HI-MIA

Dataset Description

HI-MIA is a far-field text-dependent speaker verification database used in the AISHELL Speaker Verification Challenge 2019.

The data is extracted from a larger database called AISHELL-WakeUp-1. The original resource contains wake-up words "Hi, Mia" in both Chinese and English. The challenge data provided in this resource uses the Chinese wake-up words.

The recordings were collected in real home environments using microphone arrays and a Hi-Fi microphone. The challenge data is extracted from:

• 1 Hi-Fi microphone
• 16-channel circular microphone arrays
• Recording distances of 1 meter, 3 meters, and 5 meters

Dataset Source

The dataset is available from OpenSLR:

• OpenSLR: http://openslr.org/85/
• External URL: http://aishelltech.com/wakeup_data

Dataset Structure

The dataset is divided into three subsets:

Split	Number of Speakers	Description
Train	254	Training set with speaker-dependent subfolders
Dev	42	Development set with speaker-dependent subfolders
Test	44	Test set with paired target/non-target answer for speaker verification evaluation

OpenSLR provides the following downloadable files:

File	Size	Description
train.tar.gz	36 GB	Training set with speaker-dependent subfolders
dev.tar.gz	5.1 GB	Development set with speaker-dependent subfolders
test.tar.gz	4.7 GB	Test set with target/non-target answer
test_v2.tar.gz	4.7 GB	Updated test set fixing corrupted audio files
filename_mapping.tar.gz	5.9 MB	Filename mapping rules for multi-channel information

Dataset Creation

The data was collected in real home environments. The collection process and the development of baseline systems are described in the paper cited below.

Intended Uses

This dataset is intended for research on:

• Far-field speaker verification
• Text-dependent speaker verification
• Wake-up word speaker verification
• Microphone-array based speaker verification

License

The dataset is released under the Apache License v2.0.

Citation

If you use this dataset, please cite:

@misc{himia,
    title={HI-MIA : A Far-field Text-Dependent Speaker Verification Database and the Baselines},
    author={Xiaoyi Qin and Hui Bu and Ming Li},
    year={2019},
    eprint={1912.01231},
    archivePrefix={arXiv},
    primaryClass={cs.SD}
}