andrewbawitlung/MizoSpeech

Disclaimer / Notice: Details for these are in Peer Review and publications of the paper will be made available soon for more details.

Dataset Description

The MizoSpeech is an audio corpus designed for unsupervised pre-training and linguistic research on the Mizo language and its associated dialects. It encompasses 761,053 .wav audio recordings across 10 distinct dialects within the Kuki-Chin-Mizo branch of the Tibeto-Burman family.

This dataset is distributed in Parquet format, split into ~500 MB shards.

• Language(s): Falam, Gangte, Hmar, Lai, Lushai, Mara, Paite, Ralte, Vaiphei, Zou

Repository Structure

The primary dataset repository consists entirely of processed data:

• /data — Contains the entire dataset serialized into parquet format, split into approximately 500 MB shards (e.g., mizospeech-00000-of-00408.parquet). The raw audio bytes are securely embedded within these parquet chunks alongside their corresponding metadata, removing the need to manage thousands of raw .wav files locally while facilitating high-throughput streaming and loading.

Data Fields

Field Name	Type	Description
audio	AudioFeature	A dictionary containing the decoded audio waveform. Includes the internal path string, the decoded numeric array, and the sampling_rate.
path	string	The explicit directory path mapping to the original file structure (e.g., data/Falam_00001.wav).
language	string	The specific Mizo dialect spoken in the audio recording (Falam, Gangte, Hmar, Lai, Lushai, Mara, Paite, Ralte, Vaiphei, Zou).
duration	float32	The total duration of the audio clip in seconds. Highly useful for filtering extremely long or short clips during training.

Usage and Pre-processing

from datasets import load_dataset

dataset = load_dataset("andrewbawitlung/MizoSpeech", split="mizospeech")

Dataset Statistics

Dialect	Number of Utterances	Total Audio Duration (Hours)
Lushai	425,794	1,048.34
Hmar	118,833	276.43
Lai	107,289	264.37
Paite	53,162	126.89
Falam	24,923	62.02
Vaiphei	15,806	39.66
Mara	9,111	22.73
Gangte	5,428	13.71
Ralte	492	1.22
Zou	215	0.51
Total	761,053	1,855.88

Audio Specifications

• Format: .wav embedded in Parquet
• Sampling Rate: 16,000 Hz
• Channels: Mono (1)
• Bit Depth: 16-bit