Transcribe

Transcribe

Word-level timing is what makes clean cuts possible — every cut boundary should land between words, never mid-word. gpt-4o-transcribe produces better text but does not support timestamp_granularities, so for editing we always use whisper-1 with verbose_json.

When to skip

If runs/<take>/transcript.json already exists, do not re-transcribe. The harness pre-bakes it before handing off. Re-running costs ~$0.06/min.

Procedure

1. Extract audio with ffmpeg (mono, 16 kHz, 32 kbps mp3 — keeps under the 25 MB upload cap and is plenty for whisper):

   ffmpeg -y -i raw.mp4 -vn -ac 1 -ar 16000 -b:a 32k audio.mp3
   

2. Call the API (OpenAI key is in the env):

   curl --request POST \
     --url https://api.openai.com/v1/audio/transcriptions \
     --header "Authorization: Bearer $OPENAI_API_KEY" \
     --header 'Content-Type: multipart/form-data' \
     --form file=@audio.mp3 \
     --form model=whisper-1 \
     --form response_format=verbose_json \
     --form 'timestamp_granularities[]=word' \
     --form 'timestamp_granularities[]=segment' \
     > transcript.json
   

Output shape

transcript.json contains:

• text — full transcript as a single string
• segments[] — phrase-level groupings: {id, start, end, text, …}. Skim these first to map the territory.
• words[] — every word with {word, start, end} in seconds. Use these to lock exact cut times.

Rules

• Cut boundaries must align to word start / end from the words array. Do not interpolate — the next word's start is the right boundary if you want to preserve a trailing breath.
• For non-English audio, set --form language=<iso639> to bias the model.
• If the transcript misses a proper noun, re-run with --form prompt="..." containing the spelling. Whisper only reads the last 224 tokens of the prompt, so keep it short.