@kuralle-syrinx/elevenlabs

ElevenLabs voice provider for Syrinx — multi-context WebSocket streaming TTS and Scribe v2 Realtime streaming STT.

Install

pnpm add @kuralle-syrinx/elevenlabs

Auth

Pass api_key (and voice_id for TTS) in plugin config. For local smokes, set ELEVENLABS_API_KEY in the repo-root .env (the live smoke reads EL_VOICE_ID for the TTS voice, defaulting to a free-accessible premade voice).

Streaming TTS — ElevenLabsTTSPlugin

import { ElevenLabsTTSPlugin } from "@kuralle-syrinx/elevenlabs";
import { createNodeWsSocket } from "@kuralle-syrinx/ws/node";

const tts = new ElevenLabsTTSPlugin(createNodeWsSocket);
await tts.initialize(bus, {
  api_key: process.env.ELEVENLABS_API_KEY!,
  voice_id: process.env.ELEVENLABS_VOICE_ID,
  model_id: "eleven_flash_v2_5",
  sample_rate: 16000,
});

Connects to wss://api.elevenlabs.io/v1/text-to-speech/<voice_id>/multi-stream-input (multi-stream-input, concurrent contexts keyed by context_id) via the shared tts-core deep module. Sends the required InitializeConnectionMulti-style context-init frame (voice_settings + optional generation_config) before a context's first text. output_format and generation_config are dev-configurable, not pinned. Default voice (EXAVITQu4vr4xnSDxMaL) is a premade voice accessible to free API accounts — library voices (e.g. Rachel) require a paid plan; override with voice_id.

Usage billing (usage.recorded{stage:"tts", characters}) fires on audio received, once per context — not on isFinal (ElevenLabs streams audio with isFinal:null, and a rejected generation returns an empty final that must not be billed).

Streaming STT — ElevenLabsSTTPlugin

import { ElevenLabsSTTPlugin } from "@kuralle-syrinx/elevenlabs";
import { createNodeWsSocket } from "@kuralle-syrinx/ws/node";

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

Connects to wss://api.elevenlabs.io/v1/speech-to-text/realtime (Scribe v2 Realtime, model scribe_v2_realtime by default) via the shared @kuralle-syrinx/stt-core session. Provider partial_transcript → stt.interim, committed_transcript → stt.result (speechFinal: true). Usage billing (usage.recorded{stage:"stt", audioSeconds}) fires at the final-transcript funnel with delta-billing. Owns endpointing (endpointingCapability.owner: "provider_stt").

Pricing

Cited in @kuralle-syrinx/core's DEFAULT_PRICE_CATALOG: Scribe v2 STT $0.39/hr; Flash/Turbo TTS $50/1M characters, Multilingual TTS $100/1M characters.

Status

Live-verified end-to-end (TTS audio + usage, STT transcript + usage).

Live smoke

From examples/02-hello-voice-headless (requires ELEVENLABS_API_KEY; optional EL_VOICE_ID):

pnpm -C examples/02-hello-voice-headless exec tsx scripts/spike-elevenlabs.ts

Exercises both TTS (multi-context WS) and STT (Scribe v2 realtime WS) against the live API, including usage.recorded on both stages.