generate-transcript

The generate-transcript transform runs speech-to-text over audio files listed in an upstream audio-metrics JSON document, using PCM bytes from convert-to-pcm outputs keyed by MD5.

When to use it

Use after convert-to-pcm and audio-metrics in an audio analytics pipeline when you need searchable transcripts (plain text or time-segmented JSON).

It also accepts audio directly, with no metrics document alongside it, which is how hydrate-protobuf SignalPDU feeds it. Two artefact shapes are recognised on that path:

A single WAV: transcribed whole, and the transcript is keyed by the audio md5.
An origin-audio ZIP (groupByOrigin: true): each origins/<originUuid>.wav is transcribed separately, and the transcript is keyed by originUuid.

Per-origin transcripts

Under groupByOrigin the audio artefact is a delta window, not a cumulative file, so this node carries history forward itself:

Each origin stream is transcribed, then stamped with wall clock: segments keep relative start / end seconds inside the stream and gain absolute startMs / endMs from the manifest’s streamStartUnixMs.
The node loads its own previous transcript and merges the new window into it per originUuid. Segments repeated across windows are deduplicated on (originUuid, startMs, endMs), the surviving segments are ordered by wall clock, and each entry’s text is rebuilt from them so it never disagrees with its segments.
meta.lastAudioMd5 and meta.sourceAudioMd5 record the audio this run consumed. A reprocess over unchanged audio completes without uploading, so a transcript blob is not churned and dashboards are not rescheduled.

Dashboards should prefer startMs / endMs when present and fall back to start / end seconds for legacy single-WAV transcripts.

Limitations

On the metrics-driven path, requires audio-metrics JSON and matching PCM blobs in upstream inputs; missing metrics with no direct audio payload fails the node.
Supported model values: whisperx (default) and faster-whisper, which share the speech weights baked into the runner image at build time, so they run without network access. Shipped images carry small (the default, WHISPER_MODEL_SIZE) and base (WHISPER_EXTRA_MODEL_SIZES). wav2vec downloads its model from Hugging Face on first use and whisper needs the openai-whisper package, which the image does not ship: neither works on an air-gapped cluster.
outputFormat: plain or segments (default segments). Invalid values fail the node.

Configuration (summary)

Property	Default	Description
`outputFormat`	`segments`	`plain` or `segments`
`includeTimestamps`	`true`	Include timing in segment output
`model`	`whisperx`	ASR backend (see supported list above)
`modelSize`	the image default (`WHISPER_MODEL_SIZE`, `small` in shipped images)	Model size passed to the backend. Only sizes baked into the image load offline: `small` and `base` in shipped images. `small` is more accurate than `base` and takes about two and a half times as long.
`device`	`cpu`	Inference device. Leave it at `cpu`: the shipped image carries CPU-only inference libraries, so `cuda` fails.
`language`	`en`	Transcription language code
`initialPrompt`	(none)	Text the model is primed with before each piece of audio, for example the kind of traffic (`"Military air combat radio."`). Optional.
`vocabulary`	(none)	The scenario’s callsigns, codewords and other words the transcript must get right: a list of strings, or an object of named lists (`{"callsigns": ["TAMIM 11"], "codes": ["WINCHESTER"]}`). The model is primed with every term once, after `initialPrompt`, and writes them as given far more reliably. Recorded in the transcript’s `meta`. Optional.
`dataSourceName`	(runner default)	Object store for outputs

Manifest schema lists outputFormat and includeTimestamps. The runtime also accepts model, modelSize, device, language, initialPrompt and vocabulary as used by the processor.

Keep scenario words in the session template’s node configuration, never in the runner: they change with every exercise. Whisper reads only about 220 tokens of prompt and drops the front of a longer one, so keep the vocabulary to the words that matter (a few dozen terms); the runner logs a warning when the prompt grows long. Runner builds that predate these keys ignore them.

Audio Utils
convert-to-pcm
audio-metrics
voice-analysis
hydrate-protobuf (SignalPDU audio, including groupByOrigin)

HEAT documentation team. Last updated 14 October 2026.

Opens your browser's print dialog. Choose "Save as PDF" as the destination.