Qwen3-TTS 1.7B: open-source voice cloning with emotion and accent controls

What happened and what to do

The post describes Qwen3-TTS 1.7B CustomVoice, an open-source text-to-speech model with 1.7 billion parameters aimed at voice cloning. According to the author, the model offers controls for emotion, age, tone and accent, along with multilingual support. The author's claims about realism and latency have not been verified; the model card published on the model hosting platform is identified as the primary source, and it lists the license, languages and controls.

For a company, the practical point is to evaluate the model on its own infrastructure before deciding. That means building a test set with reference audio that has documented consent, measuring quality and latency on real hardware, and checking the license and languages against the intended use cases. In parallel, it is worth putting risk controls in place: authentication for voice-based calls, consent policies and synthetic audio detection, since voice cloning makes impersonation fraud easier. An evaluation pipeline, an internal API with usage logging and a metrics dashboard are enough to decide on data.

How the consultancy can help

Wendelmaques can diagnose the exposure and opportunities of synthetic voice in your operation, assess self-hosted TTS models for license, quality and latency, and design an implementation with consent records, usage controls and monitoring, operated on your infrastructure.

Next step

Send a short description of your synthetic voice use case, the expected volume and your current hosting environment. Based on that, Wendelmaques will prepare a scoped proposal with diagnosis, steps and contracting format, per project or as a monthly retainer.