Mandarin Chinese speech data

1. Product Overview

The Mandarin Chinese speech database contains a total of 3,002 hours of recordings. The transcribed texts cover 15 domains, including smart home, autonomous driving, and industrial production. A total of 2,000 speakers from various accent regions across China participated in the recording process. Recordings were conducted in a quiet indoor environment using three different devices: a high-fidelity microphone (44.1 kHz, 16-bit, 999.7 Hz); an Android smartphone (16 kHz, 16-bit, 1000.7 Hz); and an iOS smartphone (16 kHz, 16-bit, 1002.4 Hz). This database has been professionally transcribed and annotated by skilled linguists and underwent rigorous quality assurance, achieving a text accuracy rate exceeding 97%.

2. Recording Corpus

Corpus Construction

Corpus Content

Considering the applications of speech recognition in areas such as smart homes, autonomous vehicles, and industrial manufacturing, the corpus was curated across 15 domains, comprising approximately 3 million commonly used Chinese sentences.

C0001-C1000 · C9001-C9348 · ·

Serial Number · Domain · Serial Number · Domain

1 · Home Appliances Names and Control Commands · 1 · Home

2 · POI (Geographic Information) · 2 · POI

3 · Music-related Commands · 3 · Music-related Commands

4 · Digital Commands · 4 · Digital Commands

5 · TV and Movie Titles · 5 · TV Titles

6 · Finance · 6 · Finance—Policies

7 · Technology · 7 · Finance—Real Estate

8 · Sports · 8 · Technology—International

9 · Entertainment · 9 · Technology—Domestic

10 · Current Affairs News · 10 · Sports

11 · English Pronunciation · 11 · Entertainment—Western

· · 12 · Entertainment—Hong Kong/Taiwan

· · 13 · Current Affairs News

· · 14 · General

· · 15 · English Pronunciation

Corpus Processing

• De-identification: Removal of politically sensitive content, personal privacy information, and explicit or violent material.
• Removal of symbols such as <, >, [, ], ~, /, \, =, etc.
• Exclusion of content containing languages other than Chinese and English.
• Standardization of formatting.

Structure Design of Transcribed Texts

To ensure comprehensive coverage and phonetic balance, the C0001–C1000 segment adopts a structure of 500 sentences per file, while the C9001–C9348 segment uses 520 sentences per file. The following table illustrates the distribution:

C0001-C1000 · · · C9001-C9348 · ·

Serial Number · Domain · Quantity per File/Sentence · Serial Number · Domain · Quantity per File/Sentence

1 · Home Appliances Names and Control Commands · 5 · 1 · Home · 5

2 · POI (Geographic Information) · 30 · 2 · POI · 30

3 · Music-related Commands · 46 · 3 · Music-related Commands · 46

4 · Digital Commands · 29 · 4 · Digital Commands · 29

5 · TV and Movie Titles · 10 · 5 · TV Titles · 13

6 · Finance · 132 · 6 · Finance—Policies · 20

7 · Technology · 85 · 7 · Finance—Real Estate · 20

8 · Sports · 66 · 8 · Technology—International · 15

9 · Entertainment · 27 · 9 · Technology—Domestic · 10

10 · Current Affairs News · 66 · 10 · Sports · 20

11 · English Pronunciation · 4 · 11 · Entertainment—Western · 4

· · 12 · Entertainment—Hong Kong/Taiwan · 10

· · 13 · Current Affairs News · 10

· · 14 · General · 200

· · 15 · English Pronunciation · 4

· · 16 · Q&A · 84

Total · 11 items · 500 sentences · Total · 16 items · 520 sentences

3. Speaker Information

Basic Information Records

Speaker information includes task ID, gender, age group, and place of origin.

Task ID · Age Group · Gender · Accent Region

C0001 · B · Female · Northern

Each speaker is assigned one task ID, corresponding to one set of recorded text. Each speaker may participate in only one recording session. Age groups are defined as follows: A (under 18), B (18–25), C (26–40), D (41+). Gender categories are male and female. Accent regions are classified according to the speaker’s native language area: Northern, Southern, and Other.

