Speaker Identification System

Overview

This Speaker Identification System identifies speakers by analyzing their voice embeddings. It registers speakers by recording their voices, computing embeddings with pyannote/embedding, and saving them in a YAML file. For identification, it compares new recordings against saved embeddings using cosine similarity.

Features

• Speaker Registration

  • Record and compute embeddings using pyannote/embedding.
  • Save embeddings in a YAML file.

• Speaker Identification

  • Compute embeddings for new recordings.
  • Compare with saved embeddings and return closest matches.

Embedding Model

Use the pre-trained pyannote/embedding model from Hugging Face.

• Offline Use: Download and configure the model locally for faster processing.
• Online Use: Obtain a Hugging Face token to access the model via the cloud.

Installation

1. Clone the Repository

   git clone https://github.com/Chiraz32/Speaker-Identification.git
   cd speaker-identification

2. Install Dependencies

   pip install -r requirements.txt

3. Configure the YAML File

   • Ensure speakers_embeddings.yaml is present in the root directory.

4. Start the Model

   python embedding_model/ymodel_api.py

5. Start the Main Application

   python main/main.py

Docker

The project includes Docker files to build separate containers for both models and main code:

• Embedding Model Container: Contains the model_api.py for embedding calculations.
• Main Application Container: Contains main.py for registration and identification.

To build and run the Docker containers:

docker build -t embedding-model -f embedding_model/Dockerfile .
docker build -t main-app -f main/Dockerfile .
docker run -p 8000:8000 embedding-model
docker run -p 5000:8080 main-app

Usage

Register a Speaker

• Endpoint: /register_speaker
• Method: POST
• Form Data:
  • audio: .flac file
  • speaker: Unique ID

Identify a Speaker

• Endpoint:/identify_speaker
• Method: POST
• Form Data:
  • audio: .flac file

File Structure

.
├── data
│   └── speakers_embeddings.yaml
├── embedding_model
│   ├── model_api.py
│   ├── Dockerfile
│   └── requirements.txt
├── main
│   ├── main.py
│   ├── Dockerfile
│   └── requirements.txt
└── README.md

Future Enhancements

• Support multiple recordings per speaker for better accuracy.
• Implement a GUI for easier speaker registration and identification.
• Expand to handle noisy environments and multiple languages.
• Improve frontend to specify identification thresholds and compare distances.

Pros: Requires less training data than deep learning models and provides clear results.
Cons: Can be slow with a large number of speakers due to embedding comparisons.

---