transaud 0.1.0

audio diarizer and automatic speech recognition cli tool

transaud Description

A command-line tool for speaker-attributed transcription using pyannote for speaker diarization and Whisper Turbo for speech-to-text. The tool produces readable, timestamped transcripts with speaker labels.

Overview

This tool performs the following steps:

1. Converts the input audio file to a temporary WAV format.
2. Runs speaker diarization to detect speaker turns.
3. Transcribes each speaker’s audio segment using Whisper Turbo.
4. Writes a formatted transcript grouped by speaker and timestamp. GPU acceleration is automatically enabled when available, with a CPU fallback.

Usage

poetry run transcribe -iaf /path/to/audio/file.ext

Optional arguments

• --num-speakers: Specify the number of speakers if known (improves diarization).
• --line-length: Set maximum line width for wrapped transcript text.
• --output-text-file: Specify a custom output file path.

Output

The generated transcript:

• Groups consecutive speech from the same speaker
• Includes timestamps
• Uses hanging indentation for readability If no output file is specified, a .txt file is created using the input audio filename.

Dependencies

• torch
• transformers
• pyannote.audio (for diarization output)
• tqdm
• ffmpeg (system dependency)

Environment variables.

Hugging Face Authentication (Required)
This project uses pretrained models hosted on Hugging Face.

1. Generate a new huggingface token if you don't have one from hf tokens.
2. Approve access to the pyannote diarization model - link
3. Create a .env file in the project root

   HF_TOKEN=your_token_here
   

4. Ensure environment variables are loaded into your shell:

   export HF_TOKEN=your_token_here
   

Security Notice

[!Caution] This module loads pretrained models from trusted Hugging Face repositories. Ensure model sources are trusted when running in production environments.

Known Issues

Apple Silicon (macOS)

On Apple Silicon machines, you may encounter issues related to torchcodec or audio decoding. If so, update your ffmpeg library path:

export DYLD_LIBRARY_PATH="/opt/homebrew/lib:$DYLD_LIBRARY_PATH"

Notes

• Converting audio to WAV internally ensures consistent behavior across systems.
• For best results, provide --num-speakers when the number of speakers is known.
• Long recordings may require significant memory when running on CPU.

Model Licenses

This package does not ship any pretrained models. Models are downloaded at runtime from Hugging Face and are subject to their respective licenses:

• Whisper Turbo: OpenAI license
• pyannote speaker diarization: pyannote community license Users are responsible for reviewing and complying with these licenses.