Vietnamese Voice Dataset

A Vietnamese voice dataset for text-to-speech (TTS) application

Introduction

We present a Vietnamese voice dataset for text-to-speech (TTS) application. The dataset contains 619 minutes (~10 hours) of speech data, which is recorded by a southern vietnamese female speaker. This dataset is recorded in a controlled environment with professional recording tools. The dataset is available for research purpose only.

Data preparation

The dataset is recorded in a controlled environment with professional recording tools, then processed by our home-build tool for annotation. The annotation process includes:

• Transcription: the text is transcribed by a professional transcriber.
• Alignment: the text is aligned with the audio by a professional aligner.
• Elimination: the bad audio part is removed by an user to make the dataset clean.

Data format

The dataset is organized in the following structure:

├── data
│   ├── wav
│   │   ├── 000001.wav
│   │   ├── 000002.wav
│   │   ├── 000003.wav
........
│   ├── metadata.json

The metadata.json file contains the following information:

{
    "000001": {
        "text": "Hôm nay là thứ hai",
        "audio": "wav/000001.wav",
        "n_tokens": 5,
        "duration": 2.0
    },
    "000002": {
        "text": "Hôm nay là thứ ba",
        "audio": "wav/000002.wav",
        "n_tokens": 5,
        "duration": 2.0
    },
    ...
}

Data statistics table

Dataset	text	audio	duration	n_tokens
count	10233	10233	10233	10233
mean	-	-	3.634935	16.557412
std	-	-	2.164213	9.852908
min	-	-	0.128027	1.0
25%	-	-	1.920000	9.0
50%	-	-	3.168027	15.0
75%	-	-	4.960000	22.0
max	-	-	18.210000	74.0
sum	-	-	37196.2859	169432

Samples

Here are some samples from the dataset:

• 10000.wav
• 10001.wav

Application

• To prove that this dataset is useful for TTS application, we have trained a TTS model using this dataset and create an app (baonoi.ai) using this model.
• Sample generated audio: g01.wav

Download

Please contact ai-team@codelink.io if you wish to access this dataset