Transcribe Video

Transcribe Video

Produce a plain-text transcript of a video or audio file. Outputs a .txt next to the source. Optional: also emit .srt/.vtt/.json if the user asks for timestamps.

Prerequisites

Before running any commands, verify both dependencies are installed:

command -v ffmpeg >/dev/null || { echo "ffmpeg not found"; exit 1; }
command -v whisper >/dev/null || { echo "whisper not found"; exit 1; }

If missing, guide the user:

• ffmpeg: brew install ffmpeg (macOS) / sudo apt install ffmpeg (Ubuntu/Debian)
• whisper: pip install openai-whisper

Inputs

• Media file path (the user will provide; otherwise ask). Video or audio.
• Optional: --model override. Default tiny.en for English, base otherwise.
• Optional: --language (auto-detect if omitted; non-English forces a non-.en model).
• Optional: --with-timestamps to also emit .srt (segment-level) alongside the .txt.

Working directory

./tmp// — mkdir at start, leave artifacts for debugging.

---

Step 1 — Extract audio

Whisper only needs mono 16 kHz audio. Hardware-decode the source to skip a slow HEVC software pass.

macOS (Apple Silicon):

ffmpeg -y -hwaccel videotoolbox -i "$SRC" \
  -vn -ac 1 -ar 16000 \
  ./tmp/$NAME/audio.wav

Linux / other platforms:

ffmpeg -y -i "$SRC" \
  -vn -ac 1 -ar 16000 \
  ./tmp/$NAME/audio.wav

To detect the platform: [[ "$(uname)" == "Darwin" ]] && HWACCEL="-hwaccel videotoolbox" || HWACCEL=""

Critical flags:

• -hwaccel videotoolbox (macOS only) — hardware-decodes HEVC/H.264, ~5–10× faster on Apple Silicon. Omit on Linux/Windows.
• -vn — skip the video stream entirely; we only need audio.
• -ac 1 -ar 16000 — mono 16 kHz is what whisper resamples to anyway; doing it in ffmpeg is faster than letting whisper do it.

If the source is already audio (.wav, .mp3, .m4a, .flac), still run this — it normalizes the format and gives whisper a clean input.

Step 2 — Transcribe with whisper

whisper ./tmp/$NAME/audio.wav \
  --model tiny.en \
  --output_format txt \
  --output_dir ./tmp/$NAME \
  --language en

Model picker:

• tiny.en — default for English. Fastest, quality fine for clear speech (tutorials, podcasts, voice-overs).
• base.en — bump up if tiny.en produces noticeable errors (proper nouns, technical terms).
• small.en / medium.en — only if the user explicitly asks for higher accuracy and accepts longer wait.
• turbo — best quality-to-speed ratio for any language. Use when tiny.en quality isn't good enough but the user hasn't asked for a specific model.
• base (no .en) — non-English or mixed-language. Drop --language en.

Output format picker:

• Plain .txt is the default — just the spoken words, one paragraph per whisper segment, no timestamps. This is what most users want.
• If the user asked for timestamps, use --output_format all to emit every format (.txt, .srt, .vtt, .json, .tsv). The user gets their .txt plus whichever timestamp format they need.
• If the user asked for a specific format only (e.g. just .srt), use --output_format srt.

Whisper writes audio.txt (and audio.srt/audio.vtt/audio.json if requested) into --output_dir.

Step 3 — Deliver

• Copy audio.txt to /.txt (and .srt/.vtt if produced).
• Print one line: source path → output path, word count, runtime.
• Don't auto-open the file — it's plain text, the user can cat or open it themselves. Just report the path.

---

Pitfalls — don't repeat these

• Don't use -hwaccel videotoolbox on Linux. It's macOS-only. Check uname first.
• Don't skip -hwaccel videotoolbox on macOS. Software HEVC decode on a 4K source takes minutes; with the flag, seconds.
• Don't feed whisper the raw video file. It works, but ffmpeg's resampling step is much faster than whisper's internal one — and you'd lose -hwaccel. Always extract audio.wav first.
• Don't default to small.en or larger. tiny.en is fine for English voice-overs, tutorials, podcasts. Only escalate if the user reports errors or asks for better quality.
• Don't pass --word_timestamps True unless the user asked for word-level timing. It roughly doubles whisper's runtime.
• Don't write the txt to ./tmp and stop. Always deliver alongside the source file so the user can find it.
• Don't strip whisper's paragraph breaks. The default .txt output already segments sensibly — leave it alone unless the user asks for a single blob.

Keep this skill focused on one thing: produce a plain-text transcript from a media file, fast.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: smitmartijn
• Source: smitmartijn/claude-skills
• License: MIT

Install and usage instructions live in the source repository linked above.