Mandarin neural voices can sound excellent on curated prose and still fail on the details users notice immediately. Polyphonic characters, personal names, addresses, measure words, dates and Latin abbreviations reveal whether a voice is ready for production.
Microsoft lists Xiaoxiao and many other Simplified Chinese choices in its speech catalogue. Google’s catalogue includes WaveNet and newer model families. Comparing Xiaoxiao only with an older WaveNet voice answers a narrow routing question; it does not settle the wider provider contest.
Design a harder listening set
Include characters whose pronunciation depends on context, common surnames, place names, currency, percentages, phone numbers and sentences mixing Chinese with English product terms. Add questions, dialogue and multi-paragraph narration. Punctuation and segmentation deserve their own cases because they strongly affect rhythm.
Listen for more than tones
Native reviewers should score lexical correctness, tone accuracy, phrasing, naturalness and long-form fatigue separately. A small pause error may be cosmetic; the wrong reading of a name or number can change meaning. Record both severity and frequency.
Operational proof
Nkomo currently prefers Xiaoxiao through Edge and keeps Azure as the same-voice fallback. Google remains a separate model-family fallback. Provider and voice headers make the path visible. A release test should confirm valid audio, the requested voice ID and readable pacing after forced fallback.
Mandarin TTS quality is not one score. It is correctness across the difficult tokens plus listening comfort over the actual duration.
Comments
No comments yet.