Video Transcription

Transcribe video and audio files in-place to SRT subtitle format with speaker diarization. Backed by the pi-transcribe CLI (a TypeScript port of the original standalone skill — no Python).

Usage

Run the pi-transcribe bin, optionally passing a directory or file paths. Output .mp3 (extracted audio) and .srt (subtitles) files are placed alongside the source files.

pi-transcribe [directory | file ...]

• No argument: scans ~/Movies (default)
• Single directory: scans the specified directory (e.g. a Google Recorder .m4a export folder)
• One or more file paths: transcribes exactly those files

Examples:

• /transcribe — transcribe all untranscribed files in ~/Movies
• /transcribe /path/to/recordings — transcribe files in a specific directory
• pi-transcribe "~/Movies/May 28 at 4-04 PM.m4a" "~/Movies/Feb 2 at 5-05 PM.m4a" — transcribe specific files
• MAX_CHUNK_HOURS=4 pi-transcribe ~/Movies — change the long-recording chunk size (default 4.5h)
• TRANSCRIBE_CONCURRENCY=4 pi-transcribe ~/Movies — change how many files transcribe in parallel (default 8)

Parallel processing

Files transcribe through a bounded worker pool: up to TRANSCRIBE_CONCURRENCY files (default 8) are in flight at once, overlapping the Soniox wait that dominates each file's wall-clock time. Files are dispatched oldest-first but may complete in any order. Set TRANSCRIBE_CONCURRENCY=1 for serial, deterministic behavior. The value is clamped to 1–100 (100 = the Soniox pending-job cap).

Execution

1. Run pi-transcribe, passing the user's directory/file arguments (if any)
2. The bin handles everything: file discovery, audio extraction, API transcription, idempotency (skips files that already have a sibling .srt)
3. Present the bin's summary output to the user (total found, already transcribed, newly transcribed, failed)
4. If there are failures, report which files failed and the error messages

Long recordings (>5h)

Soniox enforces a HARD per-request limit on audio duration (18000 s / 5 h), independent of file size. Recordings longer than this are automatically split into MAX_CHUNK_HOURS-sized chunks (default 4.5 h), transcribed separately, and merged into a single SRT with correct absolute timestamps — full coverage, no truncation. Override the chunk size with the MAX_CHUNK_HOURS env var.

Note: the limit is on duration, not megabytes — a long low-bitrate recording can be small in size yet still exceed 5 h, so the guard probes duration via ffprobe.

Prerequisites

• ffmpeg (with ffprobe) — declared in this package's pi.tools manifest and resolved through the dashboard tool registry (PATH or the static npm packages: ffmpeg-static for ffmpeg, @ffprobe-installer/ffprobe for ffprobe — both optionalDependencies here). When absent, video files are skipped with a warning; audio-only files still process. pi-dashboard-ensure <package-root>/package.json reports both.
• SONIOX_API_KEY — declared as an env probe in pi.tools; resolved from the environment first, then an optional gitignored .env (current directory, then the skill dir). No secret ships in the package; the bin fails fast with a clear message if the key is unresolved.

Environment overrides

Variable	Default	Meaning
SONIOX_API_KEY	(required)	Soniox API key.
MAX_CHUNK_HOURS	4.5	Chunk size for recordings over the 5 h duration limit.
MAX_AUDIO_MB	200	Reserved size guard; 0 disables.
TRANSCRIBE_CONCURRENCY	8	Files transcribed in parallel. Clamped to 1–100; 1 = serial.