Mandarin TTS in Nkomo: Why Xiaoxiao Moves Ahead of WaveNet

Nkomo previously sent Mandarin synthesis to Google `cmn-CN-Wavenet-A` before trying Microsoft's `zh-CN-XiaoxiaoNeural`. We have reversed that order: Xiaoxiao through Edge now leads, Azure provides the supported Microsoft fallback, and Google WaveNet remains third.

This is the clearest of our three comparisons. Google WaveNet is still intelligible and stable, but it is an older voice generation. Xiaoxiao has become a widely used Mandarin reading voice, and Microsoft's catalogue offers unusually broad Mandarin choice and expressive controls.

What consensus can and cannot tell us

Microsoft's voice catalogue documents numerous Simplified Mandarin voices, including voices with styles and roles. Google's catalogue shows both newer Chirp voices and the WaveNet voice Nkomo currently has configured. Therefore this comparison does not prove that every Microsoft model beats every Google model. It says Xiaoxiao is a stronger candidate than this particular older Google selection.

Chinese-language communities frequently praise Edge and Azure neural voices for fluid long-form reading. There are also reports of Edge output sounding assembled word by word on some inputs. Both observations are anecdotal, but together they suggest the right engineering posture: prefer Xiaoxiao while testing segmentation, punctuation, numerals and longer passages rather than relying on a perfect demo sentence.

The decision

Nkomo now uses Edge/Xiaoxiao, then Azure/Xiaoxiao, then Google WaveNet. Edge and Azure target the same voice ID. Azure is retained because the unauthenticated Edge transport can change independently of our application, whereas Azure supplies the official API and credentials.

Google remains useful as a diverse fallback. If both Microsoft paths fail, switching model families is better than returning silence. The API reports the provider that actually generated the audio, so operators can distinguish a successful fallback from normal service.

The next evaluation

Native review must cover tones in ambiguous contexts, erhua where appropriate, polyphonic characters, names, addresses, dates, currency, measure words, Latin abbreviations, English switching, questions and multi-paragraph reading. We should also compare Xiaoxiao with Google's current Chirp 3 HD voices rather than treating WaveNet as Google's permanent ceiling.

The present choice is therefore both confident and limited: Xiaoxiao is the better default against Nkomo's configured WaveNet-A voice. It is not evidence that Microsoft will beat Google's newest Mandarin models in every domain.