Skip to main content

​
What is Capella?

Capella is Sabi’s fine-tuned text-to-speech voice model, built for natural, expressive English speech. It supports inline audio tags — laughs, sighs, breaths, pauses, and per-sentence emotional styles — so generated speech sounds like a performance, not a readout. Capella is served through two front doors:

HTTP API

Drop-in OpenAI-compatible POST /v1/audio/speech. Point the OpenAI SDK at our base URL and it just works.

Streaming WebSocket

Low-latency, sentence-by-sentence audio frames for conversational and real-time apps.

Capella reference

Configuration parameters, environment variables, and usage examples.

Audio tags

The full vocabulary of event tags and style wraps for expressive delivery.

​
Capabilities

  • Expressive delivery — inline event tags (<|laugh|>, <|sigh|>, <|pause|> …) and sentence-level style wraps (excited, whispers, sarcastic, accents, …).
  • OpenAI compatibility — use the official OpenAI Python or Node SDK unchanged; only the base URL and API key differ.
  • Real-time streaming — a WebSocket protocol that returns one audio frame per sentence as it is synthesized.
  • Multiple output formats — wav, mp3, pcm, opus, and flac.
  • Per-request observability — every HTTP response carries x-upstream-ms and x-gateway-ms timing headers.

​
Use cases

  • Conversational agents and voice assistants that need low first-audio latency.
  • Content production: narration, voiceovers, and character work with emotional range.
  • Product experiences that read dynamic text aloud with natural pacing.