@interview-sdk/adapter-deepgram

@interview-sdk/adapter-deepgram

Deepgram voice provider adapter for @interview-sdk/core, built on @deepgram/sdk v5. Implements both transcribe() (speech-to-text, the primary use case) and synthesize() (Deepgram also ships TTS).

Install

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

Usage

import { DeepgramAdapter } from '@interview-sdk/adapter-deepgram';

const voiceAdapter = new DeepgramAdapter({ apiKey: process.env.DEEPGRAM_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)}
  // ...
/>

DeepgramAdapter accepts an optional transcribeModel (defaults to nova-3), speakModel (TTS, defaults to aura-2-thalia-en), and an optional pre-configured client for testing.

Behavior

• transcribe(audio) passes the raw ArrayBuffer/Uint8Array directly to listen.v1.media.transcribeFile — no stream wrapper needed, since v5's Uploadable type accepts binary data natively — with smart_format: true for readable punctuation/casing, and reads results.channels[0].alternatives[0].transcript from the response.
• synthesize(text) requests container: 'wav' / encoding: 'linear16' explicitly so the adapter can confidently report mimeType: 'audio/wav' back to the caller, rather than guessing from Deepgram's own defaults.
• Normalizes every DeepgramError (a single generic class with a numeric .statusCode, not a typed hierarchy) onto the shared Provider*Error taxonomy from @interview-sdk/core, the same as the text-generation adapters.
• v5 auto-retries 429/5xx/connection errors by default (maxRetries: 2), matching Claude's and OpenAI's SDK behavior out of the box — unlike Gemini, no extra opt-in was needed here.

Verified against

@deepgram/sdk@5.5.0, current as of 2026-07-04. v5 is a breaking rewrite from v4 — the client is new DeepgramClient() (not v4's createClient()), and methods live under versioned namespaces (listen.v1.media.transcribeFile, not v4's listen.prerecorded.transcribeFile). Re-check before assuming this stays accurate for long.