Meeting data

1. Product Overview

The conference speech audio database contains 639 meetings, totaling 381 effective hours. The recording language is Chinese, and the recordings were made in China. The content covers business, daily life, work, and other topics. Primarily featuring speakers from northern China, 162 speakers participated in the recording process. Recordings were conducted in real meeting environments using equipment including headset microphones, one actual meeting voice recorder, high-fidelity microphones, Android tablets, iOS phones, a 16-microphone array, and a 16-microphone circular microphone array. Audio files are stored at 16 kHz, 16-bit resolution. This database has been transcribed and annotated by professional speech annotators and rigorously quality-checked, achieving a text accuracy rate of over 95%.

2. Scenarios and Equipment

Collection Scenarios

The collection encompasses three types of meeting scenarios: small, medium, and large. Each scenario involves between 3 and 10 participants. Specific requirements for each scenario are listed below:

No. · Scenario Type · Scene Size · Number of Scenes (pcs)

1 · Small · ≥20m² room · 8

2 · Medium · 20m² < room ≤ 50m² · 8

3 · Large · Room > 50m² · 4

Meeting scene noise is categorized into two types:

No. · Type · Content

1 · Inherent Noise · Keyboard typing, fan/air conditioner sounds, minor background voices, mobile phone ringtones

2 · Artificially Added Noise Sources · Sirens, etc.

Real-world example scenarios:

Collection Equipment

Considering the unique characteristics of meeting scenarios and multi-person conversations, the data storage format is set to 16 kHz, 16-bit. We selected the following devices for recording:

No. · Equipment

1 · iFLYTEK Hearing M1

2 · Android Tablet

3 · Headset Microphone

4 · High-Fidelity Microphone

5 · iOS Phone

6 · 16-Microphone Array

7 · 16-Microphone Circular Array

3. Collection Method

In meeting scenarios, audio content is collected with real projects and business as the core focus. Participants naturally engage, with no restrictions on accent, speaking speed, or bilingualism, ensuring authenticity. Data is gathered by field staff at designated locations, with equipment arranged appropriately. ※ Small Meeting Room Audio Collection Location Diagram: ※ Medium/Large Meeting Room Audio Collection Location Diagram: Recorded data includes the following information:

Record Item · Content

Meeting Room · Type/Recording Time/Number of Participants

Recorder · Basic Personal Information (Gender/Age)

Recording Position · Device/Serial Number

Noise · Type of Noise

4. Recorder Information

Gender Ratio

The database comprises a total of 162 individuals, including 66 males and 96 females.

Gender · Male · Female · Total

Ratio · 41% · 59% · 100%

Age Distribution

The age range of recorders spans 18–64 years, with the following breakdown:

Age · Number of Individuals · Proportion

18–24 · 90 · 55.6%

25–40 · 37 · 22.8%

≥41 · 35 · 21.6%

Total · 162 · 100%

5. Annotation and Transcription Guidelines

Transcribers draft text based on recorded audio, striving to ensure consistency between written text and spoken content. Key principles include: 1) Transcriptions must exactly match the spoken words—no extra or missing words, nor errors. 2) Numbers should be written in Chinese characters, such as “一二三,” rather than “123.” Distinguish between “一” and “幺,” and “二” and “两.” 3) English words appearing in the audio should be transcribed phonetically, e.g., “thank you.” Letters pronounced individually should be rendered in uppercase followed by spaces, such as “N B A” or “U F O.” Note: No spaces between Chinese characters. For mixed Chinese-English contexts, insert spaces between English terms—for example, between letters or between letters and words. No spaces between Chinese characters and English words. 4) Symbols within sentences should be transcribed according to the speaker’s pronunciation, such as “三W点百度点com.” Unpronounced symbols should be omitted. Brand names and proper nouns should follow standard conventions, such as “QQ空间,” “iPhone,” or “喜马拉雅.” 5) The completeness of annotations must align with the original audio; no omissions allowed. 6) In overlapping speech situations, each speaker’s contribution is annotated separately. For instance, in a meeting with four speakers, there will be four distinct layers, each containing only that speaker’s utterances—even if overlapping occurs, only the overlapping portion attributed to the specific speaker is noted. Detailed scenario guidelines are as follows: ⚫ When multiple people speak simultaneously, both voices are annotated at the overlapping point and distinguished using an ampersand (&); ⚫ If one person’s entire sentence is unintelligible, their contribution is treated as background noise and not annotated; ⚫ If one speaker utters a full sentence but part of it remains unclear, only the clear portion is transcribed; ⚫ For repeated utterances by non‑main speakers, tone markers should be directly annotated—e.g., “嗯,” “啊,” “呵,” etc. {For main speakers, when clear tone markers from non‑main speakers appear, annotate them accordingly, using formats like &嗯& or &啊&; if the tone marker comes from the main speaker, simply mark it normally without adding an ampersand}[All tone markers must include the “口” radical, except for “诶.”] 7) Special symbols require separate annotation, detailed in the table below:

No. · Label · Description

1 · <sil> · Speaker hesitation, time delays, pauses, prolonged sounds

2 · <-> · Half‑tone stuttering

3 · <$> · Laughter

4 · <_> · Cutting off sound, truncating sentences at the beginning or end

5 · <%> · Coughing

6 · <#> · Background noise, such as breathing or yawning

7 · & · Overlapping sounds

6. Data File Directory

Directory Structure

The data directory tree outlines the hierarchical structure:

└─DOC（Textual Description Files）
│├─all_wav_list.txt（Audio List）
│├─spk_info.xlsx（Recorder Information）
│
└─L（Meeting Room Types）
│├─L_R003（Room ID）
│├─wav（Audio Folder）
││ ││ 20201129_L_R003S06C01.wav（Audio File）
│├─Textgrid（Text File Folder）
││ ││ 20201129_L_R003S06C01.TextGrid（Transcribed Text）
│├─L_R003.jpg（Meeting Scene Image）

6-1-1 Data Directory Structure

File Naming Rules

Directory Naming Rules

/<VENUE_TYPE>/<VENUE_ID>/<WAV>/<AUDIO_ID>e.g. L/L_R003/wav/20201129_L_R003S06C01.wav

Recording · Content · Notes

VENUE_TYPE file · Small / Medium / Large · Meeting Scene Type

VENUE_ID file · L_R003 · Venue ID

AUDIO file · wav/20201129_L_R003S06C01.wav · WAV File

File Naming Rules

/<TIME_ID>/<VENUE_TYPE>/<VENUE_ID>/<MEETING_ID>/<DEVICE_ID>.wave.g. 20201129_L_R003S06C01.wav

File · Content · Notes

TIME_ID · 20201129 · Recording Time

VENUE_TYPE · _S / M /L_ · Venue Type

VENUE_ID · R001~500 · Venue ID

MEETING_ID · S01 ~ 08 · Meeting ID

DEVICE_ID · C01 & C02 · Device ID

Equipment Information

Equipment is evenly distributed among recorders, with numbering progressing from north to south.

No. · Equipment · Serial Number

1 · iFLYTEK Hearing M1 · M01

2 · Android Tablet · T01 & 02

3 · Headset Microphone · A01

4 · High-Fidelity Microphone · H01

5 · iOS Phone · I01

6 · 16-Microphone Array · L01 & L02

7 · 16-Microphone Circular Array · C01 & C02

Data Sample Download

Click to download a sample dataset