Zekiog/media-pipeline

Media Pipeline

CPU transcription and OCR for files on this machine. FFmpeg, faster-whisper (int8), RapidOCR. Whisper model defaults to tiny.

Layout

Paths are relative to the media-pipeline checkout.

bin/mp              probe, process, translate, list, watch, selftest, doctor
bin/mp-mcp          stdio MCP server
inbox/              files for `mp watch`
work/               16 kHz wav and work/processed.json
outbox/<epoch>_<stem>/
                    normalized.mp4, thumbnail.jpg, transcript.txt,
                    transcript.srt, transcript.vtt, ocr.txt,
                    translation.<lang>.txt, manifest.json
failures/           inputs `mp watch` could not process

install.sh copies this skill to $MP_SKILLS_DIR/media-pipeline when that directory already exists. The default is $HOME/.local/share/prime-agent/0.9.6/skills.

Engines

Job	Engine	Notes
probe	ffprobe	format, duration, streams
video	FFmpeg	H.264 + AAC, yuv420p, faststart, CRF 28; thumbnail at 1s, width 1280
audio for the model	FFmpeg	16 kHz mono PCM, kept under work/
speech to text	faster-whisper	CPU, MP_WHISPER_COMPUTE default int8, model MP_WHISPER_MODEL default tiny
image text	RapidOCR	PP-OCRv4 through ONNX Runtime, written to ocr.txt
translate	Pollinations, Gemini, or LiteLLM	only when requested; see references/engines.md

Commands

bin/mp selftest
bin/mp probe FILE
bin/mp process FILE
bin/mp process FILE --translate --target tr
bin/mp translate FILE --target tr
bin/mp watch
bin/mp doctor

process leaves the input file where it is. watch moves it to outbox/ or failures/. A file whose sha256 is already in work/processed.json stays in inbox/.

doctor exits 1 when a critical check fails. One of those checks is $HOME/bin/prime-tg, which this repository does not contain.

MCP tools

media_engines, media_probe, media_process, media_translate, media_batch.

media_process accepts path, transcribe, ocr, translate, and target_language. media_batch processes inbox/ with transcription and OCR on and translation off.

Rules

• Leave MP_LITELLM_KEY and MP_GEMINI_KEY in the environment. Do not print them.
• tiny is the default. Set MP_WHISPER_MODEL to base or small when the recording needs it. medium and large-v3 are large downloads.
• An empty transcript with probe.duration_s of 0 means the audio was not decoded.
• If translation returns an error, the text is not translated. Say so.
• Repository text stays in English.

Reference

• references/engines.md for provider order, OCR limits, and the ARM host notes
• references/triggers.md for the inbox path unit and mp watch
• references/recipes.md for the venv, ffmpeg, and OCR commands