Class OpenAIWhisperSpeechToTextProvider

Speech-to-text provider that uses the OpenAI transcription API.

API Contract

  • Endpoint: POST {baseUrl}/audio/transcriptions
  • Authentication: Authorization: Bearer <apiKey>
  • Content-Type: multipart/form-data (FormData with file blob)
  • Response format: Controlled by the response_format field. When the caller names none, OpenAI's gpt- transcription models are asked for json, and every other model, whisper-1 and a whisper-compatible server's own included, for verbose_json (segments, language and duration).

Models

  • gpt-transcribe (default): the model OpenAI recommends for recorded speech. It returns the text and the detected languages, and it takes language hints as a languages list.
  • whisper-1: removed from the API on 2027-02-26. It serves segment timestamps, srt and vtt, so a call that asks for one of those formats runs on it unless the call names a model or the configured model serves the format itself. A configured gpt- model does not.

Supported Response Formats

  • verbose_json: full JSON with segments, duration and language (whisper-1)
  • json: JSON with the text (the newer models also return languages)
  • text: plain text, no JSON
  • srt: SubRip subtitles (whisper-1)
  • vtt: WebVTT subtitles (whisper-1)

When text, srt, or vtt format is used, the response is returned as plain text and segments are not available.

See

OpenAIWhisperSpeechToTextProviderConfig for configuration options See normalizeSegments() for the segment normalization logic.

Example

const provider = new OpenAIWhisperSpeechToTextProvider({
  apiKey: process.env.OPENAI_API_KEY!,
});
// Runs on gpt-transcribe and answers in json.
const result = await provider.transcribe(
  { data: audioBuffer, mimeType: 'audio/wav', fileName: 'recording.wav' },
  { language: 'en' },
);
// Segment timestamps come from whisper-1.
const timed = await provider.transcribe(
  { data: audioBuffer, mimeType: 'audio/wav', fileName: 'recording.wav' },
  { responseFormat: 'verbose_json' },
);

Implements

Constructors

constructor

Methods

getProviderName transcribe

Properties

id displayName supportsStreaming

Constructors

constructor

Methods

getProviderName

transcribe

  • transcribe(audio, options?): Promise<SpeechTranscriptionResult>

  • Transcribes an audio buffer using the OpenAI Whisper API.

    The audio is sent as a multipart form upload with the file, model, and optional parameters (language, prompt, temperature, response_format).

    Parameters

    • audio: SpeechAudioInput

      Raw audio data and metadata. The data buffer is wrapped in a Blob and sent as a form file field. If fileName is not provided, a default name is generated from the format field.

    • options: SpeechTranscriptionOptions = {}

      Optional transcription settings including language hint, context prompt, temperature for sampling, and response format.

    Returns Promise<SpeechTranscriptionResult>

    A promise resolving to the normalized transcription result.

    Throws

    When the OpenAI API returns a non-2xx status code.

    Example

    const result = await provider.transcribe(
      { data: mp3Buffer, mimeType: 'audio/mpeg', fileName: 'voice.mp3' },
      { language: 'fr', prompt: 'Discussion about AI' },
    );

Properties

Readonly id

id: "openai-whisper" = 'openai-whisper'

Unique provider identifier used for registration and resolution.

Readonly displayName

displayName: "OpenAI Whisper" = 'OpenAI Whisper'

Human-readable display name for UI and logging.

Readonly supportsStreaming

supportsStreaming: false = false

Whisper API is batch-only; streaming requires a WebSocket adapter.

Settings

Member Visibility

Theme

On This Page

constructorgetProviderNametranscribeiddisplayNamesupportsStreaming