audio-transcriber_skill

Overview

This skill extracts audio from dashcam MP4 files and produces GPU-accelerated, timestamped transcripts with optional speaker diarization. It organizes outputs by date, generates multi-format artifacts (TXT/JSON/SRT/VTT), and produces an INDEX.csv plus a results summary with GPU and quality metrics. Designed for reliable batch processing of date-organized video folders and single files.

How this skill works

The skill auto-discovers or accepts a user-specified folder of MP4s, validates files and environment (FFmpeg, GPU), then extracts audio via FFmpeg with a retry matrix. Audio is segmented (fixed chunks or VAD-based), transcribed on GPU using faster-whisper with word timestamps, and optionally passed to a diarization backend (pyannote or speechbrain). Outputs and metadata are written per-day with error handling, resume safety, and a global INDEX.csv and results JSON containing GPU metrics and processing statistics.

When to use it

• You need transcripts from dashcam MP4 footage or other video files.
• You want timestamped transcripts with optional speaker labels (diarization).
• You need batch processing for a folder or date range of videos.
• You want GPU-accelerated transcription for faster throughput.
• You need multi-format outputs (TXT/JSON/SRT/VTT) and a searchable INDEX.csv.

Best practices

• Always confirm the configuration summary before starting to avoid accidental processing.
• Provide an explicit folder path or date when possible to speed auto-discovery.
• Enable diarization only if a valid HF token and sufficient VRAM are available; otherwise use speechbrain or skip.
• Use fixed 30s segmentation for continuous speech; use VAD mode for sparse audio like parking recordings.
• Keep FFmpeg and CUDA drivers up to date and ensure adequate disk space for audio extracts and outputs.

Example use cases

• Forensic review: transcribe a day of dashcam videos and search INDEX.csv for keywords and speaker turns.
• Incident analysis: extract audio, generate timestamped transcripts and SRT subtitles for sharing with investigators.
• Bulk processing: run on a date-organized folder to produce per-video JSON with word-level timestamps and GPU metrics.
• Quality auditing: re-run with --force to regenerate transcripts after model or parameter changes while preserving FAILED logs.
• Subtitling: create SRT/VTT with normalized speaker labels for video publishing or evidence presentation.