| name | SAI-local-transcribe |
| description | Atomic SAI skill. Fully-local audio transcription via Whisper (mlx-whisper on Apple Silicon, faster-whisper fallback elsewhere). No audio leaves the machine; no API keys. Accepts a single audio file, a directory, or a podcast-download manifest.json. Writes one .md per input with YAML frontmatter compatible with a rag-index-updater pipeline. Use whenever you need to transcribe audio, run whisper on a file, transcribe all episodes in a folder, or any workflow needs text from audio. |
Trigger manifest — local-transcribe (atomic, autonomous)
Code: skills/local-transcribe/
Manifest: skill.yaml (4-tier cascade)
Entry point: runner.py → run(inputs)
Cascade plan
1. pick_backend (rules) → mlx-whisper (preferred) or faster-whisper
2. collect_inputs (rules) → manifest_path | audio_path | audio_dir
3. transcribe_each (rules) → Whisper → .md with YAML frontmatter
4. summarize (rules) → counts; ready_to_propose
All tiers deterministic. No LLM calls. No network (after the model is cached).
Invocation — chained off podcast-download
.venv/bin/python3.12 \
skills/local-transcribe/runner.py run \
--inputs-json '{
"manifest_path": "~/Media/podcasts/<show-slug>/manifest.json",
"output_dir": "~/Media/transcripts/<show-slug>",
"source_label": "Podcast"
}'
Invocation — single file
.venv/bin/python3.12 \
skills/local-transcribe/runner.py run \
--inputs-json '{
"audio_path": "~/Downloads/interview.mp3",
"output_dir": "/tmp/transcripts"
}'
Invoke runners by absolute path (folder names use hyphens, so python -m won't import).
Frontmatter shape (rag-ready)
title: "Episode Title"
date: "YYYY-MM-DD"
url: "https://..."
source: "Podcast"
show_title: "..."
episode_id: "..."
language: "en"
transcribed_by: "mlx-whisper"
transcribed_at: "..."
audio_source: "/abs/path"
Composes with
- Upstream:
SAI-podcast-download
- Downstream:
SAI-podcast-rag-index
