Qwen Unveils Open Model for Custom Voice Synthesis

Alibaba's Qwen team has expanded its open-source offerings with Qwen3-TTS, a new model dedicated to high-quality speech synthesis. Released under a permissive Apache 2.0 license, this 1.7-billion-parameter system marks a significant entry into the growing field of open text-to-speech (TTS) models.

The model's standout feature is its "Voice Design" capability. Unlike traditional TTS systems that rely on a fixed set of pre-recorded voices, Qwen3-TTS can generate speech in a novel voice by analyzing a short audio prompt. This allows developers to create unique voices or clone existing ones for custom applications, a feature previously common in proprietary, API-driven systems.

Multilingual and Prompt-Driven

Qwen3-TTS is designed to be multilingual and is controlled through a combination of text and audio inputs. A user provides the text to be spoken along with a reference audio clip, and the model generates speech that matches the voice characteristics of the reference. The "12Hz" in the model's name likely refers to the sampling rate of its internal audio representation, a technique used in modern neural audio codecs to efficiently model speech.

The release of a powerful, commercially-permissive voice design model like Qwen3-TTS is a notable development for the open-source AI community. It provides a foundational tool for a wide range of applications, including personalized digital assistants, dynamic video game character dialogue, and accessibility tools, without the restrictions of closed platforms.