Cantonese Speech Database

1. Product Overview

The Cantonese speech database comprises a total of 965 hours. It features a sampling rate of 16,000 Hz and a bit depth of 16 bits. The data has been transcribed and annotated by professional speech annotators and rigorously quality-checked, achieving a text accuracy rate exceeding 95%.

2. Speech Data Transcription

Transcribers generate textual content based on the audio they hear, striving to ensure complete consistency between the written text and the spoken audio. General guidelines are as follows: 1) The transcription must exactly match the spoken audio—no extra words, missing words, or errors. 2) Numbers should be transcribed in English Latin form; for example, “one two three” rather than “123.” 3) English text should use lowercase for words, while phonetic spellings must be written in uppercase with spaces, such as “W W W dot Google dot com” or “C E O.” 4) All annotations must fully reflect the actual pronunciation without any omissions.

3. Directory Structure and Naming Conventions

Directory Structure

Data directory tree structure:

└─DOC (textual documentation)
││├─content.txt (transcription list)
└─WAV (audio folder)
││├─Y0000001 (audio file)

Naming Conventions

CORPUS/USAGE/SPEECH_IDe.g./WAV/Y0000001.wav

Directory Name · Content · Notes

CORPUS · · Speech database identifier

USAGE · WAV · Audio folder

SPEECH_ID · Y0000001.wav · Audio file

Sample Data Download

Click to download the sample dataset (data_sample.zip)