Speaker Diarization and Audio Splitting

This project provides an automated pipeline for speaker diarization and audio splitting using a pre-trained model from Hugging Face's pyannote.audio library. The system processes an input audio file, identifies speakers, and saves their speech segments as separate .wav files.

Features

• Speaker Diarization: Automatically detects and labels different speakers in an audio file.
• Audio Splitting: Extracts and saves audio segments corresponding to each speaker.
• Pre-Trained Model: Utilizes Hugging Face's pyannote/speaker-diarization for state-of-the-art diarization.
• GPU Acceleration: Supports CUDA for faster processing when a GPU is available.

Requirements

• Python 3.8+
• Libraries:
  • torch
  • pyannote.audio
  • pydub
  • huggingface_hub

Installation

1. Clone the repository: git clone https://github.com/AmirTahaMim/speaker-diarization.git cd speaker-diarization

2. Install the required Python packages: pip install torch pyannote.audio pydub huggingface_hub

3. Install the required audio codecs for pydub:

   • For Linux: Install ffmpeg via your package manager, e.g., sudo apt install ffmpeg.
   • For Windows/Mac: Download and install ffmpeg from FFmpeg.org.

4. Authenticate with Hugging Face: from huggingface_hub import login login("your_huggingface_token")

Usage

1. Place your audio file in the working directory. The file should be in .wav format.

2. Update the audio_file variable in the script with the path to your audio file.

3. Run the script: python diarization_splitter.py

4. The script will:

   • Perform speaker diarization on the input audio file.
   • Save individual .wav files for each detected speaker in the working directory.

Output

For each detected speaker, the script generates a .wav file named <SPEAKER_ID>.wav containing all segments attributed to that speaker.

Example:

• SPEAKER_0.wav
• SPEAKER_1.wav

Example Code

Here’s a snippet of the main pipeline:

from huggingface_hub import login from pyannote.audio import Pipeline from pydub import AudioSegment import torch

Authenticate and load the pipeline

device = torch.device('cuda' if torch.cuda.is_available() else 'cpu') pipeline = Pipeline.from_pretrained("pyannote/speaker-diarization", use_auth_token="your_token") pipeline.to(device)

Process the audio file

audio_file = "path_to_your_audio.wav" diarization = pipeline(audio_file)

Split audio into speaker-specific segments

audio = AudioSegment.from_wav(audio_file) speaker_segments = {}

for segment, track, speaker in diarization.itertracks(yield_label=True): start_time = segment.start * 1000 # Convert to milliseconds end_time = segment.end * 1000 speaker_audio = audio[start_time:end_time] if speaker not in speaker_segments: speaker_segments[speaker] = [] speaker_segments[speaker].append(speaker_audio)

Save each speaker's audio

for speaker, segments in speaker_segments.items(): combined_audio = AudioSegment.empty() for segment in segments: combined_audio += segment combined_audio.export(f"{speaker}.wav", format="wav")

Notes

• Ensure your Hugging Face token has appropriate permissions to access the pre-trained model.
• Audio files should ideally be in .wav format for compatibility with pydub.

License

This project is licensed under the MIT License.

Contributing

Contributions are welcome! Feel free to submit issues or pull requests.

Acknowledgments

• Hugging Face for the pyannote.audio library.
• PyDub for audio processing.

---

Happy coding! 🎧