BreezeBlue Debuts Breeze TTS 2 to Lead Open Voice AI

BreezeBlue has released Breeze TTS 2, a new open-weights text-to-speech model that has claimed the top spot for open models on the Artificial Analysis Provider Voices leaderboard. Scoring 1,215 Elo, the model surpasses the previous open-weights leader, Fish Audio S2 Pro, by 90 points, which sits at 1,125 Elo. This score places Breeze TTS 2 at rank six overall out of more than 100 evaluated models, just behind closed systems like Sonic 3.6 at 1,283, Qwen-Audio-3.0-TTS-Plus at 1,238, Simba 3.2 at 1,238, Luna TTS at 1,223, and v3 Conversational at 1,219. On the Controlled Voices track, Breeze TTS 2 tied Fish Audio S2 Pro at 1,002 Elo, trailing Mistral's Voxtral TTS at 1,010 Elo to rank 16th out of 39.

The model supports 50 languages and excels at interactive voice tasks, combining voice design, voice direction, and low-latency streaming. On BreezeBlue's internal benchmarks, Breeze TTS 2 ranks first in Voice Design with a Role Fit score of 78.02, beating MiMo-V2.5-TTS at 72.78. It also leads in Voice Direction with a score of 4.25 compared to 3.76 for the runner-up, while maintaining a speaker similarity of 0.67. For real-time applications, it offers a p50 time to first byte of 119.4ms, a p50 time to first audio of 133.6ms, and a p95 time to first audio of 163.3ms, outperforming ElevenLabs Flash v2.5 and Fish Audio S2.1 Pro.

These capabilities come with some trade-offs in speed and cost. Breeze TTS 2 processes 45 characters per second, which is slower than the 102 characters per second achieved by Fish Audio S2 Pro. Additionally, BreezeBlue's hosted API is priced at $34 per 1 million characters. This is more expensive than Fish Audio S2 Pro at $15 per 1 million characters and significantly higher than lightweight open-weights options like Kokoro 82M v1.0, which costs $0.65 per 1 million characters. However, because the weights are openly hosted on Hugging Face, developers can bypass these API costs by running self-hosted inference.

For practitioners, this release bridges the gap between open-weights flexibility and commercial-grade voice control. Developers can now build highly responsive voice agents, such as video game NPCs requiring distinct personalities or customer service systems that adjust their tone mid-sentence, without relying on closed-source APIs.