On this page

Speech-to-Text (Whisper)

Turns the speech in your media into a transcript you can search.

The Whisper node transcribes the audio track of a media file into text using a Whisper speech-to-text model. The transcript is stored as a first-class component so it can be searched and summarized.

Kind

whisper

Applies to

Video, Audio

Input ports

audio or video — exactly one of the two is wired

Output ports

transcript — the utterances with their timings, ready to connect to LLM Enrichment, Sentiment Analysis or Text-to-Speech

Requirements

A Whisper runtime (whisper.cpp or compatible) and a model file. CPU works; a GPU accelerates transcription and is enabled via options. Storage for the model file.

Persists to

asset_transcript_comp + asset_node_result ledger

Configuration

The Whisper (Speech-to-Text) settings panel: Model Path
Figure 1. The node’s settings in the pipeline editor

Set these in the panel above, or in the node’s options block in a pipeline definition:

Option Meaning

modelPath

Path to the Whisper model file

language

Source language (or auto-detect)

useGpu / gpuDevice

Enable GPU acceleration and select the device

temperature / temperatureInc

Decoding temperature and fallback increment

Seeing it run

Turn on Debug Mode and every node keeps what it produced, on the card itself — for this node, the transcript and the segments it was cut into. Below is a real run of ggml-large-v3-turbo over a recording of a speech.

The whisper node after a run

One port, one document. transcript carries the segments as JSON — each with its text and its time span — so a downstream Script node can cut chapters out of it and Translate can take the prose.

Use Cases

  • Subtitles & captions — generate transcripts for video and audio assets.

  • Transcript search — power the chat’s search_transcript tool and the transcript-summarizer skill.

  • Downstream enrichment — connect transcript to LLM classification or summarization.

Looking for something else?

All documentation Feature overview