Transcribe Video (local Whisper)

Overview

Drives the official whisper CLI (the openai-whisper pip package) to turn a local video or audio file into transcripts/subtitles. Whisper calls ffmpeg itself to decode the media, so any container ffmpeg can read works directly — no separate audio extraction step is needed.

The flow is: take a file path → preflight the tools and the file → ask the user how they want Whisper configured → run it → report where the outputs landed (always next to the source file).

Steps

1. Get the source path

The user usually pastes a path (or passes it as the skill argument). If none was given, ask for the absolute path to the video/audio file before doing anything else.

Normalise it: strip surrounding quotes and a file:// prefix, and expand ~. Keep the absolute path — you'll need the file's parent directory for the output location.

2. Preflight — tools and file

Run these checks and stop with a clear message if any fail (don't push forward and let Whisper error out cryptically):

• whisper on PATH — command -v whisper. If missing, tell the user to install it: pip install -U openai-whisper (and that it needs Python). Stop.
• ffmpeg on PATH — command -v ffmpeg. If missing, on macOS suggest brew install ffmpeg. Whisper cannot decode media without it. Stop.
• File exists and is non-empty — confirm the path resolves to a real file. If not, surface the bad path rather than guessing at a correction.

Optionally probe the media so you can set time expectations (Whisper on CPU is roughly real-time-ish on the small models and much slower on large): ffprobe -v error -show_entries format=duration -of default=nk=1:noprint_wrappers=1 "<path>". If ffprobe isn't present, skip this — it's a nicety, not a requirement.

3. Ask for Whisper settings (interactive)

Collect the settings with the AskUserQuestion tool — one call, grouped questions, with the sensible default pre-selected as the first option in each group. This is the heart of the skill: the user wants to choose how Whisper runs without memorising flags. Present these groups:

Model — accuracy vs. speed/size tradeoff. Default turbo.

If the user is sure the audio is English, mention the English-only variants (tiny.en, base.en, small.en, medium.en) are faster/better for English — offer it, don't force it. (There is no turbo.en or large.en.)

Task — transcribe (same language as the audio, default) or translate (translate the speech into English).

Language — default auto-detect (pass nothing). Offer auto-detect plus a few likely choices, and let the user type any supported language name/code (e.g. English, en, Tagalog, tl). Specifying the language is faster and more accurate than detection when the user already knows it. The CLI accepts either the full name or the ISO code.

Output formats — default all. Options map to --output_format: txt (plain text), srt (subtitles), vtt (web captions), tsv, json (full segments with timestamps), or all. This is a multi-select intent — if the user picks several distinct formats, run Whisper once per format (the CLI's -f takes one value or all), or just use all and point them at the ones they wanted. Prefer all when they pick 3+.

You may also surface, as a lighter touch, two useful extras — only ask if it fits the context, otherwise leave at defaults:

• Word-level timestamps (--word_timestamps True) — needed for word-highlighted captions; default off.
• Initial prompt (--initial_prompt "...") — seed jargon, names, or spelling to bias recognition; default none.

4. Build and run the command

Construct the command. Key flags:

• --model <model>
• --task <transcribe|translate>
• --language <name> — omit entirely for auto-detect.
• --output_format <fmt> — omit or use all for everything.
• --output_dir "<parent dir of the source file>" — this is what puts outputs next to the video. Always set it explicitly; the CLI's default is the current working directory, which is usually wrong here.
• --device — leave it unset and let Whisper pick, unless a CUDA GPU is present and you want to force it. On a Mac there is no usable CUDA device, so it runs on CPU either way (Whisper's MPS support is not wired up in the CLI). On CPU it warns FP16 is not supported, using FP32; that warning is harmless — don't treat it as an error, and don't pass --fp16 False, the fallback is automatic.

Show the user the exact command before running it (transparency + a chance to tweak), then run it. Transcription can take a while on larger models — run it in the foreground and let the user see Whisper's progress output, or background it for very long files and poll.

Example:

whisper "/Users/me/Movies/talk.mp4" \
  --model turbo \
  --task transcribe \
  --output_format all \
  --output_dir "/Users/me/Movies"

5. Report results

List the files Whisper wrote (it names them after the source stem, e.g. talk.srt, talk.txt, in the output dir). Give the user the full paths. If a .txt was produced, optionally show the first few lines so they can sanity-check the result. If Whisper exited non-zero, surface its actual stderr — common real failures are an unreadable/corrupt file or running out of memory on large.

Notes

• Outputs always land next to the source file by design (per this skill's setup). If the user wants them elsewhere, they can say so and you adjust --output_dir.
• First use of any model downloads its weights to ~/.cache/whisper (one-time, can be large for large). Mention this if a download is clearly happening on a slow link.
• Batch: the CLI accepts multiple input files. If the user points at several videos, pass them all in one invocation with a shared --output_dir, or loop per-file when each needs different settings.