Audio and Video Workbench
Select the local media job
Inspect a user-supplied local media file, then choose audio extraction, clip cuts, frames, subtitle conversion, transcription or a combined package. Read references/input.md. Files and outputs stay in the active workspace. Use the user's requested time ranges; clarify missing ranges before cutting.
Python 3.10+ is required. Install scripts/requirements.txt in a dedicated virtual environment. Use an installed FFmpeg, or the official imageio-ffmpeg package's local executable. PyAV inspects and decodes the files. The script accepts local supported formats and disables network protocols/playlist demuxers; it does not fetch video URLs.
python3 "{baseDir}/scripts/run.py" --workspace "CURRENT_WORKSPACE" --input "request.json"
Clips are transcoded for precise requested boundaries, then their actual durations are probed. Frames become JPEGs plus a contact sheet. User-provided timestamp segments convert to UTF-8 SRT/VTT, JSON and a time-stamped transcript. For a combined package, overlapping captions are trimmed and offset relative to each clip. This exports sidecar subtitles, not burned-in captions. package also returns a ZIP of the outputs.
Optional local speech recognition
For transcribe only, install scripts/requirements-asr.txt. Choose tiny first on a CPU-only server; base or small use more resources. The engine is faster-whisper on CPU with int8 and 1..8 explicit threads. It runs locally, without a paid transcription API or MochiLab server. First model retrieval needs allowModelDownload:true; downloaded models are cached under this workspace. Prefer modelPath for a pre-downloaded, converted Whisper directory. The helper never downloads model weights when that option is false.
Transcription is segmented into 300-second chunks. Completed chunks are retained in the current workspace and keyed by source, model, decoder, language and engine version. A fresh requestId can reuse completed chunks after interruption; an identical completed requestId reuses verified outputs. No partial export is reported as a completed transcript. A local modelPath is fingerprinted from its model/config/tokenizer/vocabulary files. For a named downloaded model, the model name and engine version identify its cache; supply modelPath when exact model reproducibility matters.
Review and delivery
Open the actual clip, preview.jpg and subtitle files. Return manifest.json with source hashes, selected timestamps and measured output durations. Recognition output is labeled local-asr and needs transcript review; user-provided captions are labeled separately. A model-written summary is a later task and must not be inserted as if recognized speech. Preserve source files and do not send media to external services unless separately requested.