Speaker Demographics

Gender Ratio

The database comprises 2,000 speakers, with 863 males and 1,137 females.

Gender · Male · Female · Total

Ratio · 43% · 57% · 100%

Age Distribution

Ages under 18: 18; Ages 18–25: 1,528; Ages 26–40: 310; Ages 41+: 144.

Age Group · Number of Speakers · Proportion · Male · Female

A · 18 · 9% · 5 · 13

B · 1,528 · 76.4% · 633 · 895

C · 310 · 15.5% · 159 · 151

D · 144 · 7.2% · 66 · 78

Total · 2,000 · 100.00% · 863 · 1,137

Dialect Distribution

Dialect Region · Number of Speakers · Proportion

Northern · 1,267 · 63%

Southern · 710 · 36%

Other · 23 · 1%

4. Data Recording Environment

Recording Environment

The recording took place in a quiet indoor setting, free from noticeable background voices or other noise, and without any echoes. Speakers read aloud the provided transcripts at their normal speaking pace.

Recording Equipment

The equipment included a high-fidelity microphone and recorder, along with smartphones. The database stores data in two formats: high-fidelity recordings at 44.1 kHz, 16-bit mono, and smartphone recordings at 16 kHz, 16-bit mono.

Recording Method

Speakers sat approximately 20 cm away from the high-fidelity microphone, speaking at a normal volume and pace while reading the provided transcripts. Android and iOS smartphones were positioned about 20 cm away from the microphone respectively.

5. Annotation and Transcription Guidelines

Transcribers must faithfully reproduce the audio content, ensuring complete consistency between the written text and spoken words. Key guidelines include: 1) Transcriptions must exactly match the audio; no extra words, omissions, or errors are permitted. 2) Numbers should be spelled out in Chinese characters, e.g., “one two three,” rather than “123.” Pay special attention to distinguishing “one” from “幺,” and “two” from “两.” 3) English words appearing in sentences should be transcribed phonetically, such as “thank you.” Letters pronounced individually should be capitalized and separated by spaces, e.g., “N B A,” “U F O.” 4) Symbols within sentences should be transcribed according to the speaker’s pronunciation, e.g., “three W dot baidu dot com.” Unpronounced symbols should be omitted. Brand names and proper nouns should follow standard conventions, such as “QQ Space,” “iPhone,” “Himalaya.” 5) All annotations must fully reflect the original pronunciation without any deletions.

6. Product Directory Structure

Directory Structure

The directory structure for this dataset is organized as follows:

└─DOC（Textual Documentation）
││├─wav_list.txt（Audio File List）
││├─content.txt（Speech Transcript）
││├─spk_info.xlsx（Speaker Information）
│
└─SPEECHDATA（Data Folder）
│├─C0001（Speaker Folder）
││ ││ ANDROID（Device Folder）
││ ││ AC0001W0001.wav（Audio File）
││ ││ AC0001W0001.txt（Speech Transcript）
││ ││ IOS（Device Folder）
││ ││ IC0001W0001.wav（Audio File）
││ ││ IC0001W0001.txt（Speech Transcript）
││ ││ MIC（Device Folder）
││ │ MC0001W0001.wav（Audio File）
││ │ MC0001W0001.txt（Speech Transcript）

Naming Conventions

CORPUS/USAGE/SPEAKER_NUM/EQUIPMENT_ID/SPEECH_IDe.g./SPEECHDATA/C0001/ANDROID/ H0001A0001.wav

Directory Name · Content · Notes

CORPUS · · Voice Database Identifier

USAGE · SPEECHDATA · Folder Name

SPEAKER_NUM · C0001 · Speaker Folder

EQUIPMENT_ID · ANDROID/IOS/MIC · Device Folder

SPEECH_ID · AC0001W0001.wav · WAV File

Data Sample Download

Click to download a sample dataset