Skip to content
Syrinx

STT reconfigure

Syrinx exposes a vendor-agnostic seam to reconfigure a streaming STT provider mid-call, without tearing down the session — the mechanism behind conversation-state-driven biasing.

The seam

Section titled “The seam”

SttReconfigurePartial (in @kuralle-syrinx/core) is the vendor-neutral shape a provider applies what it can from and ignores the rest:

interface SttReconfigurePartial {
  keyterms?: readonly string[];        // bias recognition toward domain terms
  eotThreshold?: number;               // end-of-turn thresholds (Flux)
  eagerEotThreshold?: number;
  eotTimeoutMs?: number;
  endpointingMs?: number;              // silence endpointing (Nova)
  vadThreshold?: number;
  languageHints?: readonly string[];   // soft language bias
  language?: string;                   // HARD language switch (e.g. "es-ES", or Nova-3 "multi")
  contextText?: string;                // AssemblyAI-style agent-context biasing
}

A stt.reconfigure bus packet routes through VoiceAgentSession to the plugin’s sttReconfigure.reconfigure(partial).

languageHints vs language

Section titled “languageHints vs language”
  • languageHints (soft) — weight the recognizer toward one or more languages while it still auto-detects. Applied in-band by Deepgram Flux via its Configure message (no reconnect).
  • language (hard) — change the recognition language outright, e.g. "en-US" → "es-ES", or Nova-3 "multi" for code-switch. Deepgram Nova applies it on reconnect (rebuilt URL) and re-stamps stt.result.language; Flux is model-fixed (flux-general-en) and ignores it, relying on languageHints.

Provider support

Section titled “Provider support”

The seam is provider-agnostic, but only Deepgram Nova (reconnect) and Deepgram Flux (in-band Configure) implement sttReconfigure today. Grok, Google, and ElevenLabs STT don’t yet — a reconfigure is a no-op there (warned). Adding it to another provider is the Nova reconnect pattern via stt-core’s StreamingSttSession.reset().