EdgeSpeak Transcribe

Turn audio/video into a transcript, entirely on-device — the audio never leaves the machine. Under the hood it calls edgespeak-cli transcribe. When the EdgeSpeak desktop app is running, the CLI talks to its local gateway (OpenAI-compatible, 127.0.0.1:1117) and reuses the warm model (proxy mode); when the app is not running, the CLI launches the bundled on-device engine itself (standalone mode). Standalone is a normal mode, not an error.

Version compatibility. The frontmatter pins this skill's version and the oldest CLI it is written against (minCliVersion). If edgespeak-cli --version reports something older, run edgespeak-cli update (or re-run the installer) before relying on the flags documented here. Same-numbered builds can still differ, so --help is the tiebreaker: a command or flag documented here but missing from the installed --help also means update — don't route around it.

Inputs to confirm

• Media path to transcribe.
• Desired output: stdout text, .txt, .json, or .srt.
• Any requested model, word timing, sentence length, or subtitle padding options.
• Whether the user needs anonymous speaker labels (who said what) or actual names. Both require diarization at transcription time (--diarize) — adding it afterwards costs a full re-transcription. If real names are needed, also ask for a participant roster or original source/video/YouTube URL, then continue with edgespeak-name-speakers after this skill produces JSON.

If the user wants subtitles or captions, ask about cue shaping before the first run. Cue length is not guessable from the request, and discovering it afterwards costs a whole re-transcription. Ask once, in a single message:

• Max characters per cue? (--max-chars; roughly 60-90 reads well, and the default leaves whole sentences intact.)
• Any minimum length or leading/trailing padding? (--min-chars, --start-margin, --end-margin.)

Then run transcribe with the answers applied. Do not produce a plain -o out.srt first and re-run with shaping after.