@absolutejs/voice-openai

Input transcription log probabilities are requested by default and normalized as Transcript.tokens, with aggregate transcript confidence available for provider-neutral routing and calibration. Set inputTranscriptionLogprobs: false to omit the optional Realtime field.

OpenAI realtime adapter for @absolutejs/voice.

Install

bun add @absolutejs/voice @absolutejs/voice-openai

Setup

import { openai } from "@absolutejs/voice-openai";

const realtime = openai({
  apiKey: process.env.OPENAI_API_KEY!,
  model: "gpt-realtime-mini",
  voice: "marin",
  instructions: "Be concise and clear.",
  inputTranscriptionModel: "gpt-4o-mini-transcribe",
});

This adapter uses OpenAI's official Realtime WebSocket API and normalizes it into the core RealtimeAdapter contract.

What It Maps

• conversation.item.input_audio_transcription.delta -> partial
• conversation.item.input_audio_transcription.completed -> final
• completed input transcription -> normalized endOfTurn
• response.output_audio.delta -> audio
• realtime error events -> normalized error
• websocket close -> normalized close

For string input, the adapter emits a local final transcript immediately and then requests an OpenAI response. For audio input, it buffers PCM audio and commits the turn after a short inactivity window.

The response transcript events (response.audio_transcript.*, response.output_audio_transcript.*, response.output_text.*) are disabled by default to keep the adapter focused on user-input STT. Enable them only if your application intentionally consumes assistant output transcripts.

Input Audio Format

OpenAI Realtime currently expects pcm16 audio at 24kHz mono for streaming audio input. This adapter enforces that requirement when you call send(audio).

Output audio is normalized as raw pcm_s16le at 24kHz mono.

Options

Supported options include:

• apiKey
• model
• emitResponseTranscripts
• voice
• instructions
• inputTranscriptionModel
• inputTranscriptionLanguage
• inputTranscriptionPrompt
• maxOutputTokens
• autoCommitSilenceMs
• noiseReduction
• responseMode
• speed
• temperature

API Key

Set OPENAI_API_KEY in your runtime environment, or pass the key explicitly in the adapter config.

OpenAI references used for this adapter design:

• https://developers.openai.com/api/reference/resources/realtime
• https://developers.openai.com/api/docs/models