What problem does it solve? Turning recorded talks of Shri A. Nagraj into citable reference artefacts requires far more than running ASR: VAD drops doctrinally important words, whisper.cpp loops on repeated phrases, emits broken UTF-8, and injects YouTube caption boilerplate. This Skill encodes the measured configuration and the full review-and-promotion workflow so transcripts are auditable and citable. ## Core Features & Use Cases - Fetch and batch-transcribe: Build a TSV manifest, download audio with yt-dlp, and decode with faster-whisper or whisper.cpp (ROCm/HIP or Vulkan GPU backends) using Scripts/_transcribe_fetch.py and Scripts/_transcribe_batch.py with VAD off and --max-context 0. - Batch quality review: Run Scripts/_transcribe_review.py to flag anomalies via words-per-minute spread, repeated tokens, 3-gram frequency, Devanagari share, and U+FFFD counts before promoting anything. - Promotion to citable artefacts: Repair broken UTF-8, segment from native JSON timestamps, translate against a controlled terminology registry, mark every segment [R]/[P]/[U], and generate PDFs. - Use Case: A researcher adds a new recorded session to References/Madhyasth-Darshan/Nagraj-Recorded-Sessions/ by fetching the audio, transcribing on GPU, reviewing the batch, and promoting a verified transcript with reliability markings. ## Quick Start Ask the assistant to transcribe a recorded talk by providing a YouTube video ID or local audio file and requesting a reviewed Hindi transcript with English translation.
