpus is the first free, large-scale Chinese speech database suitable for building a full-fledged ASR system, with 30 hours of transcribed speech from 30 speakers.
• On clean speech, the DNN baseline achieved a word error rate (CER) of 30.11% and a phone error rate (PER) of 14.81%, indicating strong baseline performance.
• On highly corrupted white noise (0 dB SNR), the CER without noise cancellation reached 96.44%, highlighting the system’s fragility under extreme noise.
• After applying DAE-based noise cancellation, the CER on white noise dropped to 75.01%, and on car noise to 32.13%, showing significant robustness improvement.
• The DAE method effectively reduced noise corruption with minimal computational overhead and no need to retrain the main DNN model.
• The performance degradation on clean speech was attributed to the corpus’s high phonetic diversity and unusual sentence structures, which challenge standard acoustic models.