`generate_transcript`

Transcribes audio via OpenAI Whisper, producing word-level timestamps.

When to use

Run generate_transcript early in your pipeline. Its output is consumed by deadspace_trimmer, captions, broll_place, broll_veo_generate, and ai_agent.

In the editor

Add it from the Library (Transcript group), or with + add step and a search for generate_transcript, early in the pipeline: captions, b-roll placement and silence trimming all read what it produces.

Under Inputs, add source as a file input naming your recording. Set language if you know it, or leave it empty and it will be worked out from the audio. The two optional inputs are for narrower cases: cuts, a step input, transcribes only the parts an earlier step kept rather than the whole file, and footage, a step input, transcribes an assembled edit that draws on several files.

This step needs your own OpenAI key. Add it in the workspace's Variables & Service API Keys, from the editor's Settings. If it is missing, the run stops and asks for it rather than failing halfway through.

Nothing appears on the timeline. The result is the words with the exact time each one was said, which is what makes captions land in sync and lets other steps know what the video is about. Wire this step into those steps as their transcript input.

YAML Example

# Basic transcription
- id: transcribe_main
  processor: generate_transcript
  input:
    - id: source
      file: speaker-footage
  params:
    language: "en"
  cache: true

# Re-transcribe only kept segments after an AI cut
- id: transcribe_cut
  processor: generate_transcript
  input:
    - id: source
      file: speaker-footage
    - id: cuts
      step: ai_cut        # only transcribes kept segments; timestamps remapped to source
  params:
    language: "en"
  cache: true

Inputs

ID	Source	Description
`source`	file	Source video or audio file
`cuts`	step (optional)	Edit decisions from a previous step: only those segments are transcribed
`footage`	step (optional)	Edit decisions from `footage_assembler`: assembles multi-source backbone audio and transcribes it as a timeline-relative transcript

Params

Param	Type	Default	Description
`language`	string	none	ISO 639-1 language code (e.g. `"en"`, `"es"`). Empty = auto-detect.
`backbone_level`	level	none	Which timeline level to use for audio assembly when `footage` input is provided

Output

Analysis containing a Transcript with word-level timestamps.

Notes

Requires an OPENAI_API_KEY set in your workspace's Service API Keys.
Always use cache: true: transcription is slow and the result never changes for the same audio.
When cuts input is provided, only the referenced source segments are sent to Whisper. Timestamps are remapped back to source file positions.