ChildMandarin: A Comprehensive Mandarin Speech Dataset for Young Children Aged 3-5

⭐ Introduction

This repository contains the ChildMandarin dataset, a comprehensive Mandarin speech dataset specifically designed for young children aged 3 to 5. This dataset aims to address the scarcity of resources in this area and facilitate research in child speech recognition, speaker verification, and related fields.

🚀 Dataset Details

• Age Range: 3-5 years old
• Total Duration: 41.25 hours
• Number of Speakers: 397
• Geographic Coverage: 22 out of 34 provincial-level administrative divisions in China
• Gender Distribution: Balanced across all age groups
• Recording Devices: Smartphones (Android and iPhone)
• Recording Environment: Quiet indoor environments
• Annotation: Character-level manual transcriptions, age, gender, birthplace, device, accent level.
• Content: Unrestricted, focusing on age-appropriate daily communication.
• Data Format: WAV PCM, 16kHz sampling rate, 16-bit precision

Dataset Statistics

Split	# Speakers	# Utterances	Duration (hrs)	Avg. Utterance Length (s)
Train	317	32,658	33.35	3.68
Dev	39	4,057	3.78	3.35
Test	41	4,198	4.12	3.53
Sum	397	40,913	41.25	3.52

More details could be found in our paper ChildMandarin

📐 Experiments

We conducted experiments on Automatic Speech Recognition (ASR) and Speaker Verification (SV) tasks to evaluate the dataset.

1️⃣ ASR Results

Models Trained from Scratch

Encoder	Loss	# Params	Greedy	Beam	Attention	Attention Rescoring
Transformer	CTC+AED	29M	34.55	34.4	40.61	32.15
Conformer	CTC+AED	31M	28.73	28.72	31.60	27.38
Conformer	RNN-T+AED	45M	37.11	37.14	33.84	37.14
Paraformer	Paraformer	30M	31.86	28.94	-	-

Fine-tuned Pre-trained Models

Model	# Params	Zero-shot	Fine-tuning
CW	122M	18.05	13.66
Whisper-tiny	39M	67.63	28.78
Whisper-base	74M	51.49	23.33
Whisper-small	244M	37.99	17.45
Whisper-medium	769M	28.55	18.97
Whisper-large-v2	1,550M	29.43	-

More Pre-trained Models

Model	# Params	Zero-shot
Qwen-Audio	7.7B	20.39
Qwen2-Audio	8.2B	11.54
SenseVoice (Small)	234M	11.89

2️⃣ SV Results

Model	# Params	Dim	Dev (%)	EER (%)	minDCF	EER (%)	minDCF
x-vector	4.2M	512	75.4	8.91	0.7198	25.92	0.9780
ECAPA-TDNN	20.8M	192	84.6	13.72	0.8697	27.77	0.9490
ResNet-TDNN	15.5M	256	91.9	9.57	0.6597	22.11	0.9044

🤗 Dataset Download

You can access the ChildMandarin dataset on HuggingFace Datasets:

https://huggingface.co/datasets/BAAI/ChildMandarin

📚 Cite me

@article{zhou2024childmandarin,
  title={ChildMandarin: A Comprehensive Mandarin Speech Dataset for Young Children Aged 3-5},
  author={Zhou, Jiaming and Wang, Shiyao and Zhao, Shiwan and He, Jiabei and Sun, Haoqin and Wang, Hui and Liu, Cheng and Kong, Aobo and Guo, Yujie and Qin, Yong},
  journal={arXiv preprint arXiv:2409.18584},
  year={2024}
}