BreezeBlue's Breeze TTS 2 Tops Open Weights Voice AI With 1,215 Elo

The open weights text-to-speech race has a new frontrunner. Breeze TTS 2 from BreezeBlue has taken the top open weights slot on the Artificial Analysis Provider Voices leaderboard, opening a large gap over the previous leader and pushing into territory mostly occupied by closed commercial systems.

Fish Audio S2 Pro previously held the open weights lead with an Elo of 1,125. Breeze TTS 2 now sits at 1,215 Elo, a 90 point jump, and lands at rank six overall out of roughly 100 models scored. Weights are published on Hugging Face, and a hosted API is available for teams that would rather skip running inference themselves.

Reading the leaderboard

Artificial Analysis runs a blind arena where users listen to pairs of speech samples generated from the same text and pick which sounds more natural. Higher Elo scores indicate a model produces speech preferred more often by listeners. There are two tracks: Provider Voices, where each model uses its own native voices, and Controlled Voices, where all models synthesize the same reference speaker.

Results diverge sharply between the two tracks:

• Provider Voices: Breeze TTS 2 leads open weights at 1,215 Elo, ahead of Fish Audio S2 Pro (1,125) and Step Audio EditX (1,102).
• Controlled Voices: It ties Fish Audio S2 Pro at 1,002 Elo and trails Mistral's Voxtral TTS (1,010), sitting at rank 16 out of 39.
• Overall top five: Sonic 3.6 (1,283), Qwen-Audio-3.0-TTS-Plus (1,238), Simba 3.2 (1,238), Luna TTS (1,223), and v3 Conversational (1,219). Breeze TTS 2 lands one rung below that group.

The split between tracks suggests the strength here comes less from raw acoustic naturalness than from what BreezeBlue does with its own designed voices, which is what the model was built around.

Design, direct, stream

BreezeBlue frames Breeze TTS 2 as a model for interactive voice rather than one-shot narration, targeting three capabilities usually split across separate systems: voice design from a text prompt, voice direction using natural language and inline tags, and low-latency streaming. On their own benchmarks Breeze TTS 2 ranks first on both Voice Design (Role Fit 78.02, well ahead of MiMo-V2.5-TTS at 72.78) and Voice Direction (4.25 versus 3.76 for the next model), while keeping speaker similarity at 0.67.

Latency is the more concrete story for anyone building agents. Reported time to first byte at p50 is 119.4ms, with time to first audio at p50 of 133.6ms and p95 of 163.3ms, ahead of ElevenLabs Flash v2.5 and Fish Audio S2.1 Pro on that benchmark. The realtime SDK keeps a single WebSocket open across turns and streams raw PCM back while generation is still in flight:

const connection = await client.textToSpeech.realtime.connect("voc_...", {
  modelId: "breeze-tts-2",
});
connection.startTurn("turn_1");
connection.appendText("Hello from Breeze TTS 2.");
connection.flush();
connection.endTurn();

The throughput and price catch

Breeze TTS 2 processes 45 characters per second, less than half the 102 characters per second Fish Audio S2 Pro manages. On BreezeBlue's hosted endpoint it runs $34 per 1M characters, more than double Fish Audio's $15 per 1M characters and well above open weights options like Kokoro 82M v1.0 at $0.65 per 1M characters. Weights are available for self-hosting either model, so the pricing gap mostly matters if you plan to consume the API directly rather than run your own inference.

Why the release lands

Open weights TTS has been closing the naturalness gap with commercial systems for a while, but has lagged badly on the pieces that make voice agents actually work: creating a specific character voice on demand, steering that voice inside a single line, and doing it fast enough to feel like a conversation. Rolling those capabilities into one model, with weights you can inspect and run, closes a real gap.

The practical use cases follow: NPCs in games where you need thousands of distinct voices without a studio budget, customer service agents that need to shift tone within a single response, multilingual content pipelines spanning 50 languages, and any voice product where sending audio to a closed API is a non-starter for latency, cost, or compliance reasons. For teams that had been holding out for an open weights option that could stand next to the commercial leaders, the leaderboard math finally works in their favor.