Speech-to-speech (Symphony)
Need the lowest latency and the most natural turn-taking? Use Symphony, one model end to end.
speech inSymphonyspeech out
- ~1 s to first agent audio
- Realtime tool calling
- Natural interruptions and turn-taking
Voice agents that listen, think, and talk back in one API.
Native speech-to-speech agents, plus the full STT → LLM → TTS pipeline when you need control at every step.
Press record and read something out loud. The transcript writes itself.
Uses your microphone
Time to first token
- Sprag186 ms
- Deepgram364 ms
- OpenAI565 ms
- ElevenLabs584 ms
- AssemblyAI2,238 ms
Enterprise-grade infrastructure, built for speed.
Sprag serves its own models on open-source infrastructure rather than reselling somebody else's endpoint. That is where the latency goes, and it is why our audio engineers can tune a voice for you when a stock one will not do.
Time to first agent audio by provider, lowest first: Sprag 1033 milliseconds (lowest), OpenAI 1458 milliseconds, Cartesia 1896 milliseconds, AssemblyAI 3107 milliseconds, Deepgram 3299 milliseconds.
1.18s p50 end-to-end turn latency on agentic_basic, n=30, against OpenAI at 1.76s and Cartesia at 2.24s. Speech in, speech out through a single model.
See the methodologyChoose your architecture
Both run on the same API. Switch per use case, not per vendor.
Cascading pipeline (Rhythm → LLM → Chorus)
Need guardrails at the text layer, a specific LLM, or per-step logs? Use the cascade, with LLMs we serve and tune for speed, or your own.
RhythmLLMSprag-served · or bring your ownChorus
- Sprag-served LLMs, tuned to cut turn latency
- Swap any stage, including your own LLM
- Full transcripts and per-step observability
Try every building block
Each stage of the pipeline is its own API. Use them together, on their own, or alongside Symphony.
Press record and read something out loud. The transcript writes itself.
Uses your microphone
Every voice speaks every language.
53 preset voices, each with its own origin and accent. Any of them can be asked for any of the 10 output languages below, and fidelity is strongest in the one they came from.
Output languages: English, Chinese, French, German, Italian, Japanese, Korean, Portuguese, Russian, Spanish.
Eight of 53. Every voice you clone or design joins them.
Browse the full library
Founded on Open-Source
The inference engine built for every modality.
−91.4%
reduction in job completion time arXiv↗
Omni-modal
Every modality. One surface.
Text, image, audio, and video inference in a single API layer.
Any architecture
AR, DiT, and parallel generation.
Autoregressive, diffusion transformers, and non-autoregressive models in one engine.
OpenAI + ComfyUI
Works with the client you have.
Fully OpenAI-compatible API server. Native ComfyUI support.
KV cache efficiency
vLLM memory management, extended.
State-of-the-art KV cache from vLLM core, applied to multimodal workloads.