Breeze TTS 2: High Performance, Restrictive Licensing

Voice Design and Inference

Breeze TTS 2 enables voice generation through plain-English text descriptions rather than requiring reference audio. Users can control the model's adherence to these descriptions using the CFG scale parameter, where a value of 4 is recommended for stability and lower values like 1 result in more generic AI-sounding output. The model also supports inline vocal events, allowing users to embed cues like (sigh) or (laughing) directly within the text string to influence delivery.

Technically, the model is a 3-billion parameter system built on a Qwen 3 backbone, a T5 Gemma 2 text encoder, and an MI audio codec. It outputs 24 kHz mono audio and can be run locally on macOS using MLX audio.

The Licensing Constraint

While the code is released under an Apache 2.0 license, the model weights are governed by a restrictive "BreezeBlue research and non-commercial license." This license explicitly prohibits:

• Production use in any product or service.
• Hosting the model behind an API for commercial purposes.
• Using generated audio outputs for any internal operations beyond limited evaluation.
• Any commercial use, including monetized YouTube videos, podcasts, or content supported by sponsorships.

There are no exceptions for small businesses or revenue thresholds. The license clarifies that even paying for the company's official API does not grant commercial rights to the self-hosted weights or their outputs, positioning the open-weights release primarily as a demonstration tool.

Benchmark Context

Breeze TTS 2's high ranking on the Artificial Analysis leaderboard is misleading due to the methodology. While it technically holds a high ELO score, it has significantly fewer human votes compared to established systems like ElevenLabs. Furthermore, in "apples-to-apples" testing where models are forced to use identical prompts, Breeze TTS 2 ranks 16th out of 39 models, trailing behind others like Mistral's Foxtrol.