AudioEnglish corpus

Smart Home Voice Data - English

About this dataset

1. Product Overview

The smart home voice database comprises a total of 315 hours of recordings. The recording language is English, and the recordings were made in the United States. The transcribed text covers mainstream smart home control scenarios, audio‑visual entertainment, wake words, and free‑form text. A total of 207 native American speakers participated in the recording process. Recordings were conducted in real‑world home environments, with simulated smart home appliances and applications serving as recording positions. Four high‑fidelity microphones (44.1 kHz, 16-bit) were used simultaneously to capture the audio.

2. Product Catalog Structure

Catalog Structure

Data directory structure:

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

Naming Conventions

CORPUS/CORPUS_ID/SPEAKER_NUM/SPEECH_ID e.g., /SPEECHDATA/0001/H0001A/ H0001A0001.wav

Directory Name · Content · Notes

CORPUS · · Voice database ID

USAGE · SPEECHDATA · Folder name

FILE_ID · 0001 · Speaker folder name

POINT_ID · H0001A · Recording position number

SENTENCE_ID · H0001A0001.txt · TXT file

SPEECH_ID · H0001A0001.wav · WAV file

3. Text Design

Corpus Creation

The corpus includes mainstream smart home control scenarios, audio‑visual entertainment, wake words, and free‑form text. Following established rules, a corpus pool was generated and processed. • De-identification: Politically sensitive, personally private, or sexually violent content was removed. Symbols such as <, >, [, ], ~, /, \, = were also deleted. • Content containing languages other than English was excluded. • Uniform formatting was applied.

Text Structure

To ensure comprehensive speech coverage and phoneme balance, each recording session was designed with 500 sentences drawn from the corpus pool, structured as follows:

Serial Number · Content · Per Session Allocation/Number of Sentences

1 · Smart Home Control · 10

2 · Audio‑Visual Entertainment · 90

3 · Wake Words · 50

4 · Free‑Form Text · 350

Total · 4 Categories · 500 Sentences

4. Recording Environment

Recording Site

1. The recording scene replicates a real‑world home environment, including basic household items, furniture, and appliances. 2. Recording positions are set at four distances: 0.3 m, 1 m, 3 m, and 5 m. 3. After finalizing the recording location and positions, equipment is calibrated, and preparatory steps such as actual recording are completed to ensure that all devices function properly, remain stable, and meet recording requirements. 4. Speakers record their voice data while seated at designated positions.

Recording Equipment

Recording equipment consists of high‑fidelity microphones and a recorder. Data storage format for this database is high‑fidelity recording at 44.1 kHz, 16-bit mono.

Recording Method

In a home setting, four recording positions are arranged: one close‑range position and three functional positions. The close‑range position is 30 cm away from the speaker, while the functional positions are fixed at 1 m, 3 m, and 5 m, simulating typical placements of various household appliances. Speakers deliver recorded texts at normal speaking volume and pace.

Serial Number · Content

A · Close‑Range Position: 0.3 m

B · Functional Position: 1 m

C · Functional Position: 3 m

D · Functional Position: 5 m

5. Transcription Annotation Guidelines

Transcribers draft text based on the audio they hear, striving to maintain consistency between written content and spoken pronunciation. General guidelines include: 1) Transcriptions must exactly match the spoken audio—no extra words, missing words, or errors. 2) Numbers should be transcribed in English form—for example, “one two three” rather than “123”—reflecting the actual pronunciation, such as “one hundred twenty three” or “twelve and three.” 3) Symbols within sentences should be transcribed according to the speaker’s pronunciation—for instance, “triple W dot Google dot com.” Unpronounced symbols should be omitted. Brand names and proper nouns should follow standard conventions, such as “iPhone” or “PayPal.” 4) Letters pronounced individually should be transcribed in uppercase with spaces—for example, “N B A” or “D N A.” 5) Annotation completeness must align with actual pronunciation; no omissions allowed.

6. Speaker Information

Basic Information Records

Speaker information records include task ID, gender, age, and recording scenario.

Task ID · Gender · Age Range · Scenario

0001 · M · B · Scenario 1

Task ID: Each speaker receives one task ID, corresponding to one transcription assignment. Each speaker may participate in only one recording session. Gender: Male or female. Accent region: Speakers are categorized by their native regional accent—North, South, or Other. Age range: A (16–25), B (26–40), C (41+). Recording scenarios: There are four distinct settings, with room details listed below.

Scenario · Number of People · Room Dimensions

Scenario 1 · 27 · 15 m long × 20 m wide × 3.5 m high

Scenario 2 · 20 · 6 m long × 4 m wide × 3 m high

Scenario 3 · 142 · 5.5 m long × 4 m wide × 2.8 m high

Scenario 4 · 11 · 5.2 m long × 3.5 m wide × 3.2 m high