STT providers

Every STT plugin here is a wire protocol over the shared @kuralle-syrinx/stt-core lifecycle — reconnect, the interim/final funnel, and usage billing all work identically regardless of which one you pick.

Deepgram Nova

The flagship cascade STT. Silence-based endpointing with a Finalize handshake state machine for accurate turn boundaries.

import { DeepgramSTTPlugin } from '@kuralle-syrinx/deepgram';

const stt = new DeepgramSTTPlugin(socketFactory);
await stt.initialize(bus, {
  api_key: process.env.DEEPGRAM_API_KEY!,
  model: 'nova-3',
  language: 'en-US',
  sample_rate: 16000,
  endpointing: 300,
  keyterm: ['Syrinx', 'onboarding'],
});

Key	Default	Notes
api_key	—	Required.
model	nova-3
language	en-US
sample_rate	16000
endpointing	300	Silence, in ms, before a turn is considered finished.
keyterm	—	Repeatable; biases recognition toward domain terms.
encoding	linear16

Supports mid-call reconfigure (keyterms, endpointing, language) via a reconnect at the turn boundary.

Deepgram Flux

A turn-aware model — one API call handles transcription and semantic end-of-turn, so you don’t need a local endpointer. Runs over a plain WebSocket, which makes it the only semantic-EOT option available on the Cloudflare Workers edge.

import { DeepgramFluxSTTPlugin } from '@kuralle-syrinx/deepgram';

const stt = new DeepgramFluxSTTPlugin(socketFactory);
await stt.initialize(bus, {
  api_key: process.env.DEEPGRAM_API_KEY!,
  eot_threshold: 0.7,
  eager_eot_threshold: 0.5,
  eot_timeout_ms: 5000,
});

Key	Default	Notes
api_key	—	Required.
model	flux-general-en
eot_threshold	0.7	Confidence required to fire EndOfTurn.
eager_eot_threshold	unset	Set to enable eager mode — an early, retractable end-of-turn signal for speculative generation.
eot_timeout_ms	5000
keyterm	—
language_hint	—	Soft language bias — see languageHints vs language.

Supports mid-call reconfigure (keyterms, EOT thresholds, language hints) in-band, without a reconnect.

ElevenLabs — Scribe v2 Realtime

import { ElevenLabsSTTPlugin } from '@kuralle-syrinx/elevenlabs';

const stt = new ElevenLabsSTTPlugin(socketFactory);
await stt.initialize(bus, {
  api_key: process.env.ELEVENLABS_API_KEY!,
  sample_rate: 16000,
});

Key	Default	Notes
api_key	—	Required.
sample_rate	16000

Owns its own endpointing (provider_stt) — you don’t need a separate VAD or endpointer alongside it.

Google Cloud Speech-to-Text v2

import { GoogleSTTPlugin } from '@kuralle-syrinx/google';

const stt = new GoogleSTTPlugin(socketFactory);
await stt.initialize(bus, {
  api_key: process.env.GOOGLE_API_KEY!,
  project_id: process.env.GOOGLE_PROJECT_ID!,
  language: 'en-US',
  model: 'latest_long',
  confidence_threshold: 0,
});

Key	Default	Notes
api_key	—	Required.
project_id	—	Required.
language / language_codes	en-US	language_codes accepts multiple for auto-detect.
model	latest_long
location	global
recognizer	_
confidence_threshold	0	Results below this confidence are dropped rather than emitted.
encoding	LINEAR16

Grok (xAI)

import { GrokSTTPlugin } from '@kuralle-syrinx/grok/stt';

const stt = new GrokSTTPlugin(socketFactory);
await stt.initialize(bus, {
  api_key: process.env.XAI_API_KEY!,
  language: 'en',
  sample_rate: 16000,
});

Key	Default	Notes
api_key	—	Required.
language	en
sample_rate	16000

Choosing between them

• Turn-taking already solved and want to run on the Workers edge? Deepgram Flux is the only semantic-EOT provider that doesn’t need a local ONNX model.
• Want the most battle-tested silence-based endpointing? Deepgram Nova.
• Already on ElevenLabs for TTS? ElevenLabs STT shares your account and billing.
• On GCP already? Google Cloud Speech-to-Text v2.

See STT reconfigure for which of these support mid-call biasing, and Usage & pricing for cited rates.