BreezeBlue (@BreezeBlueX) released Breeze TTS 2, a text-to-speech model supporting 50 languages with three modes: Voice Clone (reference audio plus its exact transcript, preserving timbre and delivery), Voice Design (a natural-language description generates a voice from scratch, no reference audio needed), and Voice Direction (clone a voice from reference audio while steering tone, emotion, and pace via text prompt). Artificial Analysis — an independent benchmark aggregator, not BreezeBlue itself — confirmed the headline claim: Breeze TTS 2 is now the top-ranked open-weights model on its Provider Voices Speech Arena, beating the previous leader, Fish Audio S2 Pro, by 90 Elo points (1,215 vs 1,125).
The claim holds — on one of Artificial Analysis's two leaderboards
Worth stating plainly: the #1 ranking is real, independently measured, and a legitimate result. But Artificial Analysis runs two separate TTS leaderboards that test different things, and the two tell noticeably different stories about the same model. Provider Voices scores each model using its own native voice options — where Breeze TTS 2 tops the open-weights field. Controlled Voices standardizes the voice across models to isolate raw synthesis quality — and there, Breeze TTS 2 ranks #3 among open-weights models, tied with the very model it just beat, Fish Audio S2 Pro, both at 1,002 Elo, trailing Mistral's Voxtral TTS at 1,010. Same two models, opposite relative ordering, depending on which leaderboard is doing the comparing. Overall — open and closed models combined — Breeze TTS 2 lands 16th out of 39 on Controlled Voices, a meaningfully different picture than "the leading open-weights TTS model" on its own conveys.
Speed and price didn't improve alongside the Elo win
Artificial Analysis's own measurements on throughput and cost are the more concrete story here. Breeze TTS 2 processes 45 characters per second, against Fish Audio S2 Pro's 102 — less than half the generation speed of the model it just outranked on Provider Voices. On price, BreezeBlue's hosted endpoint charges $34 per million characters, against Fish Audio S2 Pro's $15 — more than double. None of that contradicts the Elo result; a model can win a quality comparison while losing on speed and cost. But a launch announcement built entirely around the Elo win, without mentioning either trade-off, gives an incomplete picture of what adopting the model actually costs in practice.
"Open weights" here means something narrower than usual
The model card's license terms are worth reading precisely rather than assuming from the "open weights" framing. The source code is Apache 2.0. The weights, checkpoints, adapters, and any derivative models or self-hosted outputs are governed separately by a BreezeBlue Research and Non-Commercial License — commercial use requires written authorization from RESONIA, INC. That's a materially different arrangement than MIT or Apache-licensed weights: downloadable and inspectable, yes; freely usable in a commercial product, no. Worth naming directly given how often "open weights" gets used as shorthand for "you can just use this" — the same gap this blog has flagged before between technically-open and practically-usable releases.
What to expect next
- Watch whether Controlled Voices or Provider Voices becomes the reference leaderboard cited going forward — they measure genuinely different things, and which one a given announcement leads with shapes the story readers get.
- Watch for a throughput or pricing update from BreezeBlue. A model that just took the top open-weights Elo spot on one leaderboard has a clear incentive to close the speed and cost gap with Fish Audio S2 Pro next.
- Watch for clarity on commercial licensing terms — the non-commercial default with a discretionary authorization path is a common pattern for "open" model releases this year, and it's worth tracking whether BreezeBlue publishes concrete terms for that authorization or leaves it case-by-case.
References: BreezeBlue/Breeze-TTS-2 on Hugging Face · Artificial Analysis — Text to Speech Open Weights Leaderboard · Elo, speed, and pricing figures independently measured and published by Artificial Analysis, not BreezeBlue's own numbers · related coverage: Frontier Arcade: trends & predictions