@interview-sdk/adapter-elevenlabs

ElevenLabs voice provider adapter for @interview-sdk/core, built on @elevenlabs/elevenlabs-js — the current package (the unscoped elevenlabs package is deprecated and has moved here). Implements both synthesize() (text-to-speech, the primary use case) and transcribe() (ElevenLabs also ships speech-to-text).

Install

npm install @interview-sdk/core @interview-sdk/adapter-elevenlabs

Usage

import { ElevenLabsAdapter } from '@interview-sdk/adapter-elevenlabs';

const voiceAdapter = new ElevenLabsAdapter({ apiKey: process.env.ELEVENLABS_API_KEY });

<InterviewWidget>'s transcribe/synthesize props are plain functions ((audio: Blob) => Promise<string> / (text: string) => Promise<SynthesisResult>), not a VoiceProviderAdapter instance — synthesize matches directly, but transcribe needs a one-line adapter since VoiceProviderAdapter.transcribe takes an ArrayBuffer/Uint8Array and returns a TranscriptResult:

<InterviewWidget
  transcribe={async (audio) => (await voiceAdapter.transcribe(await audio.arrayBuffer())).text}
  synthesize={voiceAdapter.synthesize.bind(voiceAdapter)}
  // ...
/>

ElevenLabsAdapter accepts an optional voiceId (defaults to ElevenLabs' prebuilt "Rachel" voice), speakModel (TTS, defaults to eleven_multilingual_v2), transcribeModel (STT, defaults to scribe_v2), and an optional pre-configured client for testing.

Behavior

• synthesize(text) requests outputFormat: 'mp3_44100_128' explicitly (so the adapter can confidently report mimeType: 'audio/mpeg') and drains the SDK's ReadableStream<Uint8Array> response into a single ArrayBuffer.
• transcribe(audio) passes the raw ArrayBuffer/Uint8Array directly to speechToText.convert — no stream wrapper needed, the SDK's Uploadable type accepts binary data natively.
• Normalizes every ElevenLabsError (a base class with a numeric .statusCode; typed subclasses exist per status but this adapter branches on the code directly, matching the Deepgram adapter's pattern) onto the shared Provider*Error taxonomy from @interview-sdk/core.
• Auto-retries 408/409/429/5xx by default (maxRetries: 2), matching Claude's and OpenAI's SDK behavior out of the box.

Verified against

@elevenlabs/[email protected], current as of 2026-07-04 — the package itself was renamed from unscoped elevenlabs; model names and output formats shift quickly. Re-check before assuming this stays accurate for long.