Models and technology
Some voice models focus on language coverage, others on emotional performance or local deployment. Explore the official research, concrete strengths, version differences, and ways to get started.
Describe a character or speaking style in an instruction, add emotion and laughter tags to a script, and guide cloned voices into Chinese dialects.
Xiaomi's open-source TTS spans 600+ languages, clones from short recordings, and adds voice attributes, duration control, and pronunciation correction.
A stability-focused 1.x release with improved English and mixed Chinese-character/pinyin input, useful when narration needs predictable pronunciation.
Keep a character's timbre while borrowing emotion from another clip, setting eight emotion intensities, or describing the desired feeling in text.
Generate continuous speech representations without discrete audio tokens: 1.5 offers 44.1 kHz Chinese/English cloning; version 2 adds 30 languages and 48 kHz.
DS Speech currently provides Qwen-Audio-3.0-TTS online. Other models in this library are documented for learning and self-deployment.