Speech Service: Mandarin SSML with SSML marks — Tone accuracy issues in HD voice

Elvin Guo-Jin Chen 0 Reputation points
2026-11-19T05:07:47.57+00:00

Problem description

I am experiencing issues with Azure Speech Text to Speech for Mandarin Chinese (zh-CN) using an HD voice. When I wrap the word '偏好' in SSML phoneme markup with alphabet='sapi', the voice produces incorrect Chinese tones. However, the same SSML works correctly with neural voices like zh-CN-XiaoxiaoNeural.

Environment

Azure Speech Text to Speech; Language/locale: zh-CN; Voice: Dragon HD Flash; Region: eastus (assumed from context).

What I've already tried

I reproduced the issue by wrapping the word '偏好' in SSML phoneme markup with alphabet='sapi' and observed incorrect tones in the HD voice. When I send the same word as plain text without SSML, it is pronounced correctly. I also tested with neural voices like zh-CN-XiaoxiaoNeural, which handle SSML correctly. I confirmed that the phoneme markup is used with the correct sapi alphabet and tone numbers in the phoneme attribute.

Current status

I believe this is a voice-specific pronunciation issue in the Dragon HD Flash voice for Mandarin, and looking for options to fix it

Foundry Tools
Foundry Tools

Formerly known as Azure AI Services or Azure Cognitive Services is a unified collection of prebuilt AI capabilities within the Microsoft Foundry platform


Your answer

Answers can be marked as 'Accepted' by the question author and 'Recommended' by moderators, which helps users know the answer solved the author's problem.