URL/Local file → Audio → Whisper Transcript

Download/clip audio from a URL or local file and generate a timestamped transcript with faster-whisper. Minimal, fast, and scriptable.

• URL or local input: YouTube links or files on disk.
• Clip range: --start/--end in HH:MM:SS (or use defaults in the script).
• Job-based runs: each source → its own run_YYYYMMDD_HHMMSS_xxx/.
• GPU optional: works on CPU; CUDA speeds things up if available.
• Straightforward outputs: audio.mp3 and transcript.txt.

Pipeline (diagram)

flowchart LR
  A["URL or local file"] --> B{"Range known?"}
  B -- "Yes" --> C["yt-dlp/ffmpeg → MP3 clip"]
  B -- "No" --> D["Probe duration (yt-dlp/ffprobe)"] --> C
  C --> E["faster-whisper (batch pipeline)"]
  E --> F["timestamped transcript.txt"]

Loading

Requirements

• Python ≥ 3.11
• ffmpeg on PATH
• pip install -r requirements.txt
• (Optional) PyTorch with CUDA for GPU
• (Recommended) Keep yt-dlp up to date

Quickstart

One or more sources:

python clip2txt.py "https://youtu.be/CPBJgpK0Ulc" "C:\path\to\local\audio.mp3"

Clip range, output dir, model:

python clip2txt.py "SOURCE_URL_OR_PATH" --start 00:01:00 --end 00:03:30 -o output -m medium --compute-type int8_float32

CLI

python clip2txt.py [SOURCES ...]
                   [--start HH:MM:SS]
                   [--end HH:MM:SS]
                   [-o OUTDIR]
                   [-m MODEL]
                   [--compute-type TYPE]

Option	Description	Default (from script)
SOURCES ...	URLs or file paths (zero or more)	SOURCES
--start HH:MM:SS	Clip start	START_TS
--end HH:MM:SS	Clip end	END_TS
-o, --outdir PATH	Output folder	OUTDIR
-m, --model NAME	Whisper model (tiny … large-v2)	MODEL
--compute-type T	Compute type (int8, int8_float32, …)	COMP_TYPE

Outputs per source end up in output/run_YYYYMMDD_HHMMSS_xxx/ with audio.mp3 and transcript.txt.