Model Card｜BreezeBlue Breeze TTS 2

TL;DR Breeze TTS 2: open weights (Apache 2.0 code, research/non-commercial model license), #1 open-weights model on Artificial Analysis Provider Voices (1,215 Elo, +90 over Fish Audio S2 Pro), #1 on both Voice Design (Role Fit 78.02) and Voice Direction (4.25) benchmarks; TTFA p50 133.6ms / p95 163.3ms, RTF 0.32 on H100; hosted API priced at $34 per 1M characters (over 2x Fish Audio S2 Pro); supports 50 languages, commercial use requires a separate license from RESONIA, INC.

#ai-agent #model-release #daily #breezeblue #model-family-breeze-tts

Table of Contents

Model Information
Highlights
Benchmark Results
Comparison with Predecessor/Competitors
Implications for Agent Development
Today's Insight
References

Series: AI Model Tracker (10 / 10)

Model Card｜Tencent Hy4 Preview

🌏 中文版

Model Information

Field	Value
Model ID	`BreezeBlue/Breeze-TTS-2` (hosted API model id `breeze-tts-2`)
Vendor	BreezeBlue (parent company RESONIA, INC.)
Parameters	Not fully disclosed; backbone is a Qwen3-architecture model at `llama-1B` scale (28 layers, hidden size 2048), plus a depth decoder (`llama-100M` scale), a T5Gemma2-based text encoder, and a Kyutai Mimi audio codec (24kHz, 32 quantizers)
Context Window	Not applicable (TTS model; streaming latency is the primary metric, not context length)
Input pricing (USD/1M tokens)	Not applicable; hosted API is priced at $34.00 / 1M characters
Output pricing (USD/1M tokens)	Not applicable (single billing dimension, same as above)
Open source	Yes (code under Apache License 2.0; model weights under the BreezeBlue Research and Non-Commercial License — commercial use requires separate authorization from RESONIA, INC.)
Release date	2026-08-17 (official announcement/press release); weights and PyTorch inference code open-sourced 2026-08-25
Official announcement	Introducing Breeze TTS 2
HuggingFace	BreezeBlue/Breeze-TTS-2
Family	Breeze TTS series (Breeze TTS 2 is the latest generation)

Highlights

Tops the Artificial Analysis Provider Voices blind Speech Arena among open-weights models: 1,215 Elo, 90 points ahead of the previous open-weights leader Fish Audio S2 Pro (1,125 Elo), ranking #6 overall out of roughly 100 models (open and closed combined)
#1 on BreezeBlue's own Voice Design benchmark (generating a voice from a text description with no reference audio): Role Fit score 78.02, 5.24 points ahead of runner-up MiMo-V2.5-TTS (72.78), and produces 39% more distinct voices (Voice Diversity 708) than the closest competitor
#1 on the Voice Direction benchmark (steering tone, emotion, and pace via natural language while preserving a reference voice's identity): score 4.25, 13% ahead of runner-up MiMo-v2.5-TTS (3.76), while keeping speaker similarity (SPK_SIM) at 0.67
Low-latency streaming: time-to-first-audio (TTFA) p50 133.6ms / p95 163.3ms, time-to-first-byte (TTFB) p50 119.4ms, and a warmed-up real-time factor (RTF) of 0.32 (~3.1x real time) on an H100 — faster than ElevenLabs Flash v2.5 and Fish Audio S2.1 Pro on the same latency benchmark
Natural speech across 50 languages, with inline vocal-event tags — parentheses in English (e.g. (sigh)) or square brackets in Chinese (e.g. [叹气]) — for laughs, coughs, sighs, and similar cues embedded directly in the script

