ATT-Whisper Subtitle Generator

Generate SRT (and VTT) subtitles from audio or video files using OpenAI Whisper.

Features

• Single file or whole folder processing (recursive media discovery)
• Choose Whisper model size (tiny .. large-v3)
• Spanish by default (override with -l <lang>)
• Transcribe or translate (--task translate → English)
• Export .srt, .vtt, or both
• GPU auto-detection (CUDA → FP16 for speed) or force CPU
• Custom model weights directory via --model-dir
• Generate short, TikTok/Shorts-style subtitle blocks using --max-words and --max-chars

Requirements

• Windows (tested) / Linux / macOS
• Python 3.9+ (you are using 3.12)
• FFmpeg installed and on PATH
• Python packages: openai-whisper, torch (+ CUDA build if using GPU)

Install packages:

py -3.12 -m pip install -U pip openai-whisper

If you need CUDA-enabled Torch for a new setup, follow instructions at: https://pytorch.org/

Basic Usage

Single file:

py -3.12 generate_srt.py "input.mp3" -m small -l es

Folder (creates subs/ inside folder):

py -3.12 generate_srt.py "D:\Media\Interviews" -m medium -l es -f both

Force CPU:

py -3.12 generate_srt.py "clip.mp4" -m base --device cpu

Translate to English:

py -3.12 generate_srt.py "entrevista.wav" -m medium --task translate -l es

Chunked Captions

To generate very short subtitle blocks (e.g., 3–5 words) that look better on vertical videos:

# Maximum 4 words per block
py -3.12 generate_srt.py "video.mp4" -m small -l es --max-words 4

# Maximum 5 words or 28 characters (whichever comes first)
py -3.12 generate_srt.py "video.mp4" -m medium -l es --max-words 5 --max-chars 28

# Character Limit Only (No Word Limit)
py -3.12 generate_srt.py "video.mp4" -m base -l es --max-chars 32

Additional files produced (besides the standard ones):

• name_chunked.srt
• name_chunked.vtt (if you use -f vtt or -f both)

How chunking works:

1. Enables word_timestamps to get per-word timing (if supported by the backend).
2. Accumulates words until --max-words or --max-chars would be exceeded.
3. Closes the block and starts a new one.
4. If the model does not return individual words, it approximates the split within the segment.

Tips:

• 3–5 words per block are usually readable for fast-paced content.
• Combine word and character limits to avoid overly long lines.
• Review the output and adjust values to fit your video style.

Example (Your Workflow)

Exact command you ran:

py -3.12 generate_srt.py "C:\PATH" -m medium -l es -f srt --device auto --model-dir ".\models" --max-words 4 --max-chars 10

What happens step-by-step:

1. Script resolves device: selects cuda if available else cpu.
2. Creates/uses custom model cache folder: .\models (so weights like medium.pt live in the project).
3. Loads the Whisper medium model (downloads once if missing).
4. Processes the WAV file (or all supported media if a folder given).
5. Generates segments and writes name.srt next to the input (or subs/ for folder mode).
6. You can import the SRT into Adobe Premiere Pro.

Supported Media Extensions

Audio: mp3, wav, m4a, aac, flac, ogg, wma Video: mp4, mov, mkv, m4v, webm, avi

Output Location Rules

• Single file input → subtitle saved alongside the media.
• Folder input → a subs/ directory is created inside that folder unless -o is provided.
• Custom output directory: use -o path/to/output.

Custom Model Directory

Use --model-dir to store/download model weights locally (helpful for portability or offline use):

py -3.12 generate_srt.py input.wav -m medium --model-dir .\models

Pre-seed by moving an existing weight file:

Move-Item $env:USERPROFILE\.cache\whisper\medium.pt .\models\medium.pt

Choosing a Model

Model	Size	Notes
tiny	~75 MB	Fastest, lowest accuracy
base	~142 MB	Lightweight
small	~466 MB	Good balance
medium	~1.5 GB	Higher accuracy (current choice)
large-v2	~3.1 GB	Older large option
large-v3	~3.2 GB	Latest high accuracy

If VRAM is limited or you only need a quick draft, try small.

Language and Translation

• -l es sets source language (Speeds up + improves accuracy vs auto-detect).
• --task translate outputs English text from the source language.

Performance Tips

• Use GPU (cuda) for large models. Automatically chosen when available.
• FP16 is enabled on CUDA to reduce memory and improve speed.
• For long batches of files, consider smaller models to reduce load times.

Troubleshooting

Issue	Cause	Fix
FFmpeg not found	Not installed / not on PATH	Install from https://www.gyan.dev/ffmpeg/builds/ and add ffmpeg\bin to PATH
CUDA not used	Driver / toolkit mismatch	Update NVIDIA driver; ensure you installed correct Torch CUDA build
Out of VRAM	Model too large	Try small or medium instead of large-*
Slow CPU run	Using big model on CPU	Switch to smaller model
Wrong language output	Language mis-specified	Adjust -l or remove to auto-detect

Script Arguments Summary

positional:
  input                 File or folder

options:
  -o, --output_dir      Output directory
  -m, --model           Model size (default: large-v3)
  -l, --language        Language code (default: es)
      --task            transcribe | translate (default: transcribe)
  -f, --format          srt | vtt | both (default: srt)
      --device          auto | cuda | cpu (default: auto)
      --temperature     Sampling temperature (default: 0.0)
      --beam_size       Beam size (default: 5)
      --verbose         Print segments to console
      --model-dir       Custom model download/cache directory
  --max-words       Máximo de palabras por bloque (genera *_chunked.srt/.vtt)
  --max-chars       Máximo de caracteres por bloque (aplicado después del límite de palabras)

Roadmap Ideas

• Add progress bar per file (currently only the model download + decode progress are shown)
• Optional JSON export of segments
• Auto language detection option

License / Usage

Use responsibly. Whisper models are released under the MIT license (see upstream repo). Check media rights before transcribing.