Text-to-speech models

Two questions narrow the field: do you need a custom voice or will a built-in voice work, and do you need real-time streaming?

​

Built-in or custom voice?

​

Built-in voices

Pick a voice from the library and start synthesizing immediately.

• CosyVoice — rich voice library, high quality, no setup beyond picking a voice
• Qwen3-TTS — low-latency streaming; add -instruct for natural-language control over speed, emotion, and style

​

Custom voice

Need a voice that doesn't exist in the library?

• Voice Cloning — reproduce a specific person's voice from audio samples. Use when you have a target voice to match.
• Voice Design — create a new voice from a text description (e.g., "a warm, low-pitched female voice"). Use when you want a brand voice without audio samples.

​

Controlling how the voice sounds

Three approaches, ranked by flexibility:

1. Instruction control (qwen3-tts-instruct-flash, qwen3-tts-instruct-flash-realtime) — Describe the desired delivery in natural language. Control speed, emotion, and style per request. Most flexible.
2. Voice design (qwen3-tts-vd-*) — Generate a custom voice from a text description. Good for creating a brand voice without audio samples.
3. Voice cloning (qwen3-tts-vc-*) — Reproduce an existing voice from audio samples. Best when you need to match a specific person's voice.

​

Recommended models

​

All models

​

Learn more