Azure AI Speech — Speech-to-Text

Transcribe audio to text with Azure Fast Transcription — synchronous,
word-level timestamps, speaker diarization, and multi-language identification.
In OpenMontage this is exposed through the azure_stt tool (capability=analysis,
provider=azure). It is an optional cloud STT provider — when
AZURE_SPEECH_KEY is configured, prefer it for cloud transcription. The local
transcriber tool (faster-whisper) remains the default offline path and the
fallback when Azure is unavailable.

Docs: Fast Transcription · Speech service overview

Why Fast Transcription (not Batch)

Azure exposes three STT surfaces. OpenMontage uses Fast Transcription because
the pipeline transcribes local audio files:

Surface	Input	Latency	Needs
Fast Transcription (used here)	local file, multipart POST	synchronous, sub-real-time	key + region
Batch Transcription	audio at a URL (Blob + SAS)	async job + polling	Blob storage plumbing
Speech SDK (spx)	mic / stream / file	streaming	native azure-cognitiveservices-speech package

Fast Transcription needs no Blob storage, no SAS URLs, and no native SDK — just
requests and the two env vars.

Setup

Create a Speech resource in the Azure portal;
copy the key and region from its Keys and Endpoint page.

export AZURE_SPEECH_KEY=your_speech_resource_key
export AZURE_SPEECH_REGION=eastus          # your resource's region
# export AZURE_SPEECH_ENDPOINT=https://...  # optional: overrides region

azure_stt reports AVAILABLE once AZURE_SPEECH_KEY plus either
AZURE_SPEECH_REGION or AZURE_SPEECH_ENDPOINT are set.

Using it in a pipeline

Prefer azure_stt over transcriber unless the run must be offline. Its output
matches the transcriber schema exactly, so it is a drop-in for subtitle_gen
and any stage that consumes a transcript.

from tools.tool_registry import registry
registry.discover()
stt = registry._tools["azure_stt"]

result = stt.execute({
    "input_path": "projects/my-video/assets/audio/narration.mp3",
    # "language": "en",          # ISO 639-1 or BCP-47 ("en-US"); omit for auto-ID
    # "diarize": True,           # speaker labels, no HuggingFace token needed
    # "max_speakers": 4,
    "output_dir": "projects/my-video/artifacts",
})
if result.success:
    segs = result.data["segments"]          # [{id,start,end,text,words:[...]}]
    words = result.data["word_timestamps"]  # flat [{word,start,end,probability}]

If azure_stt is unavailable (no key) or errors, fall back to transcriber
(local whisper) — its execute signature and output are identical.

Parameters that matter

• language — pass an ISO code ("en") or a full locale ("en-US"). Pin it
  when you know the language; it is faster and more accurate than auto-ID.
• candidate_locales — when language is omitted, Azure runs language
  identification across this shortlist. Narrow it to the languages you actually
  expect; a huge list slows detection and invites misclassification.
• diarize / max_speakers — enable for multi-speaker audio (interviews,
  podcasts). Set max_speakers to the real upper bound.
• profanity_filter — None | Masked (default) | Removed | Tags.

Response shape (mapped to the transcriber schema)

The raw Azure response (phrases[] with offsetMilliseconds / words[]) is
converted to seconds and the OpenMontage transcript schema:

{
  "segments": [
    {"id": 0, "start": 0.0, "end": 2.4, "text": "Hello world",
     "speaker": 1,
     "words": [{"word": "Hello", "start": 0.0, "end": 0.5, "probability": 0.98}]}
  ],
  "word_timestamps": [{"word": "Hello", "start": 0.0, "end": 0.5, "probability": 0.98}],
  "language": "en-US",
  "duration_seconds": 2.4,
  "provider": "azure"
}

Note: Fast Transcription has no per-word confidence, so each word carries the
phrase confidence in probability.

Limits & tips

• Single file up to ~2 hours / a few hundred MB per request. For longer or bulk
  jobs, use Azure Batch Transcription instead.
• Send clean audio (16 kHz+ mono is plenty). Transcode video to audio first if
  you only need speech — smaller upload, same result.
• Verify timing: word timestamps drive subtitle cues in subtitle_gen. Spot-check
  the first and last cues against the source audio.