Real‑time conversational systems need speech that starts immediately, flows naturally, and scales to many concurrent callers. Gepard rethinks TTS for dialogue by making the model produce whole audio frames as soon as text begins arriving, trading a bit of speaker similarity and word accuracy for immediate, natural‑sounding streaming.
What Sets It Apart
- Streaming‑first autoregressive design: generates entire audio frames (FSQ codec frames) in one step so output can begin before a full sentence is available — typical time‑to‑first‑chunk ≈ 50 ms on target hardware. This reduces perceived latency compared with two‑stage TTS pipelines that wait for full utterances.
- Single model that learned text and speech jointly: prosody, timing and pacing emerge from the same network rather than being stitched from separate duration/pitch/stats models, producing more conversational rhythm.
- Practical performance points: backbone is a Qwen3.5 full‑attention decoder (~14 layers, hidden 1024, 8 heads) with total ≈ 555.7M parameters; audio codec is NVIDIA NeMo NanoCodec (FSQ, 22.05 kHz, 21.5 fps, 1.89 kbps); sample rate 22,050 Hz.
- Real‑world throughput: ~25× real time on an RTX 5090; a 96 GB GPU configuration can serve many simultaneous conversations (model reports up to ~256 parallel sessions in that environment). vLLM‑native optimizations are provided for low latency and high concurrency.
- Quality vs tradeoffs: ranks top on perceived naturalness (NISQA‑MOS) and noise/color/discontinuity metrics in published evaluations but shows lower speaker similarity and slightly higher WER compared to some non‑streaming baselines — an explicit design choice favoring live conversational feel.
- Practical features: baked‑in CFG refinement option (two‑pass mode available as a quality dial), short‑clip zero‑shot voice cloning, multilingual support (EN, ES‑MX, PT‑BR, NL) and a production API/service option for managed deployment.
Who It's For & Trade‑offs
Great fit if you need an immediately responsive conversational voice agent or voice bot where natural prosody and low perceived latency matter more than perfect speaker mimicry or absolute ASR accuracy. Ideal for voice assistants, dialog systems, live streaming voice agents, and call‑center automation at scale.
Look elsewhere if your top priorities are exact speaker reproduction, the lowest possible WER, or you require languages/accents outside the supported set — some voices and non‑English languages are noted to vary in quality. Also note the model uses NVIDIA’s NeMo NanoCodec, which carries its own license terms.
Use responsibly: do not clone voices without consent and follow applicable laws and license terms.
