Vietnamese speech dataset

1. Product Overview

A Vietnamese speech database totaling 100 hours, with a sampling rate of 16,000 Hz and a bit depth of 16 bits. The speakers are native Vietnamese speakers from Hanoi and Ho Chi Minh City, aged between 7 and 70 years old. Transcribed and annotated by professional speech annotators, the data has undergone rigorous quality assurance, achieving a text accuracy rate exceeding 95%.

2. Text Design

Text Domains

Considering real-world application scenarios, the text domains cover a broad range, primarily including:

• General Scenarios: everyday conversations, family interactions, news broadcasts, voice-controlled commands, human-computer interaction.
• Commercial Scenarios: e-commerce, catering, finance, logistics, meetings.
• Industrial Scenarios: agriculture, electronics manufacturing, automotive, textiles.
• Public Service Scenarios: government consultations, hospital consultations, school teaching.
• Complex Noise Scenarios: street scenes, buses, subways, conference rooms, indoor home environments.

Given the wide age range of speakers and the diverse application contexts, text domains have been tailored to suit different speaker groups as follows:

• Children: Primarily focused on command-based interactions, supplemented by simple conversational texts, covering home, in-car, human-computer interaction, and general daily conversation.
• Young and Middle-Aged Adults: This group represents the highest usage rate and the broadest coverage, encompassing home, in-car, human-computer interaction commands, dining, finance, logistics, medical consultations, industrial manufacturing, education, and casual daily chats.
• Seniors: Their domain lies between children and young/middle-aged adults, covering home, in-car, human-computer interaction commands, dining, news, medical consultations/healthcare, entertainment, travel, and casual daily conversations. For simple command phrases, fast, medium, and slow speaking speeds are distinguished.

The specific number of sentences assigned to each category is as follows:

Domain · Children · Young & Middle-Aged · Seniors

Home Commands · 25*3 · 5*3 · 20*3

In-Car Commands · 25*3 · 5*3 ·

Human-Computer Interaction Commands · 25*3 · 5*3 · 20*3

Family Conversations · 50 · ·

Dining · · 100 · 55

Finance · · 90 ·

Logistics · · 100 ·

Medical Consultations/Healthcare · · 100 · 55

News · · · 55

Entertainment · · · 55

Travel · · · 55

E-Commerce · · 100 ·

Industrial Manufacturing · · 80 ·

Education and Teaching · · 80 ·

General Daily Texts · 50 · 105 · 55

Summary · 325 · 800 · 450

In addition to reading aloud, children’s recordings also include guided dialogues. The thematic coverage for each individual is as follows:

Domain · Example · Hours

Family Conversation (Parent-Child) · Parent: “What’s your favorite game? Do you enjoy playing it?” Child: “I love building with blocks—when I finish constructing a castle, I feel so accomplished!” · 0.1

Dialogue · ·

Educational Interaction (Teacher-Student) · Teacher: “How did you approach this math problem? Is there anything you don’t understand?” Child: “I can do addition, but I always get subtraction mixed up. Could you explain it again, please?” · 0.1

Casual Chat (Classmates, Friends, etc.) · A: “That game today was really fun—let’s play together again tomorrow.” B: “I might not be able to make it tomorrow, but how about the day after?” · 0.05

Transcription Guidelines

Transcribers draft the text based on the audio they hear, striving to ensure complete consistency between the written content and the spoken words. General guidelines are as follows:

1. The transcribed text must exactly match the spoken audio; no extra words, missing words, or errors are permitted.
2. When symbols are pronounced verbally, half-width symbols should be retained, with one space on either side—for example, “Giảm 10 % giá sản phẩm,” where the percent sign is attached directly to the number without a space.
3. Proper English abbreviations should retain their full original spelling, such as USB or Wi‑Fi; when consecutive letters are pronounced, insert spaces—for example, “A B C.”
4. Basic punctuation marks follow Western conventions.

3. Directory Structure and Naming Conventions

Directory Structure

The data directory tree structure is as follows:

└─DOC (Documentation Files)
││├─content.txt (List of Transcribed Content)
││├─spk-info.txt (Speaker Information Table)
└─SPEECHDATA (Audio Folder)
││├─V0001 (Speaker ID)
││ │├─V0001W0001.wav (Audio File)

Naming Conventions

CORPUS/USAGE/SPEAKER_ID/SPEECH_IDe.g./SPEECHDATA/V0001/V0001W0001.wav

Directory Name · Content · Notes

CORPUS · · Speech Database Identifier

USAGE · SPEECHDATA · Audio Folder

SPEAKER_ID · V0001 · Folder for Speaker ID

SPEECH_ID · V0001W0001.wav · Audio File

4. Speaker Information

Gender Information

The database contains a total of 123 speakers, including 60 males and 63 females, as shown in the table below:

Gender · Male · Female · Total

Proportion · 48.8% · 51.2% · 100%

Age Information

The ages of the speakers span from 7 to 70 years old, with the following distribution:

Age Group · Number of Speakers · Percentage (%)

7–18 Years Old · 10 · 8.1%

18–35 Years Old · 105 · 85.4%

36–60 Years Old · 6 · 4.9%

61–70 Years Old · 2 · 1.6%

Total · 123 · 100%

5. Recording Plan

Recording Environment

Speakers record in homes or other quiet indoor settings, free from significant background noise.

Recording Method

Speakers use self-developed recording tools for close-range recording. Recording formats include reading aloud, natural pronunciation, and dialogue:

• Reading Aloud: Reading pre-assigned texts.
• Natural Pronunciation: Given a scenario and text, speakers produce sentences naturally and fluently without strict adherence to the provided text.
• Dialogue: Guided conversational exchanges.

Data Sample Download

Click to download a sample dataset