BreezeBlue is a real-time voice AI team that grew out of building virtual-character voices.

On August 25, BreezeBlue open-sourced both the weights and the PyTorch inference code for Breeze TTS 2, its second-generation flagship TTS model. The model is just 3B parameters, and a single checkpoint covers 50 languages including Chinese and English. It can do reference-audio-based voice cloning, zero-shot voice design from a natural-language description, and it can switch emotion, pace and delivery on demand while keeping the speaker identity stable. The key headline number is that the median Time To First Audio (TTFA) sits at 133.6 ms — and with the warmed-up "fast" path it has gone under 40 ms — putting it in the same range as top paid services such as ElevenLabs Flash v2.5 and Fish Audio S2.1 Pro.

The "three crowns" all come from BreezeBlue's own benchmarks.

Alongside the model, BreezeBlue open-sourced three TTS evaluation suites — voice design, voice direction, and latency — and ran a head-to-head comparison against real competitors. On the voice design benchmark, Breeze TTS 2 ranks first with a Role Fit score of 78.02 and 708 distinct voices, beating the second-place MiMo-v2.5-TTS by 5.24 Role Fit points and 39% more voice diversity. On voice direction it leads with 4.25, 13% ahead of the next entry, while keeping speaker similarity (SPK_SIM) at 0.67. On latency, the model posts TTFB p50 119.4 ms, TTFA p50 133.6 ms, and TTFA p95 163.3 ms — all lower than ElevenLabs Flash v2.5, Fish Audio S2.1 Pro, Cartesia Sonic 3.5, xAI TTS, Speechify Simba 3.2, and Async Flash 1.5.

One important caveat: these three benchmarks were open-sourced by the BreezeBlue team itself, so all three first-place scores come from that team rather than from any cross-independent evaluation. Readers doing a horizontal comparison should keep that in mind.

How can a 3B model do "write a voice from text" and "play it in real time" at the same time?

Breeze TTS 2 packs four input conventions into one checkpoint:

  • Voice Clone: feed a reference audio clip with its exact transcript and the model preserves timbre, rhythm, and emotion.
  • Voice Design: no reference audio needed; describe the speaker in natural language — "middle-aged male, magnetic bass, staccato cadence" — and generate a brand-new voice.
  • Voice Direction: clone a voice, then steer each line's emotion, intensity, and pacing with natural-language instructions.
  • Vocal Events: inline tags such as (laugh) / (sigh) / (cough) in English or [笑] / [叹气] in Chinese become actual sound events inside the synthesized audio.

The inference stack is Apache 2.0 PyTorch. On NVIDIA H100, the eager path uses roughly 7.7 GiB of GPU memory; enabling --fast-all (full CUDA-Graph compilation) raises this to 14.4 GiB while shaving more off the first-audio latency. Chinese and English share the same checkpoint, and the same model generates Japanese, Spanish, French, German, Korean, Portuguese, Hindi and other languages up to 50 in total. The launch page demonstrates voice personas across animation, games, film and TV, literature and stage, mythology and folklore, and news broadcasting.

The weights are not Apache 2.0 — commercial deployment still requires a license.

The code is Apache 2.0, but the model weights, derivative models, and self-hosted outputs sit under BreezeBlue's own "research and non-commercial" license, and commercial deployment requires written authorization from RESONIA. The structure is the same one we have seen on MiniMax-Music3 and IndexTTS-2.5 recently — research users can download freely, enterprises still have to negotiate. Compared with fully closed commercial offerings from ElevenLabs and Cartesia, the open-weights TTS space is settling into a "free for research, paid for commercial" two-tier structure, and platform-side budgets will increasingly tilt toward the former.

The model card also writes voice-cloning compliance boundaries directly into the license: unauthorized voice cloning, impersonation, fraud, and other unlawful or harmful uses are prohibited. This matters for any Chinese team building AI customer service, digital humans, or short-video voice-overs — regulators in China have been visibly stricter on "realistic cloning of another person's voice" over the past year, and you have to clear the compliance question before shipping, not just glance at the license text.

Real-time voice AI is being driven back to "low latency".

The current wave of TTS releases (MiniMax-Music3, IndexTTS-2.5, Qwen-Audio-3.0-TTS, Breeze TTS 2) all share one direction — push TTFA / TTFB from "over one second" to "under 200 ms". The reason is simple: in real-time voice agents, game NPCs, and interactive stories, the natural conversational rhythm needs a first-audio latency of ≤ 300 ms, otherwise users feel an obvious "stutter". Every team is racing along three axes: voice diversity (thousands of personas from one model), emotional controllability (let the voice actually "act"), and streaming protocols (single WebSocket session across many turns, raw PCM flowing back while the model is still generating).

Breeze TTS 2 setting TTFA p50 at 133.6 ms effectively raises the engineering bar for real-time voice AI by another notch. Two numbers to watch next: first, whether third-party independent benchmarks (outside LMSYS and Artificial Analysis) can reproduce the "three crowns"; second, how the commercial license treats hardware, language, and voice-clone-count caps. Those two variables will decide its real place on the enterprise SDK shortlist.

Original announcement: https://breezeblue.ai/breeze-tts-2