Whisper Transcription Service

Available from Release 2026.06

Overview

Starting with release 2026.06, the speech service supports OpenAI Whisper as an alternative transcription backend alongside Microsoft Azure Speech-to-Text. Whisper can be deployed in two ways:

• Azure OpenAI Whisper — cloud-managed via your existing Azure OpenAI resource

• Whisper OX — self-hosted using faster-whisper via Ray Serve, for on-premise / data-sovereignty scenarios

For general voice architecture, provisioning, and troubleshooting, refer to: - Voice Infrastructure — architecture diagrams, deployment, and MS STT setup - Voice Administration — admin configuration and permissions - Using Voice — end-user documentation

When to Choose Whisper

How Whisper Differs from MS STT

The existing MS STT provider streams audio in real-time with interim results. Whisper providers use chunked batch processing:

1. Audio arrives from the browser as PCM chunks via WebSocket (same as MS STT)

2. The speech service buffers audio for a configurable duration

3. The buffer is converted to WAV and sent to the Whisper endpoint

4. A transcript is returned to the browser

5. On stop, remaining audio is sent as a final chunk

The chunk duration is controlled by WHISPER_CHUNK_DURATION_SECONDS:

Option A: Azure OpenAI Whisper

Use this if you already have an Azure OpenAI resource.

Prerequisites

1. An Azure OpenAI resource with a Whisper model deployment

2. API key for the resource

3. Network connectivity from the speech service pods to the Azure OpenAI endpoint

Provisioning

1. In the Azure Portal, navigate to your Azure OpenAI resource

2. Under Model Deployments, deploy the whisper model

3. Note the endpoint URL and API key

Environment Variables (Speech Service)

Budget

Refer to Azure OpenAI pricing for current Whisper rates.

Option B: Whisper OX (Self-Hosted)

Use this for on-premise or data-sovereignty scenarios where audio must not leave your network.

Prerequisites

1. A server or Kubernetes cluster with GPU support (recommended) or sufficient CPU

2. Network connectivity from the speech service to the Whisper OX endpoint

3. A deployed Whisper OX server implementing the API contract below

API Contract

The endpoint must accept POST with multipart/form-data:

Parameters sent in generate_kwargs:

{
  "multilingual": true,
  "vad_filter": true,
  "task": "transcribe",
  "language": "en",
  "batch_size": 8,
  "beam_size": 5,
  "condition_on_previous_text": true
}

Expected response:

{
  "segments": [
    {
      "start": 0.0,
      "end": 2.5,
      "text": "Hello world",
      "sequence_confidence": -0.25,
      "words": null
    }
  ],
  "language": "en",
  "duration": 2.5
}

 Sizing Guide

"Nx realtime" = N seconds of audio transcribed per 1 second of wall-clock time.

A reference implementation is available in the Unique monorepo at .local/whisper-ox/.

Environment Variables (Speech Service)

Security

• The endpoint supports optional API key authentication. Set WHISPER_OX_API_KEY and optionally WHISPER_OX_API_KEY_HEADER (defaults to X-API-Key) on the speech service, and the corresponding WHISPER_API_KEY / WHISPER_API_KEY_HEADER on the Whisper OX server. When unset, no authentication is enforced — secure at the network level (private endpoints, NetworkPolicy, service mesh).

• Ensure TLS if traffic crosses network boundaries

• Audio is sent as raw WAV — treat the endpoint as sensitive infrastructure

Frontend: Enabling Whisper Mode on Assistants

The assistant must be configured to use Whisper mode. Set the following in assistant settings:

When set, the frontend shows a waveform during recording and a processing indicator while waiting for the transcript. The WebSocket URL (SPEECH_BACKEND_API_URL) does not change — the backend handles provider routing transparently.

To revert an assistant to live streaming, set speechToTextMode back to live.

Troubleshooting

Enable debug logging with LOG_LEVEL=debug on the speech service for detailed chunk sizes, API calls, and transcription timings.

Migration Checklist

Enabling Whisper (from MS STT)

1. Deploy the Whisper backend (Azure OpenAI or Whisper OX)

2. Set TRANSCRIPTION_SERVICE and endpoint env vars on the speech service

3. Restart the speech service

4. Set sttConfig.speechToTextMode: "whisper" on target assistants

5. No frontend deployment changes needed

Rolling Back to MS STT

1. Remove or set TRANSCRIPTION_SERVICE=MS_STT

2. Restart the speech service

3. Set sttConfig.speechToTextMode: "live" on assistants