Sub-100ms streaming text-to-speech that sounds human — five languages, ~22 hours of audio for $5, one API call away.
No account needed
The two ways TTS speed gets measured — vendor model-only latency, and independent end-to-end from the Coval benchmark. Lower is faster.
Model latency is each vendor's own model-only figure (excludes network) — ours is ~35 ms. End-to-end is independent Coval P50 TTFA, benchmarks.coval.ai/tts. Our ~90 ms (marked *) is self-measured and pending an independent benchmark. Qwen's model figure is the hosted Qwen-Audio-3.0-TTS-Flash (~300 ms, the build Coval tested); its open-weights variant claims ~97 ms. Grok publishes no model-only figure. Cartesia additionally self-reports #1 naturalness. Latency is measured differently by every vendor — treat these as directional, not exact.
(NEW TTS) Gepard 1.0: very fast and sounds great + Apache 2.0 - demo available on Hugging Face ⤵️
We open-sourced Gepard 1.0 - a 555M streaming TTS that starts talking as text arrives. ~50ms to TTFA, ~20x RTF on 1 RTX 5090. Voice cloning from seconds of audio. vLLM native. Apache 2.0. Demo + weights: huggingface.co/nineninesix/ge…
Ever wanted to clone voices with just a few seconds of audio? Meet Gepard 1.0, a text-to-speech model that combines autoregressive generation with voice cloning. It's built on Qwen 3.5 and uses Safetensors for efficient inference. 819 downloads and counting!
Call the models through our API, or download the open weights from Hugging Face.
