Smart Home Voice Data – Chinese

1. Product Overview

The smart home voice database comprises a total of 1,852 hours of recordings. The spoken language is Chinese, and the recordings were made in China. The transcribed text covers 11 categories, including intelligent control and commands for mainstream home scenarios. Primarily recorded in northern Chinese accents, the dataset involved 646 speakers. Recording took place in real‑world home environments, simulating the use of smart home appliances and applications, with seven recording positions set up. Among these, five used high‑fidelity microphones (44.1 kHz, 16-bit; each microphone recorded between 217 and 284 hours), one used an iOS smartphone (16 kHz, 16-bit; 282 hours), and another used an Android smartphone (16 kHz, 16-bit; 281 hours). This database has been professionally transcribed and annotated by skilled linguists and underwent rigorous quality assurance, achieving a text accuracy rate exceeding 95%.

2. Product Catalog Structure

Catalog Structure

Data catalog structure

└─DOC (textual documentation)
││├─wav_list.txt (audio file list)
││├─content.txt (speech text content)
││├─spk_info.xlsx (speaker information)
││├─room_info.xlsx (scene information)
│
└─SPEECHDATA (data folder)
│├─0001 (speaker folder)
││ ││ H0001A (audio folder)
││ ││ H0001A0001.wav (audio file)
││ ││ H0001B (audio folder)
││ ││ H0001B0001.wav (audio file)

Naming Conventions

CORPUS/USAGE/SPEAKER_NUM/POINT_ID/SPEECH_IDe.g./SPEECHDATA/0001/H0001A/ H0001A0001.wav

Directory Name · Content · Notes

CORPUS · · Voice database ID

USAGE · SPEECHDATA · Folder name

SPEAKER_NUM · 0001 · Speaker folder name

POINT_ID · H0001A · Recording position number

SPEECH_ID · H0001A0001.wav · WAV file

3. Text Design

Corpus Construction

Considering the application of speech recognition in areas such as voice control of smart devices, command words, and current affairs texts within mainstream home scenarios, the corpus was selected across 11 domains. The corpus pool was processed according to established rules.

Serial Number · Domain

1 · Robots

2 · Air conditioners

3 · Televisions

4 · Smart speakers

5 · Controllers

6 · Table lamps

7 · Digital category

8 · Music category

9 · TV and film category

10 · Radio category

11 · Current affairs category

• De-identification processing: Removal of politically sensitive, personally private, or sexually violent content.
• Removal of symbols such as <, >, [ , ] , ～ , / , \ , =.
• Removal of content containing languages other than Chinese and English.
• Removal of sentences longer than 25 characters.
• Standardization of format.

Text Structure

Considering speech coverage and phoneme balance, the database’s audio transcripts are designed with 520 sentences per file, drawn from the corpus pool, structured as follows.

Serial Number · Domain · Distribution per file/sentence

1 · Robots · 11

2 · Wake-up 1 · 13

3 · Controller · 25

4 · Table lamp · 20

5 · Wake-up 2-1 · 6

6 · Wake-up 2-2 · 6

7 · Wake-up 3 · 6

8 · Speaker · 25

9 · Wake-up, Command 4 · 50

10 · Radio · 5

11 · Air conditioner · 50

12 · Television · 40

13 · Music · 40

14 · Digital · 40

15 · Movies, Videos · 40

16 · Broad‑area textual content · 143

Total · 11 items · 520 sentences

4. Recording Environment

Recording Site

1. The recording environment consists of real‑world home settings, including basic household items, appliances, tables, and chairs. 2. Specific recording positions are determined based on the actual site conditions, including distance and height considerations. 3. Recorded distances vary between 0.5 m, 1.5 m, 2 m, 2.5 m, 3 m, 3.5 m, 4 m, and 5 m, depending on the scene. 4. After finalizing the recording location and positions, equipment is calibrated, and preparatory steps such as actual recording are completed to ensure that all recording devices operate normally, stably, and meet the required standards. 5. Speakers record their voice data at designated positions.

