SingMOS-v1: 3,421 Chinese and Japanese Singing Voice Clips

SingMOS-v1 is a preview version of the SingMOS-Pro dataset containing 3,421 Chinese and Japanese vocal clips. The clips have a sample rate of 16 kHz and total 4.25 hours in duration. The dataset was created by TangRain and last updated on Hugging Face in October 2025.

Use Cases

• Train models for singing voice quality assessment based on the provided audio clips.
• Benchmark MOS prediction algorithms using the structured train/dev/test splits mentioned in the description.
• Analyze acoustic features of Chinese and Japanese singing voices for quality evaluation.
• Develop cross-lingual audio quality models leveraging the bilingual nature of the dataset.

Strengths

• Contains 3,421 individual audio clips.
• Totals 4.25 hours of audio data.
• Provides structured splits for machine learning (train, dev, test).

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Data may reflect geographic bias inherent to its Chinese and Japanese source material.

Provenance

Source

TangRain on Hugging Face.

Freshness

Last updated 2025-10-02 07:04:08; freshness should be verified.

Geography

Likely China and Japan, based on language of vocal clips.