@absolutejs/voice-openai-whisper

Buffered OpenAI transcription adapter for @absolutejs/voice. It collects raw audio for a turn, wraps PCM input in a WAV container, and sends one transcription request when you flush or close the session.

Install

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

Setup

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

const stt = openaiWhisper({
  apiKey: process.env.OPENAI_API_KEY!,
  model: "gpt-4o-mini-transcribe",
  language: "en",
});

const session = stt.open({
  format: {
    channels: 1,
    encoding: "pcm_s16le",
    sampleRateHz: 16_000,
  },
});

session.on("final", ({ transcript }) => {
  console.log(transcript.text);
});

await session.send(audioChunk);
await session.flush();
await session.close("turn-complete");

Buffering behavior

This is a batch adapter, not a realtime streaming transport. send() appends audio to the current turn. flush() transcribes the accumulated buffer and emits final followed by endOfTurn. Closing flushes by default; set flushOnClose: false when the caller owns that lifecycle explicitly.

The default buffer ceiling is 20 MiB. Set maxBufferedBytes to a lower bound for latency or memory control, and flush between turns. Exceeding the bound throws before more audio is retained.

Options

• apiKey — required OpenAI API key.
• model — defaults to whisper-1; transcription models are accepted.
• language — optional language hint; fixed language strategies from voice core take precedence.
• prompt, temperature, and responseFormat — forwarded to the transcription endpoint.
• organization — sends the OpenAI-Organization header.
• baseUrl — targets an OpenAI-compatible endpoint.
• fetch — injects a custom fetch implementation for policy, tracing, or tests.
• flushOnClose and maxBufferedBytes — control buffering lifecycle.

Errors and events

HTTP failures and malformed provider responses emit a non-recoverable error. Successful responses emit normalized final and endOfTurn events through the standard @absolutejs/voice STTAdapter contract. The adapter does not emit partial transcripts because the provider request happens only at flush time.