Gepard 1.0 open-sources a streaming TTS model for real-time dialogue

Researchers have open-sourced Gepard 1.0, a streaming-first text-to-speech model designed for real-time conversation that generates audio frame by frame as text arrives. The model is built on a Qwen3.5 0.8B backbone with Nemo NanoCodec and supports zero-shot voice cloning across English, Spanish, Portuguese, and Dutch.

• Approximately 555 million parameters using a Qwen3.5 0.8B backbone and Nemo NanoCodec (FSQ, 22.05kHz)
• Achieves ~20x real-time factor and ~50ms time-to-first-audio on one RTX 5090 via vLLM
• Supports up to 256 parallel sequences on a single RTX Pro 6000 Blackwell with 96GB VRAM
• Leads Seed-TTS-eval benchmarks with a NISQA-MOS of 4.25, outperforming VoxCPM2, Fish-S2, OmniVoice, Qwen3-TTS, Echo-TTS, and Chatterbox Turbo on perceived quality
• Licensed under Apache 2.0 with code available for inference, vLLM serving, and training

The streaming-first design prioritizes natural real-time voice interaction over exact speaker similarity, making it suitable for applications where low latency is more critical than perfect voice-matching.