Clean Common Voice 24.0: 10K-100K Taiwan Chinese Voice Seeds

Encompassing between 10,000 and 100,000 cleaned audio clips derived from the Mozilla Common Voice 24.0 Chinese (Taiwan) corpus. Released by OKHand in early 2026, it provides 'Voice Seeds' processed through Silero-VAD to remove silence and environmental noise for generative speech tasks.

Use Cases

• Training text-to-speech (TTS) models using the cleaned audio and text pairs
• Voice cloning applications requiring high-quality voice seeds without background silence
• Fine-tuning speech recognition models specifically for the Chinese (Taiwan) dialect

Strengths

• Processed with Silero-VAD (0.5 threshold) to ensure clean voice segments
• Optimized Parquet format for efficient data streaming and storage
• Scale of 10,000 to 100,000 high-quality voice seeds

Limitations

• Automated cleaning via Silero-VAD may introduce minor audio artifacts or clip speech in rare edge cases
• Geographic bias toward the Taiwan dialect of Chinese, which may not represent other Mandarin variants

Provenance

Source

Mozilla Common Voice 24.0

Collection Method

Automated processing of existing crowdsourced audio using Silero-VAD for silence removal.

Freshness

Last updated March 2, 2026.

Geography

Taiwan