µSubs.py

Whisper Powered .SRT Subtitle File Generator

A quick and dirty, vibe-coded .srt maker with a few nifty features grown over time.
Buffers the audio/video file to a local .wav file at Whisper-native bitrate.
Samples 10 different spots in the audio to smartly detect the correct language.
You can override that detection, of course, and bilingual transcript/subs are supported.
To quell dreamt up transcripts, the audio is pre-screened for vocal sections with Silero.

The built-in readme lays it all out, and follows here:

µSubs Adaptive Silicon Edition

COMMAND LINE ARGUMENTS: input_files Paths to video/audio files for batch processing.
--cores [n/max] Explicitly set CPU thread count (e.g., --cores 4).
--language [code] Force source language (e.g. --language ja).
--force-model [m] Override automatic selection.
--max Alias for --cores max. Run at maximum threads.
--transcribe Original language only ([file].srt).
--bilingual English Top / Original Bottom ([file].srt).
--quiet Suppress per-line terminal output.

WHISPER MODEL SELECTION TABLE:

MODEL	VRAM USAGE	SPEED	ACCURACY NOTES
tiny / .en	~1.0 GB	32x	Lowest accuracy, lightning fast.
base / .en	~1.0 GB	16x	Good for English, weak for others.
small / .en	~2.0 GB	6x	Great balance for clear audio.
medium / .en	~5.0 GB	2x	High fidelity, slower on CPU.
turbo	~6.0 GB	8x	Large-v3 fidelity, highly optimized.
large-v3	~10.0 GB	1x	Maximum accuracy, expensive on CPU.

TOP 20 LANGUAGE REFERENCE (ISO-639-1):

CODE	LANGUAGE	CODE	LANGUAGE	CODE	LANGUAGE	CODE	LANGUAGE
en	English	zh	Chinese	hi	Hindi	es	Spanish
fr	French	ar	Arabic	bn	Bengali	pt	Portuguese
ru	Russian	ur	Urdu	id	Indonesian	de	German
ja	Japanese	pa	Punjabi	mr	Marathi	te	Telugu
tr	Turkish	ta	Tamil	vi	Vietnamese	ko	Korean

LINGUISTIC ENGINE STANDARDS:

• Audio Logic: FFmpeg Reverted Pipeline (16kHz Mono).
• Silero VAD: Min Silence 500ms | Threshold 0.25 (Aggressive splitting).
• Logic Profile: All Cores - 2 (Default).
• Power State: Auto-engages caffeinate -i to prevent idle sleep.

---