Far-field voiceprint speech data
About this dataset
1. Product Overview
The voice database comprises a total of 20,253 hours of recordings. The recording language is Chinese, and the recordings were collected in China. A total of 511 speakers were invited, each recording three separate sessions on different days, spanning 7 to 25 days. The speakers range in age from 17 to 66 years old. This database can be used for research in areas such as speaker recognition and speech recognition. All recordings have been transcribed and annotated by professional linguists and rigorously quality-checked, achieving a text accuracy rate exceeding 95%.
2. Recording Corpus
Corpus Pool Construction
Design of Recording Texts
This database contains approximately 330 sentences per recording session, with each speaker contributing about 990 sentences across three sessions. The recorded content covers wake-up phrases and related interactive requests at fast, medium, and slow speaking rates; numerical sequences; non-repetitive multi-domain vocabulary; and natural conversational exchanges. The specific structure of the results is as follows:
Domain · Content · ID
Wake-Up · Hello, Miya (fast) · 0001–0010
Wake-Up · Hello, Miya (medium) · 0011–0020
Wake-Up · Hello, Miya (slow) · 0021–0030
Wake-Up Interaction · Hello, Miya + Interaction (fast) · 0031–0050
Wake-Up Interaction · Hello, Miya + Interaction (medium) · 0051–0070
Wake-Up Interaction · Hello, Miya + Interaction (slow) · 0071–0090
Numbers · Consecutive number strings · 0091–0110
Mobile & Tablet Voice Control · Operational control commands · 0111–0170
Speaker Interaction · Voice control and commands for smart speakers · 0171–0210
Domain-Independent · Broad-range texts · 0211–0270
Natural Conversation · Natural interactive sentences · 0271–0330
· Total · 330
Corpus Processing
- De-identification: Removal of politically sensitive, personally private, or sexually explicit/violent content.
- Removal of symbols such as <, >, [, ], ~, /, \, =, except for numeric strings.
- Exclusion of content containing languages other than Chinese and English.
- Elimination of single sentences exceeding 25 characters.
- Standardization of formatting.
3. Speaker Information
Speaker Demographic Characteristics
Gender Ratio
The database includes a total of 511 speakers—211 males and 300 females.
Gender · Male · Female · Total
Ratio · 41.3% · 58.7% · 100%
Age Distribution
The speakers span ages from 17 to 66, distributed as follows: Group A (16–25 years), Group B (26–40 years), and Group C (41 years and older).
· Age Group · Number of Speakers · Percentage
Group A · 16–25 years · 481 · 94%
Group B · 26–40 years · 15 · 3%
Group C · 41 years and older · 15 · 3%
Total · 100 · 100% ·
4. Data Recording Environment
Recording Environment
A quiet indoor setting with no echoes, supplemented by household and conference noise depending on the scenario. Speakers read the recorded texts aloud at slow, normal, and fast speeds.
Recording Equipment
The equipment includes high-fidelity audio devices, Android smartphones, iOS smartphones, Android tablets, and a circular microphone array with a 5 cm radius.
Recording Method
A total of seven recording positions are utilized, including close-range recording, 1.5 meters behind, 0.25 meters away, 1 meter away, and 3 meters left and right. The device placement diagram is shown below:
Note: Noise1 refers to fan noise; Noise2 refers to TV, music, and office noise.
5. Speech Data Transcription
Transcribers convert audio into written text, striving to ensure consistency between the written content and the spoken audio. General guidelines include: 1) The transcription must exactly match the spoken words, without adding, omitting, or misrepresenting any words. 2) Numerals should be transcribed in Chinese characters—for example, “one two three” rather than “123.” Pay special attention to distinguishing “one” from “幺,” and “two” from “两.” 3) Any English pronunciation in the audio should be rendered in corresponding Chinese characters or retained in English. Specific cases include: all letters or words in URLs should be capitalized—for example, “www.abc.com” should be transcribed as “Three W dot A B C dot com”; English words within the audio should be transcribed entirely in lowercase; English letters should be transcribed entirely in uppercase. For certain proper nouns or acronyms that are capitalized, add spaces—for example, “C E O,” “C C T V,” etc. 4) The completeness of annotations must align with actual pronunciation, with no deletions allowed.
6. Data File Directory
Directory Structure
The data directory tree is structured as follows:
└─DOC（Textual Description Files）
││├─wav_list.txt（Audio List）
││├─spkrinfo.xlsx（Speaker Information）
│
└─SPEECHDATA（Data Folder）
│├─0548 (Speaker ID)
│├─F0548（First Recording）
│├─548A0.25M（Equipment and Position Information）
││ ││ F0548_548A0.25M_1_0019_normal.wav（Audio File）
│├─F0548.txt（Speech Content Text）
│├─S0548（Second Recording）
│├─T0548（Third Recording）
Naming Rules
Directory Naming Convention
<USAGE>/<SpeakerID>/<Time>/<Recording_info>/<wav_ID>e.g.SPEECHDATA/0548/F0548/548A0.25M/F0548_548A0.25M_1_0019_normal.wav
Directory · Content · Notes
USAGE · SPEECHDATA · Folder Name
SpeakerID · 0548 · Speaker ID
Time · F0548 · Recording Session Number
Recording_info · 548A0.25M · Recording Details
wav_ID · F0548_548A0.25M_1_0019_normal.wav · Audio File
File Naming Convention
<Time><Speaker_ID><Device_ID><Point_ID><Channel_ID><WAV_NUM><Speed>e.g. F0548_548A0.25M_1_0019_normal.wav
File · Content · Notes
Time · First（F）、Second（S）、Third（T） · Recording Sessions
SPEAKER_ID · 0001–0550 · Speaker IDs
Device_ID · A: Android Phone, I: iOS Phone, MIC: High-Fidelity Microphone, PAD: Android Tablet, PCM: 16-Mic Circular Array · Device Information
Point_ID · 0.25M、-1.5M、1M、3M、L3M（Left）、R3M（Right）、5M · Position Details
Channel · 1–16 · Channel Information
WAV_NUM · 0019 · WAV Number
Speed · normal、fast、slow · Speaking Speed
Data Sample Download
Samples
- Audio sample 1
- Audio sample 2