GPU-Optimized Bulk Video Transcription Tool

This is a high-performance Python application for transcribing video files in bulk. It leverages faster-whisper for GPU-optimized transcription and supports advanced features like speaker diarization and gender classification.

Features

• High-Speed Transcription: Utilizes faster-whisper and CTranslate2 for optimal performance on NVIDIA GPUs (with float16/int8 precision).
• Bulk Processing: Transcribe a single video file or an entire directory of videos.
• Speaker Diarization: Automatically identify and label different speakers in the audio.
• Gender Classification: Classify speakers as MALE or FEMALE.
• Multiple Output Formats: Generate transcripts in SRT, VTT, TXT, or JSON formats.
• Intelligent Subtitle Splitting: Automatically breaks long transcription segments into shorter, more readable subtitles.
• Subtitle Embedding: Option to embed the generated subtitles as a soft-track into the original video file.
• Flexible File Handling: Choose how to handle existing transcript files (overwrite, skip, or rename).

Prerequisites

Before you begin, ensure you have the following:

1. FFmpeg: The script requires ffmpeg to extract audio from video files. Ensure it is installed on your system and accessible in your system's PATH.

   • You can download it from https://ffmpeg.org/download.html.

2. Hugging Face Account (for Speaker Diarization): The speaker diarization feature uses a gated model from Hugging Face. To use it, you must:

   • A. Have a Hugging Face Account: If you don't have one, create one at https://huggingface.co/join.
   • B. Accept Model Terms: Visit the page for the diarization model and accept the terms of service. You must be logged in.
     • Model URL: https://huggingface.co/pyannote/speaker-diarization-3.1
   • C. Get an Access Token: Create an access token to authenticate with the Hugging Face Hub.
     • Token URL: https://huggingface.co/settings/tokens

You will need to provide this token to the script using the --hf-token option.

Installation and Usage

There are two primary ways to use this application.

Method 1: As a Command-Line Tool (Recommended)

This method installs the application as a system-wide command, allowing you to run it from any directory.

1. Installation

Install the application directly from GitHub using pip:

pip install git+https://github.com/VyanP/gpu_transcriber.git#egg=gpu-video-transcriber

2. Usage

Once installed, run the application using the gpu-transcriber command:

gpu-transcriber <INPUT_PATH> [OPTIONS]

Examples:

# Transcribe a single file
gpu-transcriber "C:\path\to\my\video.mp4"

# Transcribe a directory with diarization
gpu-transcriber "/path/to/videos/" --enable-diarization --hf-token YOUR_HF_TOKEN

---

Method 2: Directly from Source

This method is for developers or those who want to run the script without installing it as a package.

1. Setup

First, get the source code and install the required dependencies.

# 1. Clone the repository
git clone https://github.com/VyanP/gpu_transcriber.git

# 2. Navigate into the project directory
cd gpu_transcriber

# 3. Install dependencies
# The "-e" flag makes it "editable", linking to your source files
pip install -e .

2. Usage

Run the application as a Python module from within the project's root directory:

python -m gpu_transcriber_app.main <INPUT_PATH> [OPTIONS]

Examples:

# Transcribe a single file from source
python -m gpu_transcriber_app.main "C:\path\to\my\video.mp4"

# Transcribe a directory with diarization from source
python -m gpu_transcriber_app.main "/path/to/videos/" --enable-diarization --hf-token YOUR_HF_TOKEN

Command-Line Options

Option	Alias	Description
--model	-m	Whisper model to use. Larger models are more accurate but slower. (Default: large-v3)
--language	-l	Force a specific language for transcription (e.g., "en", "es").
--output-format	-f	The output format for the transcript file. (Default: srt)
--embed / --no-embed		Embed subtitles into video files as a soft subtitle stream.
--device	-d	Processing device (auto, cuda, cpu). (Default: auto)
--batch-size	-b	Batch size for GPU processing. (Default: 16)
--compute-type		Compute precision (float16, float32, int8). (Default: float16)
--disable-vad / --enable-vad		Disable/Enable Voice Activity Detection. Disabling can be useful for music or mixed content.
--enable-diarization		Enable speaker diarization to identify different speakers. Requires Hugging Face setup.
--enable-gender-classification		Enable gender classification. Works with or without diarization.
--max-subtitle-duration		Maximum duration for a single subtitle segment in seconds. (Default: 5.0)
--max-subtitle-chars		Maximum characters per subtitle line. (Default: 80)
--file-conflict		How to handle existing output files (ask, overwrite, rename, skip). (Default: ask)
--hf-token		Your HuggingFace access token. Required for speaker diarization. See Prerequisites.
--verbose	-v	Enable verbose logging for debugging purposes.