AudioChinese dialects

Cantonese Dialect Speech Database

About this dataset

1. Product Overview

The Cantonese dialect speech database contains a total of 768 hours of audio. The data types include read speech and conversational speech, with 618.8 hours of read speech and 149.4 hours of conversational speech. The recording language is Cantonese, and the recordings were made in Guangzhou, China. A total of 2,672 speakers were invited to record in quiet environments using their mobile phones. This database has undergone professional proofreading, achieving a text accuracy rate of over 95%, and can be used for research in areas such as speaker recognition and speech recognition.

2. Recording Corpus

Recording Text Results

In the read-speech portion of this database, each speaker recorded 470 sentences, covering wake-up words, command words, free-form text, and more. Another set of data was primarily collected for in-vehicle voice control, with each speaker recording approximately 10–15 minutes. The specific distribution is as follows:

PY0001-PY3400 · PY3501-PY5128 · · ·

Domain · Example Sentences · ID Numbers · Domain ·

Wake-Up Words · Xiao Ai Tongxue, Xiao Du Xiao Du · 0001-0021 · ·

POI · Navigate to Science Square · 0022-0044 · ·

Home Control · Please help me turn on video two · 0045-0051 · ·

Phonetic Spelling · Q M P E · 0052-0054 · ·

Numbers · Eight hundred sixty zero nine thousand five hundred seventy-seven · 0055-0074 · ·

Q&A · What movies did Chen Shisan star in? · 0075-0089 · ·

Music · Sing a song by Wang Sutong · 0090-0119 · ·

TV & Movies · The Game Girl Cantonese version · 0120-0139 · In-Vehicle Voice Control · 0001-0300

Radio · I want to tune AM to 915.5 MHz · 0140-0144 · ·

Entertainment · Ouyang Nana later worked in mainland China · 0145-0160 · ·

Finance · Year-on-year growth of 1.9% · 0161-0240 · ·

Technology · Total shipments fell short of Apple Watch · 0241-0290 · ·

Sports · Going forward, Alas will base itself on North马 · 0291-0340 · ·

Current Affairs News · A notice was posted on the glass door · 0341-0390 · ·

General · For you, tonight must be very long · 0391-0470 · ·

Total · · 470 · · 300

The in-vehicle voice-control segment is distributed among vehicle control, music, navigation, inquiries, etc., as follows:

Domain · Proportion

Music · 50.0%

POI · 27.4%

Inquiries · 3.8%

Vehicle Control · 7.8%

Other · 11.0%

Total · 100%

During conversations, speakers freely recorded based on given topics, which are roughly categorized as follows:

Domain · Proportion

E-commerce Retail · 14%

Finance · 12%

Education · 13%

Technology · 7%

Life Services · 13%

Automotive Sales · 7%

Real Estate · 9%

Logistics · 5%

Other · 20%

Corpus Pool Processing

 De-identification processing: Remove politically sensitive, personally private, pornographic, or violent content.  Delete symbols such as <, >, [, ], ~, /, \, =, except for numeric strings.  Remove content containing languages other than Chinese and English.  Standardize formatting.

3. Speaker Information

Gender Ratio

The database includes a total of 2,672 speakers, with 1,125 males and 1,547 females.

Gender · Male · Female · Total

Proportion · 42.1% · 57.9% · 100%

· Age Group · Number of Speakers · Proportion

A · 18–25 years old · 686 · 25.6%

B · 26–40 years old · 1,708 · 63.9%

C · 41–50 years old · 223 · 8.4%

D · 51+ years old · 55 · 2.1%

Total · 2,672 · 100% ·

Age Distribution

The age distribution of speakers is as follows: A (18–25 years old); B (26–40 years old); C (41–50 years old); D (51+ years old).

4. Data Recording Environment

Recording Environment

A quiet indoor setting, free from echoes and other voices.

Recording Equipment

Common smartphones available on the market, such as Apple, Huawei, Xiaomi, OPPO, etc. The data storage format is 16 kHz, 16-bit WAV files.

Recording Method

For the read-speech portion, speakers recorded in quiet indoor settings using smartphone apps, reading texts at normal speed. For the conversation portion, speakers recorded via phone calls in quiet indoor settings.

5. Speech Data Transcription

Transcribers write down the content based on what they hear, striving to ensure that the written text matches the spoken audio exactly. General guidelines are as follows: 1) The transcription must match the spoken audio word-for-word; no extra words, missing words, or errors are allowed. 2) Numbers should be transcribed into Chinese characters, e.g., “一二三” instead of “123.” Pay special attention to distinguishing between “一” and “幺,” and “二” and “两.” 3) If the audio contains English pronunciation, it should be transcribed into corresponding Chinese characters or left in English. Specific cases are as follows: All letters or words included in URLs should be capitalized. For example, if the audio says “www.abc.com,” it should be transcribed as “三W点A B C点com.” English words within the audio should be transcribed entirely in lowercase. English letters within the audio should be transcribed entirely in uppercase. For certain proper nouns written in all caps, keep them capitalized without spaces, such as CEO, NBA, etc. 4) The completeness of annotations must match the actual pronunciation; no deletions are permitted.

6. Data File Directory

Directory Structure

The directory tree structure of the data is as follows:

└─dialogue（Conversation Data）
│├─DOC（Text Folder）
│├─wav_list.txt（Audio List）
│├─spkrinfo.xlsx（Speaker Information）
│├─Textgrid（Transcription Text Folder）
│├─PYD_0623.Textgrid（Transcription Text）
│├─SPEECHDATA（Data Folder）
││ ││ PYD_0623.wav（Audio File）
│
└─reading（Read-Speech Data）

│├─DOC (Text Folder)
│├─content.txt（Transcription Text）
│├─wav_list.xlsx（Audio List）
│├─spkrinfo.xlsx（Speaker Information）
│├─SPEECHDATA (Data Folder)
│├─PY0028 (Speaker ID)
││ ││ PY0028.wav（Audio File）