Gepard 1.0, a Streaming TTS Model for Real-Time Conversation, Released as Open Source

Gepard 1.0, designed for real-time conversational AI, has been released as open source under the Apache 2.0 license. The model adopts a streaming-first approach that generates audio frame-by-frame as text comes in, without waiting for a sentence to be completed.

Key Technical Specifications and Performance:

• Model Architecture: A roughly 555M-parameter model based on a Qwen3.5 0.8B backbone (14 layers) and Nemo NanoCodec (FSQ, 22.05kHz).
• Inference Speed: Achieved a 20x real-time factor (RTF) and ~50ms time-to-first-audio (TTFA) on an RTX 5090.
• Scalability: Can handle up to 256 parallel sequences on a single RTX Pro 6000 Blackwell (96GB VRAM).
• Features: Supports zero-shot voice cloning using just a few seconds of reference audio, and supports English, Spanish, Portuguese, and Dutch.

Performance Evaluation: On the Seed-TTS-eval benchmark, the model achieved best-in-class perceptual quality with a NISQA-MOS of 4.25, outperforming existing models such as VoxCPM2 and Fish-S2, and delivered the cleanest audio quality in terms of noise and discontinuity. However, due to streaming optimization, there are some trade-offs in speaker similarity (SIM 0.585) and word error rate (WER 0.036).