BreezeTTS2 - 100% Local Real-Time Voice

Introduction to Breeze TTS 2

Breeze TTS 2 is a new open-weight text-to-speech (TTS) model developed by the Chinese startup BreezeBlue. It is a 3-billion-parameter model that currently ranks first among open-weight models on the Artificial Analysis leaderboard.

Key Capabilities

• Voice Design: Generates a unique voice based solely on a descriptive text prompt, without requiring reference audio. It is highly capable of producing specific accents.
• Voice Direction: Clones a voice from a short audio reference while allowing users to steer the tone, emotion, and delivery style through natural language instructions.
• Vocal Events: Integrates expressive inline vocalizations such as laughing, coughing, sighing, and clearing the throat directly into the generated speech.
• Multilingual Support: Capable of generating expressive speech across 50 different languages.
• Low Latency and Streaming: Achieves a very fast time-to-first-audio, making it suitable for real-time conversational applications.

Limitations and Licensing

The primary drawback of Breeze TTS 2 is its research and non-commercial license, which restricts commercial use and prevents distillation. Additionally, because the model was trained without removing signal processing from the training data, some generated voices retain environmental artifacts, making them sound as though they were recorded through specific microphones or played through small speakers.

Local Performance and Quantization

The model can be run locally on hardware with a capable GPU. In a real-time demonstration combining a Gemma language model with Breeze TTS 2, the system generated spoken responses with millisecond-level latency. While the full 16-bit version delivers high quality, running 8-bit or 4-bit quantized versions can significantly reduce file size and processing time while maintaining comparable audio standards.