Breeze TTS 2 Guide 2026: The Top Open-Weight Speech Synthesis Model

On August 25, 2026, BreezeBlue open-sourced the Breeze TTS 2 model weights and inference code. It ranks #1 among open-weight models on the Artificial Analysis TTS leaderboard — even outperforming frontier proprietary systems.

The short version: Breeze TTS 2 is an open-weight text-to-speech model built for real-time interaction. Its natural-language instruction-following supports reference-free "Voice Design" and reference-guided "Voice Direction", with under 40ms time-to-first-audio on an NVIDIA H100.

This article covers what Breeze TTS 2 is, its four voice features, performance, setup, and licensing.

What is Breeze TTS 2

Breeze TTS 2 is an open-weight text-to-speech (TTS) model developed by BreezeBlue (RESONIA, INC.).

Item	Value
Model	BreezeBlue/Breeze-TTS-2 (HuggingFace, ModelScope)
Developer	BreezeBlue (RESONIA, INC.)
Est. params	~1.74B (BF16)
Model size	~3.5GB (BF16/F32)
Languages	English & Chinese (bilingual, one model)
Released	August 25, 2026
License	Code: Apache 2.0 · Weights: research & non-commercial
Features	Voice Clone / Design / Direction / ultra-low latency

Key highlights

Top open-weight TTS: #1 among open-weight models on the Artificial Analysis leaderboard, beating frontier commercial systems
Four voice features: clone, design, direction, and vocal events
Ultra-low latency: under 40ms time-to-first-audio (TTFA) on H100
Real-time generation: RTF 0.32 (~3.1× real time)
GPU-efficient: ~7.7 GiB for eager inference; runs on a 12 GB GPU

Four voice features

The standout capability: you control the voice in natural language.

① Voice Clone

Pass reference audio plus its exact transcript to preserve the speaker's timbre, rhythm, emotion, and style.

python infer.py ../breeze-tts-2 \
  --ref-audio reference_en.wav \
  --ref-text "This is the exact transcript of the English reference audio." \
  --text "(sigh) It is good to hear your voice again after all this time." \
  --output outputs/voice_clone_en.wav

Reference audio must be clean speech with minimal background noise.

② Voice Design

Create a distinctive voice from a natural-language description, with no reference audio — a capability traditional TTS lacks.

python infer.py ../breeze-tts-2 \
  --text "(sigh) Welcome aboard. Your journey begins now." \
  --instruction "A warm, thoughtful young woman with a clear voice and a calm, reflective delivery." \
  --cfg-scale 4 \
  --output outputs/voice_design_en.wav

Use --cfg-scale 4 to strengthen instruction-following.

③ Voice Direction

Keep the cloned voice's identity while steering tone, emotion, pace, and delivery — like directing a voice actor.

python infer.py ../breeze-tts-2 \
  --ref-audio reference.wav \
  --ref-text "This is the exact transcript of the reference audio." \
  --text "(clears throat) We need to discuss what happened last night." \
  --instruction "Speak slowly with a restrained, serious tone." \
  --cfg-scale 4 \
  --output outputs/voice_direction.wav

④ Vocal Events

Embed expressive events directly in the text: English parentheses (laugh), (cough), (clears throat), (sigh); Chinese square brackets [笑], [咳嗽], [清嗓子], [叹气].

Performance: ultra-low latency, real-time generation

Metric	Value	Notes
TTFA	<40ms	Time to first audio (H100, fast path)
RTF	0.32	~3.1× real time generation (H100, fast path)
GPU memory	7.7 GiB	Eager inference (runs on 12 GB GPU)
GPU memory (fast)	14.4 GiB	With --fast-all (24 GB GPU recommended)

Requirements

OS: Linux, Python 3.10+
GPU: CUDA-capable NVIDIA GPU (12 GB+ recommended; 24 GB for fast path)
Framework: PyTorch (transformers)

Setup

1. Install

git clone https://github.com/breezeblue-ai/breeze-tts.git
cd breeze-tts
python -m pip install -r requirements.txt

2. Docker (recommended)

bash docker/build.sh

Default targets H100/Hopper (sm90). For A100:

FLASH_ATTN_CUDA_ARCHS=80 bash docker/build.sh

3. Streaming API (OpenAI-compatible)

python -m breeze_infer.api ../breeze-tts-2 --host 0.0.0.0 --port 7860

curl -X POST http://127.0.0.1:7860/v1/audio/speech \
  -F "cfg_scale=4" \
  -F "ref_audio=@reference.wav" \
  -F "ref_text=This is the exact transcript of the reference audio." \
  -F "text=(clears throat) We need to discuss what happened last night." \
  -F "instruction=Speak slowly with a restrained, serious tone." \
  -F "seed=42" \
  --output voice_direction.pcm

The response is streaming mono 24 kHz, 16-bit little-endian PCM.

4. Fast inference (--fast-all)

Eager streaming is the default (no graph warmup). Add --fast-all to optimize every stage (text encoder, backbone prefill/decode, depth decoder, codec) when the extra cold-start is acceptable.

License & caveats

Important constraints:

Code: Apache License 2.0 (commercial OK)
Model weights, checkpoints, adapters, derivatives, and self-hosted outputs: research & non-commercial only
Commercial use requires written authorization from RESONIA, INC. (contact@breeze.blue)

Unauthorized voice cloning, impersonation, and fraud are prohibited.

Summary

Breeze TTS 2 delivers TTS that doesn't just read — it performs.

✅ #1 open-weight model on the Artificial Analysis TTS leaderboard
✅ Voice Design creates new voices from text descriptions
✅ Voice Direction steers cloned voices with acting instructions
✅ Under 40ms TTFA on H100
✅ Light enough for a 12 GB GPU (~1.74B params)
✅ Bilingual English & Chinese

For real-time conversational voices, character voice design, and narrated performances, Breeze TTS 2 is a top contender. But the weights are non-commercial — contact RESONIA before any commercial use.