Source-linked AI summary
THCHS-30 : A Free Chinese Speech Corpus
Dong Wang, Xuewei Zhang
TL;DR
ASR research requires substantial speech data, yet licensing costs can restrict access for beginning researchers. The paper addresses this gap by releasing THCHS-30 with supporting resources and baseline evaluations on clean and noisy speech. The release is intended to provide a freely accessible basis for building and comparing Chinese ASR systems.
Problem
Large speech datasets are essential for ASR, but expensive database licenses create an access barrier for researchers beginning in the field.
Method
The paper releases THCHS-30, associated lexica, language models, training recipes, tools, noisy data, and a baseline ASR system.
Results
The authors present baseline performance on both clean and noisy data, including highly noisy 0 db conditions.
Takeaways & Limitations
THCHS-30 provides a free Chinese speech resource that can support practical ASR development and shared comparison through challenges.
Abstract
from arXiv · showhide
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.
1 Introduction
ASR research depends heavily on large speech datasets, but expensive licensing limits access for beginners. The paper advocates free data and introduces THCHS-30 as a freely available Chinese corpus with resources for practical ASR and comparative challenges.
- Speech recognition relies heavily on large amounts of data because human speech contains complex patterns that must be learned.
- Commercial speech databases support comparable evaluation but are often expensive, creating an access barrier for beginning researchers.The paper gives the WSJ corpus's 2,500 US dollar license as an example.
- The authors advocate freely publishing speech data, especially when collection is supported by public funds, to broaden research access.
- Existing free-speech efforts include AMI/AMID, VoxForge, OpenSLR, LibriSpeech, and Scandinavian databases, but the authors identify no sufficiently capable free Chinese corpus.
- THCHS-30 releases a Chinese speech database with lexicon, language model, and training recipe resources for building a practical large-vocabulary continuous ASR system.The authors describe it as the first release they know of with this capability and propose it as a standard reference.
- The paper describes the database, reports a baseline system, and organizes challenges around the released resources.
2 Features of THCHS-30
THCHS-30 is a freely released Mandarin corpus built to improve phone coverage and support complete Chinese ASR development. It includes organized speech data, language resources, recipes, tools, and separately generated noisy conditions for recognition challenges.
- The corpus originated as TCMSD, a supplemental database designed to maximize phone coverage alongside the 863 database, and was later renamed THCHS-30.
- THCHS-30 contains more than 30 hours of Mandarin speech recorded from fluent speakers at 16,000 Hz with 16-bit samples.The recordings use a single carbon microphone in a silent office, and 1,000 sentences were selected to augment phone coverage.
- Training uses groups A, B, and C with 30 speakers and 10,893 utterances, while test group D contains 10 speakers and 2,496 utterances.
- The release provides lexica, language models, training recipes, useful tools, and additional noisy-condition data to support practical Chinese ASR construction.
- The release includes a 48k-word word-based 3-gram language model and a phone-based 3-gram model covering 218 Chinese tonal finial-initials.
- The authors provide a free download, complete Chinese ASR training recipes based on Kaldi, and challenges for large-vocabulary and phone recognition.Challenge evaluation includes clean and 0 db conditions with white, car, and cafeteria noise.
- Noise conditions separately corrupt training and test data with white, car, and cafeteria noise, focusing on 0 db where noise and speech energies are equal.The noisy data are created by simple wave mixing using signals from the public DEMAND repository.
3 Baseline system
The paper establishes a Kaldi-based HMM-DNN baseline for THCHS-30 and evaluates it on clean and noisy speech. It also applies a DAE front end that significantly improves noisy-speech performance and can be adapted using a small noise example.
- Framework and setting: The baseline uses monophone and triphone GMMs to generate alignments for a DNN acoustic model.The GMM stages use MFCC-based features, CMN, LDA, and MLLT before DNN training.
- Framework and setting: The DNN uses spliced 40-dimensional FBank features, LDA reduction to 200 dimensions, four 1200-unit hidden layers, and 3386 output units.Training uses cross-entropy with stochastic gradient descent.
- Preliminary results: Noisy-speech performance is much lower than clean-speech performance, particularly under white noise, while clean-speech CER and PER remain relatively high.The authors attribute the high errors partly to peculiar test sentences selected for maximal phone coverage.
- Preliminary results: The test-set peculiarity makes recognition highly challenging and may require techniques such as more aggressive phone sharing.The issue concerns unusual pronunciation and spelling in the selected sentences.
- Noise cancellation with DAE: A simple DAE is trained on artificially noise-corrupted speech to learn a nonlinear transform for noise cancellation.The approach requires finding a noise signal and using it to corrupt clean speech; new noise types may otherwise require DNN retraining.
- Noise cancellation with DAE: The DAE maps an 11-frame FBank window to a noise-removed central-frame feature, which is then processed before the clean-trained DNN.The outputs are spliced, LDA transformed, globally normalized, and fed to the acoustic model.
- Noise cancellation with DAE: DAE-based noise cancellation significantly improves the baseline CER/PER results on noisy data and supports noise-specific adaptation from a small noise example.The comparison is reported between the noisy-data results in Tables 3 and 4.
4 Conclusions
The paper releases THCHS-30 as a free Chinese speech database with resources for constructing practical ASR systems. It presents baseline results on clean and noisy data and positions the release as a basis for challenges and collaboration.
- Conclusions: THCHS-30 is released as a free 30-hour Chinese speech database for building Chinese ASR systems.The release is intended to support new researchers entering speech recognition.
- Conclusions: The release includes lexica, language models, training recipes, useful tools, and additional noisy-condition data.These resources are provided to assist construction of a practical Chinese ASR system.
- Conclusions: The paper demonstrates the building process and reports baseline results on both clean and noisy data.
- Conclusions: The authors describe THCHS-30 as, to their knowledge, the first free Chinese database sufficient for a practical Chinese ASR system.They also call for challenges based on the database to encourage innovation and collaboration.