audio tldr

Transcribing and summarizing audio/video content is a resource-intensive chore that often requires juggling Python scripts, FFmpeg, and LLM APIs. This skill automates the entire pipeline, turning URLs or local files into structured takeaways while maintaining a local cache to avoid redundant transcription costs.

What it does

• Automated Transcription uses Whisper backends to convert video and audio files into text locally.
• Intelligent Summarization distills long transcripts into meeting minutes, analysis reports, or key takeaways.
• Smart Name Correction cross-references transcriptions against media metadata and user glossaries to fix misheard names.
• Visual Extraction captures slides or frames from video sources to illustrate the digest.
• Translation provides full-transcript or summary translations into target languages.
• Privacy-first: audio is transcribed locally and never uploaded; an optional mode keeps the digest on your own Ollama server too.

How it works

1. Transcribe. Provide a YouTube URL, Apple Podcast link, or local file. The skill runs a cached transcription phase using the best available local Whisper backend.
2. Configure. Select a template like meeting-minutes or key-summary, or describe a custom focus for the digest.
3. Digest. The skill processes the transcript through a subagent or local Ollama, applying name-correction rules and visual frames if requested.
4. Export. The final digest is saved to a designated output folder in Markdown or HTML format.

Frameworks & tools

Built with Python, utilizing yt-dlp for media fetching and ffmpeg for processing. It supports various Whisper backends including mlx-whisper, faster-whisper, and whisper.cpp. Integrates with Ollama for local LLM workflows.

Why this beats prompting it yourself

Manual prompting requires you to handle file downloads, manage context window limits for long transcripts, and manually fix common speech-to-text errors. This skill manages the local cache, automates chunking for long content, and applies a conservative name-correction rule that fixes misheard names from the source's own details and lists every correction it makes.

Use cases

• Generate structured meeting minutes from Zoom or Google Meet recordings.
• Convert hour-long technical podcasts into five-minute skimmed summaries.
• Extract visual slides and key points from recorded webinars or tutorials.
• Create localized translations of foreign language interviews with timestamped references.

Known limitations

Requires local installation of FFmpeg, yt-dlp, and at least one Whisper backend. Transcription speed depends heavily on local hardware. Subtitles for a source already transcribed as plain text need one re-transcription.