TaiMECS: Taiwan Mandarin–English Code-Switching Sentences

A set of sentences in Taiwanese Mandarin with embedded English (code-switching), spoken by a single Taiwanese speaker. Designed for training and evaluating Taiwan-specific Mandarin speech models, with deliberate emphasis on words whose Taiwan readings differ from Mainland Mandarin.

Pronunciations are hand-curated and dictionary-verified with a focus on hard, easy-to-get-wrong cases (Taiwan/Mainland differences, polyphonic characters, common model misreadings), making this a small but high-quality resource for Taiwan-specific speech. See Pronunciation.

Composition

Clips	100 (001.mp3–100.mp3)
Audio	MP3, 48 kHz, mono, 256 kbps
Loudness	Normalized to −16 LUFS, true-peak −1 dBTP (EBU R128)
Duration	5–14 s each (~13 min total)
Speaker	1
human	20 — recorded by the speaker
synthetic	80 — generated with openbmb/VoxCPM2, using a voice cloned from the human recordings
Language	Traditional Chinese (zh-Hant, Taiwan) + English

All clips are loudness-normalized to −16 LUFS (integrated) with a −1 dBTP true-peak ceiling, using EBU R128 two-pass linear normalization (measure, then apply a single linear gain) so loudness is uniform across human and synthetic clips without altering dynamics.

Format

HuggingFace AudioFolder layout: audio files plus metadata.jsonl, with one line per clip:

{"file_name": "048.mp3",
 "text": "升息的循環還沒結束，房貸的 interest rate 一直在漲，很多人的還款壓力突然變得很大。",
 "source": "synthetic",
 "tts_model": "openbmb/VoxCPM2",
 "reference_audio": "001.mp3",
 "tts_text": "升{xi2}的循環還沒結束，房{dai4}的 interest rate 一直在{zhang3}，很多人的{huan2}款壓力{tu2}然變得很大。"}

Field	Meaning
file_name	Audio file
text	Clean transcript/label
source	human or synthetic
tts_model / reference_audio / tts_text	Provenance for synthetic clips: the model, the speaker’s clip used as the voice reference, and the inline-pinyin text actually fed to the TTS system (null for human clips)

Pronunciation

The text field is the clean transcript/label. The audio follows Taiwan readings, which were curated by hand with a deliberate focus on the hardest cases — situations where the correct Taiwan reading is easy to get wrong. This curation is what makes the dataset high-quality despite its small size.

For the synthetic clips, Taiwan readings were enforced with inline pinyin (e.g., 企 → {qi4}; recorded in the tts_text field), covering four categories of difficult cases:

• Taiwan vs. Mainland readings — characters whose standard Taiwan reading differs from Mainland Mandarin (e.g. 垃圾 lè sè not lā jī, 企業 qì not qǐ, 研究 jiù not jiū, conjunctive 和 hàn not hé).
• Polyphonic characters — characters read differently by word sense, independent of the Taiwan/Mainland split (e.g. 還款 huán vs. 還是 hái, 調整 tiáo vs. 調度 diào, 成分 fèn vs. 分鐘 fēn).
• Model misreadings — single-reading characters the strong TTS model still got wrong, surfaced through ASR comparison and spot listening, then forced to the correct reading (e.g. 房貸 dài, 崩潰 kuì, 肩頸 jǐng).
• Special tokens — non-normalized tokens rewritten for correct delivery (e.g. 7-11 → "seven eleven").

Every reading above was cross-checked against the Taiwanese Ministry of Education dictionary (《重編國語辭典修訂本》). The synthetic voice’s soft retroflex realization of zh/ch/sh is an intended feature of the Taiwanese accent, not a defect.

Intended Use

Training and evaluating Taiwan Mandarin speech models, especially on Mandarin–English code-switching and Taiwan-specific readings. Use the source field to keep human and synthetic data separate, such as evaluating on human clips only.

Limitations

• Single speaker and small dataset.
• Part of the dataset is synthetic and generated through voice cloning.
• Synthetic readings were verified through ASR comparison and spot listening, but not exhaustively.