Taiwanese Hokkien Synthetic Speech Dataset from Qwen3-TTS

33,228 synthetic audio clips for Taiwanese Hokkien text-to-speech, generated using the Qwen3-TTS-1.7B-Base model with voice cloning. The dataset was created by lianghsun and last updated in March 2026.

Use Cases

• Train or fine-tune text-to-speech models using the `text` and `audio` columns.
• Benchmark voice cloning techniques using the `speaker_id` and `seed_audio_id` fields.
• Analyze the relationship between audio `duration` and linguistic content in the `text` field.
• Study speaker characteristics in synthetic audio by grouping samples by `speaker_id`.

Strengths

• Contains 33,228 audio-text pairs (722 from TAT seed + 32,506 from Common Voice seed).
• Audio is standardized at a 12000 Hz sample rate.
• Provides traceability to source data via `seed_audio_id` and `seed_text_id`.

Limitations

• All audio is synthetically generated, not natural human speech.
• Dataset size is moderate for TTS training; the 722 TAT seed samples represent a limited speaker pool.
• Potential quality limitations inherited from the voice cloning model and seed data sources.

Provenance

Source

Text from lianghsun/tw-hokkien-seed-text; voice seeds from lianghsun/tat_open_source and OKHand/Clean_Common_Voice_Speech_24.0-TW.

Collection Method

Synthesized using Qwen/Qwen3-TTS-12Hz-1.7B-Base model with x-vector voice cloning.

Freshness

Last updated March 2026.

Geography

Taiwan (Taiwanese Hokkien language).