Universal Speech-to-Text Transcriber and Tranlsator

A Python-based Proof of Concept for transcribing any audio files using the Whisper model locally and translate it to any language. This tool supports both long-form audio processing and sample creation, with options for chunked processing of larger files.

Features

• Convert M4A files to MP3 format
• Create sample clips from longer audio files
• Local speech-to-text processing using Whisper
• Local trnaslation of transcribed text to chosen language
• Support for both timestamped and clean text output
• Chunked processing for handling large audio files
• Completely offline operation - no API keys needed

Prerequisites

• Python 3.10+
• FFmpeg
• Sufficient disk space for Whisper model

Configuration

Adjust these parameters in the script for different use cases:

• model_size: Choose from "tiny", "base", "small", "medium", "large-v2"
• chunk_duration: Duration in minutes for splitting large files
• use_chunks: Boolean to enable/disable chunked processing

Usage

To run the script, use the following command:

python transcribe.py <input_file> <source_lang> <target_lang>

• <input_file>: Path to the input audio file (MP3 format)
• <source_lang>: Source language code (e.g., "de" for German)
• <target_lang>: Target language code (e.g., "hu" for Hungarian)

Example:

python transcribe.py test.mp3 de hu

hu This will transcribe the test.mp3 file from German (de) to Hungarian (hu).

Performance Notes

• Recommended to use "medium" model for balance of speed and accuracy
• M1 Mac users can utilize MPS acceleration
• Large files can be processed either directly or in chunks based on available memory