Benchmark Results

Benchmark	Breeze TTS 2	Runner-up	Notes
Artificial Analysis Provider Voices (Elo, open weights)	1,215 (#6 overall out of ~100 models)	Fish Audio S2 Pro: 1,125	Each model uses its own native voices
Artificial Analysis Controlled Voices (Elo)	1,002 (tied; #16/39 among open weights)	Voxtral TTS (Mistral): 1,010	All models synthesize the same reference speaker; Breeze TTS 2 ties Fish Audio S2 Pro here
TTS Voice Design Benchmark (Role Fit)	78.02	MiMo-V2.5-TTS: 72.78	BreezeBlue's own open-sourced benchmark
TTS Voice Direction Benchmark	4.25	MiMo-v2.5-TTS: 3.76	BreezeBlue's own open-sourced benchmark; SPK_SIM 0.67
Throughput (characters/second)	45	Fish Audio S2 Pro: 102	Breeze TTS 2's throughput is less than half of the competitor's

⚠️ Provider Voices / Controlled Voices Elo scores come from Artificial Analysis, an independent third-party blind arena, and carry higher confidence. The Voice Design / Voice Direction / Latency benchmarks are BreezeBlue's own — but with open-sourced evaluation code, not a black-box self-report — though they haven't yet seen large-scale independent reproduction. Throughput figures are sourced from AlphaSignal's reporting on Artificial Analysis data.

Comparison with Predecessor/Competitors

Breeze TTS 2 is BreezeBlue's second generation, and this is the first time its weights have been made public (previously it was only available as a hosted commercial service). Compared to the incumbent open-weights leader Fish Audio S2 Pro, Breeze TTS 2 leads decisively (+90 Elo) in the "using its own designed voices" track, but only ties — and even trails Mistral's Voxtral TTS — on the Controlled Voices track, where every model synthesizes the same reference speaker. That split suggests Breeze TTS 2's edge isn't raw acoustic naturalness, but rather the product-oriented combination of voice design and voice direction — consistent with BreezeBlue's own positioning of the model for interactive voice experiences (game characters, digital companions, narrative content) rather than one-shot narration.

The trade-off is speed and price: throughput of 45 characters/second is less than half of Fish Audio S2 Pro's 102 characters/second, and the hosted API's $34 per 1M characters is more than double Fish Audio's $15 per 1M characters, and far above lightweight open-weights options like Kokoro 82M v1.0 at $0.65 per 1M characters. In short, Breeze TTS 2 trades throughput and cost for controllability and latency in voice design/direction — whether that trade is worth it depends on whether your application genuinely needs "a distinct voice per character" rather than simply high-throughput narration.

Implications for Agent Development

Breeze TTS 2's positioning is clear: it's not a general-purpose narration tool, but a model built for agents and applications that need real-time interactive voice. A p50 TTFA of 133.6ms combined with a single WebSocket connection that streams continuously across multiple conversation turns (connection.appendText() / flush() / endTurn()) is an API design purpose-built for multi-turn voice agent conversations.

If you're building game NPCs or digital companions: Voice Design lets you generate a character-specific voice from a text description (e.g. "a gravelly, world-weary storyteller") without a voice library, and a Voice Diversity score of 708 means you're less likely to collide on voices even with many characters
If you're building customer service or multi-turn voice agents: Voice Direction lets you adjust tone in real time (e.g. shifting from confident to panicked) within the same voice identity, without redesigning the voice, and low-latency streaming suits real-time voice interaction
Not a fit: high-throughput batch narration of long-form content (audiobooks, long-form voiceover) — at 45 characters/second and $34 per 1M characters, Breeze TTS 2 will be far more expensive than throughput-optimized models like Fish Audio S2 Pro or Kokoro; commercial deployment also requires separate authorization from RESONIA, INC. — only non-commercial testing is covered by the open weights license

Today's Insight

Most TTS benchmarks historically measure just one thing: how human the speech sounds (naturalness). But Breeze TTS 2's split performance across two tracks — dominant on Provider Voices, merely tied on Controlled Voices — points to something easy to miss: "good" for a voice model isn't a single dimension. When the product need is "design a plausible-sounding voice for each character" rather than "mimic one specific real speaker," traditional speaker-similarity/naturalness evaluations may not even capture the real differentiator — which is presumably why BreezeBlue built and open-sourced its own Voice Design and Voice Direction benchmarks instead.