Zh-Taiwan: Traditional Chinese Speech Dataset for TTS and ASR

2,740 audio recordings at 16 kHz form a dataset for traditional Chinese speech synthesis and recognition. Each entry includes the audio, its length, traditional Chinese text, and a corresponding normalized simplified Chinese text. The dataset, originally from ivanzhu109/zh-taiwan, is mirrored and formatted by lianghsun.

Use Cases

• Train text-to-speech models based on the provided 16 kHz audio and traditional Chinese text pairs.
• Train automatic speech recognition models based on the audio and its corresponding text transcriptions.
• Benchmark model performance across traditional and simplified Chinese scripts using the parallel normalized_text field.
• Study the impact of text normalization on speech synthesis or recognition for Chinese languages.

Strengths

• Contains 2,740 audio-text pairs, split into train (2,698), validation (14), and test (28) subsets.
• Provides parallel text in both traditional and normalized simplified Chinese, enabling cross-script applications.
• Audio is recorded at a standard 16 kHz sampling rate suitable for speech processing tasks.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count for the primary data splits is known, but the total number of rows per column is unspecified.
• The validation and test sets are very small (14 and 28 samples), which may limit robust evaluation.

Provenance

Source

Original source is ivanzhu109/zh-taiwan; this repository is a mirror and format-organized version by lianghsun.

Collection Method

Likely gathered from speech recordings paired with text transcriptions.

Freshness

Last updated 2026-04-14 23:00:17; freshness should be verified.

Geography

The dataset name suggests a focus on Taiwanese Mandarin or Traditional Chinese, but specific geographic coverage is not detailed.