EmotiVoice 2026：带真实情感的开源中文语音合成
中文语音合成太机械，讲故事尤其明显。网易有道的EmotiVoice能生成带可控情感的中文语音。
EmotiVoice: Mandarin TTS With Feelings
EmotiVoice, open-sourced by NetEase Youdao (8,507 GitHub stars as of August 2026), is a text-to-speech system focused on one thing most Chinese TTS models neglect: emotion. It lets you specify an emotion label - happy, sad, angry, surprised, fearful, disgusted, neutral - and control emotional intensity, producing Mandarin speech that actually sounds expressive instead of flat.
Why Emotion Matters for TTS
Audiobooks, game characters, dubbing, and voice assistants all fail when delivery is monotone. Most open-source Chinese TTS (like the older chinese TTS models in Coqui) reads text correctly but with a single flat tone. EmotiVoice's dual approach - a base speech model plus an emotion control module - makes expressive output a parameter you set, not a lucky accident.
How to Use It
git clone https://github.com/netease-youdao/EmotiVoice
git submodule update --init --recursive
# run the web demo and type text + choose emotion
It ships with a Gradio web interface and a command-line synthesis script. Model weights are large (several GB) but inference runs on a single consumer GPU, and CPU inference works for short clips.
The Practical Comparison
| Need | Tool |
|---|---|
| Expressive Mandarin with emotion control | EmotiVoice |
| English + Chinese general TTS | Edge-TTS / Azure / ElevenLabs |
| Voice cloning | OpenVoice / RVC |
| Fastest Chinese TTS (less emotion) | Native cloud APIs |
Limitations to Know
- Long text synthesis can lose emotional consistency - keep paragraphs short.
- Some emotions (surprised, fearful) are less natural than happy/sad in early versions.
- It is primarily Mandarin; English support exists but is not the focus.
FAQ
Is EmotiVoice free? Yes, open source (Apache-2.0 style license per the repo); you pay for compute.
Does it support voice cloning? No - that is a different tool (see OpenVoice).
Can I use it commercially? Yes, for your own generated audio; verify the license terms in the repo.
How much VRAM do I need? Around 8-12GB for the full model; quantized versions run on less.
