The Voice You Can Describe, Not Clone
New contender in text-to-speech, breeze TTS 2, made waves when it supposedly beat ElevenLabs. Released by BreezeBlue on August 25th, this 3-gigabyte model runs entirely on your local machine, even a MacBook. Initial reports from artificial analysis benchmarks positioned it above industry giants, hinting at a breakthrough for local, open-weights TTS. But does its performance live up to the hype?
Breeze's real innovation lies in its "voice design" capability. Instead of cloning an existing voice from reference audio, you describe the voice you want using text. Imagine typing "A warm, thoughtful, wise old man with a calm, reflective delivery" and getting a unique voice generated from that prompt. This eliminates the need for initial audio samples, a common hurdle for many voice cloners, and fills a real gap in open-weights TTS.
Running breeze TTS 2 locally is surprisingly simple. Mac users can get it going with a quick terminal setup, using pip install MLX audio. Other platforms like Linux or Nvidia GPUs have equally straightforward installation paths. A crucial parameter for fine-tuning your generated voice is the CFG scale, which dictates how strongly the model adheres to your descriptive prompt. A lower CFG scale, like 1, results in a more generic AI voice, while the recommended default of 4 maintains fidelity to your description.
The Truth Behind the Leaderboard
Yes, breeze did technically 'beat' ElevenLabs on artificial analysis's Speech Arena leaderboard, but don't get excited. It edged out ElevenLabs Conversational by a mere five Elo points, 1215 to 1210. This slim victory, however, comes with significant asterisks.
That five-point lead looks less impressive when you consider the data behind it. Breeze's score rests on about 1200 votes, while ElevenLabs boasts over 4500. This makes breeze's ranking far less statistically settled on the board.
Furthermore, breeze isn't even the top dog on that specific leaderboard. Cartesia Sonic 3.6 holds a clear lead at 1282 Elo, making the initial 'beat ElevenLabs' headline highly selective and misleading.
The true test of a model's mettle comes from an apples-to-apples comparison. On a different artificial analysis leaderboard, where every model uses the same prompt, breeze's performance drops dramatically. It scores a mere 1002, plummeting to 16th place overall out of 39 models and third among open-weight models, behind Mistral's Voxtral.
Read the Fine Print: The Open-Weights Trap
Forget the leaderboard. The true catch with breeze TTS 2 hides in its fine print: the 'BreezeBlue Research and Non-Commercial License.' This isn't an open source license, as many users mistakenly believe. BreezeBlue states this explicitly: "It does not grant any commercial rights and is not an open source license." This is a critical distinction you shouldn't overlook.
The license imposes severe restrictions, extending far beyond merely preventing the resale of the model weights. It expressly forbids using the generated audio outputs for any commercial purpose. This includes production use, integration into a product or service, hosting it behind an API, or even using the outputs for internal operations beyond limited evaluation.
Consider what this bans: you cannot use a breeze voice in a monetized YouTube video, a sponsored podcast, or any commercial application. The license specifically names sponsorship and subscription revenue as prohibited. There's no creator exception, no small business exception, and no revenue threshold.
This makes the "free" open-weights model essentially a demo for BreezeBlue’s paid API, which costs $34 per million characters. For those seeking a truly open solution, this license is a trap, not an opportunity. For more technical details on the model itself, you can visit Breeze TTS 2: Real Time AI Voice Generation - BreezeBlue.
Enjoying this? Get one like it in your inbox each morning.
one email a day · unsubscribe in two clicks · no third-party tracking
The License Is the New Benchmark
Breeze TTS 2 presents a technically interesting package. Its voice design capability, which allows generating voices from mere descriptions like "a warm, thoughtful, wise old man," is a significant step forward, bypassing the need for reference audio. However, this innovation is ultimately crippled by the "BreezeBlue Research and Non-Commercial license." This isn't a license for community use; it's a lead-generation tool, explicitly stating it is "not an open source license" and prohibiting commercial output.
Nowadays, the distinction between available "open weights" and genuinely "open source" software is paramount. Breeze TTS 2 perfectly illustrates this, making its weights accessible but its output commercially unusable. For any practical application, a model's license is just as important as its raw performance metrics, like its artificial analysis Speech Arena score. One tells you what the technology can achieve; the other tells you what you are legally permitted to achieve with it.
Developers and creators must prioritize license terms over leaderboard scores. If your project requires commercial rights or production use, breeze is a non-starter. Instead, choose truly open-source alternatives with permissive licenses, such as:
- VoxCPM2
- Kokoro (Apache 2.0)
- Chatterbox (MIT)
These models empower you to build without legal encumbrance. The benchmark tells you what a model can do; the license tells you what you can do.
Frequently Asked Questions
What is Breeze TTS 2?
Breeze TTS 2 is a 3B parameter open-weights text-to-speech model from BreezeBlue. Its standout feature is 'voice design,' allowing users to generate a voice from a text description rather than cloning an audio sample.
Does Breeze TTS 2 actually beat ElevenLabs?
Technically, yes, on one specific user-voted leaderboard by a very small margin and with fewer votes. On a direct, apples-to-apples benchmark, it ranks 16th overall and falls behind other models like Mistral's Voxtral.
Why is the Breeze TTS 2 license so restrictive?
Its 'Research and Non-Commercial' license explicitly forbids using the model's outputs (the generated audio) for any commercial or internal business purpose, including monetized content or products. This makes it a demo for their paid API, not a tool for creators.
What are some truly open-source alternatives to Breeze TTS?
For developers seeking permissive licenses, models like Kokoro (Apache 2.0), Chatterbox (MIT), and VoxCPM2 offer powerful TTS capabilities without the restrictive commercial limitations of Breeze.