Qwen3-TTS-12Hz-1.7B-CustomVoice targets real-time, instruction-driven text-to-speech: it lets you produce multilingual, style-controllable audio with low end-to-end latency and reusable voice prompts. Its design focuses on shipping usable custom timbres and streaming-first generation for interactive applications rather than research-only benchmarks.
Key Capabilities
- End-to-end discrete-codebook TTS with the Qwen3-TTS-Tokenizer-12Hz so text → codes → audio without an explicit intermediate vocoder stage, enabling compact transmission and low-latency synthesis. This enables first-packet streaming (reported ~97ms) and fast incremental generation.
- CustomVoice interface: nine premium speaker presets (Vivian, Serena, Uncle_Fu, Dylan, Eric, Ryan, Aiden, Ono_Anna, Sohee) plus instruction-driven adjustments to timbre, emotion, and prosody. Useful for voice design, localized personas, and multi-language delivery.
- Voice-design and voice-clone workflows: a VoiceDesign model can synthesize a reference clip that the Base model then ingests for reusable voice cloning prompts (3-second cloning support). Integrations and examples exist for qwen-tts Python package, gradio demo, vLLM offline inference and Alibaba Cloud real-time APIs.
- Multilingual and cross-lingual performance: reported evaluations show competitive WER and speaker-similarity scores across 10 major languages and favorable cross-lingual transfer in many en/zh/ja/ko directions.
Who it's for and trade-offs
Great fit if you need low-latency, instruction-controllable TTS for demos, product prototypes, or service integration (real-time avatars, IVR, game/dialogue characters). It’s practical when you want explicit speaker presets plus the option to create reusable cloned personas. Look elsewhere if you require fully open-weight research reproducibility (some weights and tooling are tailored to the Qwen ecosystem), or if you need the absolute best long-form stability or the smallest possible runtime footprint for tiny-device deployment.
Where it fits
Positioned between SaaS TTS endpoints and heavyweight research stacks: closer to production-grade TTS with streaming and API support (e.g., Alibaba Cloud/DashScope, vLLM examples) while exposing local model usage through qwen-tts and Hugging Face.