Recording Equipment

Recording equipment includes high‑fidelity microphones, tape recorders, and 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

In a home setting, seven recording positions are arranged, comprising three close‑range positions and four functional positions. Functional positions are fixed at air conditioners, televisions, table lamps, and sweeping robots, while close‑range positions are located 30 cm away from the speaker, equipped with one high‑fidelity microphone and two smartphones (iOS and Android). Speakers deliver their recordings at normal volume and pace, reading aloud the provided text.

Serial Number · Content

A · Recording position 1: Air conditioner

B · Recording position 2: Television

C · Recording position 3: Table lamp

D · Recording position 4: Sweeping robot and others

E · Recording position 5: Close‑range iOS

F · Recording position 6: Close‑range Android

G · Recording position 7: Close‑range microphone

5. Annotation and Transcription Guidelines

Transcribers draft text based on the audio they hear, striving to maintain complete consistency between written text and spoken content. Key guidelines include: 1) Transcriptions must exactly match the spoken audio—no extra words, missing words, or errors. 2) Numbers should be written in Chinese characters, e.g., “一二三,” rather than “123.” Pay special attention to distinguishing “一” from “幺,” and “二” from “两.” 3) English words appearing in sentences should be transcribed phonetically, e.g., “thank you.” Letters pronounced individually should be capitalized and separated by spaces, e.g., “N B A,” “U F O.” 4) Symbols appearing in sentences should be transcribed according to the speaker’s pronunciation, e.g., “三W点百度点com.” Unpronounced symbols should be omitted. Brand names and proper nouns should be transcribed in their commonly used forms, e.g., “QQ空间,” “iPhone,” “喜马拉雅.” 5) The completeness of annotations must align with the actual pronunciation; no omissions allowed.

6. Speaker Information

Basic Information Records

Speaker information records include task number, gender, accent region, age range, and scene environment.

Task Number · Gender · Accent Region · Age Range · Scene Environment

0001 · Male · Northern · B · Scene One

Each speaker receives one task number, corresponding to one transcription file. Each speaker may participate in only one recording session. Gender: male or female. Accent region: categorized into Northern, Southern, and Other based on the speaker’s native dialect. Age range: A (under 15), B (16–25), C (26–40), D (41+). Scene environment: Twelve recording scenes are defined, with scene details recorded in the attached room_info file.

Scene · Number · Scene · Number

Scene One · bedroom_1 · Scene Seven · diningroom_1

Scene Two · bedroom_2 · Scene Eight · drawingroom_1

Scene Three · bedroom_3 · Scene Nine · drawingroom_2

Scene Four · bedroom_4 · Scene Ten · drawingroom_3

Scene Five · bedroom_5 · Scene Eleven · drawingroom_4

Scene Six · bedroom_6 · Scene Twelve · drawingroom_5

Speaker Demographic Characteristics

Gender Ratio

The database contains a total of 646 speakers, including 276 males and 370 females.

Gender · Male · Female · Total

Ratio · 43% · 57% · 100%

Age Distribution

A (under 15): 0 people; B (16–25): 631 people; C (26–40): 12 people; D (41+): 3 people.

Age Range · Age Group · Number · Ratio · Male · Female

A · ＜15 years old · 0 · 0 · 0 · 0

B · 16–25 years old · 631 · 97.7% · 269 · 362

C · 26–40 years old · 12 · 1.9% · 6 · 6

D · ＞41 years old · 3 · 0.5% · 1 · 2

Total · 646 · 100% · 276 · 370

Dialect Distribution

North: 545 people; South: 91 people; Other: 10 people.

Region · Number · Ratio

North · 545 · 84.4%

South · 91 · 14.1%

Other · 10 · 1.5%

Total · 646 · 100%

Sample Data Download

Click to download sample data