OpenAI-Compatible Transcription Endpoint Integration

Overview

Kamiwaza can register any external endpoint that speaks the OpenAI /v1/audio/transcriptions wire format and proxy it through the same deployment, audit, and access-control surfaces as other models. This covers OpenAI's Whisper, Microsoft Azure's Whisper and gpt-4o-transcribe deployments, and any self-hosted or third-party Whisper-compatible service.

Once registered, the endpoint exposes the standard /runtime/models/<id>/v1/audio/transcriptions route. Kamiwaza's Context Service uses this same route when transcribing audio uploaded by users in workrooms.

Prerequisites

• A provider API key with permission to call the audio-transcription endpoint.
• The transcription model is enabled on the account (or, for Azure, the deployment exists in the project).
• Outbound HTTPS from the Kamiwaza control plane to the provider hostname is permitted (see Provider examples below for hostnames).

OpenAI-specific: Restricted API keys must be granted the Audio scope in addition to chat. Keys created with the default restricted scope set return 401 on transcription calls. Either widen the scope or create a separate key with the Audio permission enabled.

Quick Start (UI)

The wizard is identical to the chat-endpoint wizard — only the models you tick on the Review step differ. See OpenAI-Compatible Chat for screenshots of each step.

1. In Kamiwaza, go to Models, click Add Model, then click Add external inference endpoint in the dialog header.

2. Source — choose OpenAI, Azure, or Other (OpenAI-compatible) under Where is your model hosted?, then click Next.

3. Setup — configure the endpoint:

   • Inference endpoint URL — The provider's API root. Examples:
     • OpenAI: https://api.openai.com/v1
     • Azure OpenAI Service: https://<your-resource>.openai.azure.com
     • Azure AI Foundry: https://<your-project>.services.ai.azure.com/models
     • Self-hosted Whisper: your internal hostname
   • Credentials — paste the provider API key, or pick Use existing credential if one is already stored for this URL.

   Click Connect. Kamiwaza validates the URL, authenticates, and discovers the models the endpoint exposes.

4. Review — tick the transcription model(s) you want to register (whisper-1, gpt-4o-transcribe, an Azure deployment name, etc.). Expand a row to customize the display name or per-model defaults. Click Register N models when done.

5. Deploy each newly registered model from the Models list.

Provider examples

Provider	Source	Inference endpoint URL	Underlying model
OpenAI Whisper	OpenAI	https://api.openai.com/v1	whisper-1, gpt-4o-transcribe
Azure OpenAI (Whisper, gpt-4o-transcribe)	Azure	https://<your-resource>.openai.azure.com	The Azure deployment name
Azure AI Foundry (gpt-4o-mini-transcribe)	Azure	https://<your-project>.services.ai.azure.com/models	The Foundry deployment name
Self-hosted Whisper	Other (OpenAI-compatible)	Your internal hostname	The service's published model identifier

If you've already registered a chat endpoint against the same provider URL, Kamiwaza recognizes the existing credential and offers Use existing credential. For OpenAI specifically, double-check the key has the Audio scope before reusing.

Credentials

Credentials registered through the form are encrypted on save and stored in the Kamiwaza secret catalog, keyed by base URL. Catalog reuse, rotation, and the five-minute propagation window are described in External endpoints overview.

API Usage

curl -X POST "https://<your-domain>/runtime/models/<deployment-id>/v1/audio/transcriptions" \
  -H "Authorization: Bearer $TOKEN" \
  -F "file=@recording.wav" \
  -F "model=whisper-1" \
  -F "response_format=json"

Supported request fields: file (the audio), model (the underlying model), language (BCP-47 code), and response_format. Other OpenAI Whisper fields (prompt, temperature) are not forwarded by Kamiwaza on this path. response_format accepts json (default), text, verbose_json, srt, and vtt; values outside that set are normalized to json.

Operational notes

• Audio uploads are streamed through Kamiwaza to the upstream provider. Kamiwaza does not retain audio bytes after the response is returned to the caller.
• Transcription errors from the provider are surfaced through the Kamiwaza response body. The requester, workroom, and deployment ID are recorded in the Kamiwaza audit log under event_type=model_invocation, regardless of upstream success or failure.
• This path is request/response only — Kamiwaza does not stream OpenAI-compatible transcription. For streaming speech-to-text, register an AWS Transcribe endpoint instead.

Next steps

• For the streaming and batch capabilities of AWS's managed transcription service, see AWS Transcribe.
• Register a chat endpoint against the same provider — see OpenAI-Compatible Chat.
• Contact Kamiwaza Support for help with non-standard providers, sovereign-cloud hostnames, or private connectivity.