Skip to content

Latest commit

 

History

History
132 lines (90 loc) · 3.16 KB

README-transcribe.md

File metadata and controls

132 lines (90 loc) · 3.16 KB

Azure Speech Batch Transcription Function

Small Azure Functions HTTP endpoint for Azure Speech-to-Text Batch Transcription. Used as part of the video-annotator project.

Workflow:

  1. Provide a Blob SAS URL to an audio file (media_url)
  2. Function submits a batch job and returns a job_url
  3. Call again with job_url to get status; when Succeeded, it returns normalized transcript output.

Repo layout

  • TranscribeHttp/

    • __init__.py — HTTP function (submit + status/result)
    • function.json — HTTP trigger bindings

  • shared/speech_batch.py — Speech + Blob helpers (submit job, list files, normalize output)

  • make_sas.py — generate a SAS URL for a blob (for testing)

  • test_transcribe.py — local test driver

  • local.settings.json — local dev settings (do not commit secrets)

  • requirements.txt, host.json

Prereqs

  • Python 3.11+
  • Azure Functions Core Tools v4
  • Azure Speech resource (SPEECH_KEY, region or endpoint)
  • Azure Storage account with a container that holds your audio (defaults assume speech-input)

Install deps:

python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

Configuration

Local config is read from local.settings.json:

  • SPEECH_KEY

  • SPEECH_REGION (optional if SPEECH_ENDPOINT is set)

  • SPEECH_ENDPOINT (e.g. https://eastus.api.cognitive.microsoft.com/)

  • SPEECH_API_VERSION (e.g. 2025-10-15)

  • Storage vars if you’re using the helper scripts:

    • AZURE_STORAGE_ACCOUNT
    • AZURE_STORAGE_KEY
    • INPUT_CONTAINER (default speech-input)
    • OUTPUT_CONTAINER (default speech-output)

Run locally

Start the Functions host:

func start

1) Get a SAS URL for an audio blob

If you already uploaded an audio file to Blob Storage, run:

python make_sas.py

It prints a SAS URL you can use as media_url.

2) Submit a transcription job

curl -sS -X POST "http://localhost:7071/api/TranscribeHttp" \
  -H "Content-Type: application/json" \
  -d '{
    "media_url": "https://<account>.blob.core.windows.net/<container>/<file>.m4a?<sas>",
    "locale": "en-US",
    "display_name": "measles_short"
  }' | jq

Response (202):

{ "job_url": "https://.../speechtotext/transcriptions/<id>?api-version=..." }

3) Check status / fetch transcript

curl -sS -X POST "http://localhost:7071/api/TranscribeHttp" \
  -H "Content-Type: application/json" \
  -d '{ "job_url": "PASTE_JOB_URL_HERE" }' | jq

Possible responses:

  • Running:
{ "status": "Running", "job_url": "..." }
  • Failed:
{ "status": "Failed", "job": { ... } }
  • Succeeded:
{ "status": "Succeeded", "result": { "utterances": [...], "words": [...] } }

Notes

  • Batch jobs may sit in Running due to queueing; poll until Succeeded/Failed.
  • Normalization returns utterances and (if present) words with timestamps in milliseconds.
  • The shared submit payload uses channels: [0] to avoid duplicated results from stereo audio.

Deploy

Set the same configuration values as app settings in your Azure Function App, then publish:

func azure functionapp publish <APP_NAME> --python