Transcription Skill

Transcription skill

Transcribe a YouTube URL, a local video file, or a local audio file to plain text using Google's gemini-3.5-flash on Replicate.

export REPLICATE_API_TOKEN=...
python3 <skill-directory>/scripts/transcribe.py <youtube-url | video-file | audio-file> [output-file] [--from TIME] [--to TIME]

Replace <skill-directory> with the directory containing this SKILL.md.

If the script exits saying yt-dlp or ffmpeg is missing, offer to install it for the user (e.g. brew install yt-dlp or brew install ffmpeg on macOS) before retrying, rather than just reporting the error.

Transcribing part of a video

When the user names a time range, pass it through with --from and --to rather than transcribing the whole thing. Times can be SS, MM:SS, or HH:MM:SS, and either flag can be used alone.

python3 <skill-directory>/scripts/transcribe.py "https://youtu.be/abc123" --from 15:08 --to 16:22

For a YouTube URL this passes --download-sections to yt-dlp, so only the requested range crosses the network. On a 108-minute podcast, pulling a 74-second clip moves about 23 MB instead of the full download, and the prediction sees 74 seconds of audio instead of an hour and a half.

Clipped files carry the range in their names (e.g. <slug>-15m08s-16m22s.m4a), so clips never overwrite a full-length transcript of the same video. Trimming a local file needs ffmpeg even when the input is already audio.

A range that starts past the end of the media is rejected before anything is downloaded or cut. A range that merely ends past it warns and transcribes up to the end.

What it does, in order

1. If the input is a YouTube URL, downloads it with yt-dlp into the current directory under a slugified filename (lowercased title + video id, e.g. my-video-title-abc123.mp4), fetching only the --from/--to range if one was given.
2. If the input is a video file (or was just downloaded), extracts its audio with ffmpeg via stream copy (no re-encoding) and saves it alongside the video as <slug>.m4a. A local audio input with a time range is trimmed the same way.
3. Base64-encodes the audio and sends it to google/gemini-3.5-flash on Replicate with a verbatim-transcription prompt.
4. Polls until the prediction completes and writes the transcript to <slug>.txt (or the given output path).

An audio file input skips straight to step 3. google/gemini-3.5-flash is hardcoded as the transcription model; see the root README.md for how that choice was benchmarked against other models.

Known gotchas

• ffmpeg seeking past the end of a file doesn't fail. With -ss beyond the media's duration and -acodec copy, ffmpeg exits 0 and writes the tail of the stream with negative timestamps (time=-02:59:00.00) instead of an empty file. Gemini then hallucinates a plausible sentence over that garbage. This is why the requested range is checked against the source duration up front rather than by inspecting the resulting clip.

• YouTube 403s. yt-dlp's extractor breaks against YouTube frequently. If downloads fail with HTTP Error 403: Forbidden, run brew upgrade yt-dlp (or update however it was installed) and retry.

• Gemini can't fetch Replicate's own file URLs. Uploading audio to Replicate's /v1/files and passing that URL as the audio input fails because Gemini's backend can't authenticate the fetch, and gets back a 401 JSON response it reports as "could not determine mimetype". Send the audio as a base64 data: URI directly in the prediction input instead.

• .m4a mimetype sniffing. Python's mimetypes.guess_type reports .m4a as audio/mp4a-latm, which Gemini rejects. Use audio/mp4.

• api.replicate.com blocks requests with no User-Agent header (Cloudflare error 1010). Always set one.

• Pin to the model's current latest_version. Don't hardcode a version id — Google ships new Gemini versions on Replicate often enough that hardcoded ids go stale.