vocalid 0.1.8

Personal voice verification toolkit using ECAPA-TDNN embeddings

VocalID: A Lightweight Voice Authentication Toolkit

VocalID is a compact and practical voice authentication library that combines ECAPA-TDNN embeddings with a simple classifier to verify user identity from audio recordings. It supports file-based verification and real-time microphone input. The project is designed to be easy to train, deploy, and# VocalID: A Lightweight Voice Authentication Toolkit

VocalID is a compact and practical voice authentication library that combines ECAPA-TDNN embeddings with a simple classifier to verify user identity from audio recordings. It supports file-based verification and real-time microphone input. The project is designed to be easy to train, deploy, and extend.

---

Features

• ECAPA-TDNN embeddings using speechbrain/spkrec-ecapa-voxceleb
• Training with positive (owner) and negative (impostor) audio samples
• Evaluation with accuracy and classification metrics
• Verification from audio files or live microphone input
• CLI toolkit for training, evaluating, and verifying
• Modular design with trainer, verifier, embeddings, config, and utilities
• Simple model storage using pickle-based persistence
• Full test suite included

---

How It Works

1. Audio Processing

Audio is loaded or recorded, resampled, and normalized.

2. Embedding Extraction

ECAPA-TDNN generates fixed-dimensional speaker embeddings. These embeddings represent unique speaker characteristics.

3. Feature Preparation

Positive and negative embeddings are labeled and fed into the trainer.

4. Classification Model

A simple Logistic Regression model is trained on the embeddings.

5. Verification

During verification:

1. Extract embeddings for the new audio.
2. Predict with trained classifier.
3. Return a confidence score.
4. Compare score with threshold from config.py.

---

Package Structure

VocalID
    └── voice_verifier/
        │
        ├── trainer.py         # Training logic, evaluation, model save/load
        ├── verifier.py        # File and waveform verification
        ├── embeddings.py      # ECAPA-TDNN embedding extraction
        ├── audio_utils.py     # Audio loading and microphone recording
        ├── config.py          # Threshold + ECAPA model configuration
        ├── model_store.py     # Model checkpoint loader
        ├── cli.py             # Command-line interface
    └── tests/                 # Full pytest suite
    └── examples/
    └── requirements.txt
    └── api/
        ├── app.py
    └── README.md

---

Components

VoiceTrainer

Functions: - train() - evaluate() - prepare_features() - save() - load()

VoiceVerifier

Methods: - verify_file(path) - verify_array(audio_tensor)

EmbeddingExtractor

• embed_file(path)
• embed_waveform(waveform, sr)

Audio Utilities

• load_audio(path)
• record_audio(seconds)

---

Installation

    pip install vocalid

---

Example usage script(Python)

Directory Structure Example

Assume your dataset looks like this:

Voice tip: Each voice sample of 5-6 seconds with different tone/ bg noise/ accent/ microphone

└── dataset/
    └── my_voice/               <-- positive class (your voice)
        sample1.wav
        sample2.wav
        sample3.wav
        sample4.wav
        
    └── other_voices/           <-- negative class(other's voices)
        voice1.wav
        voice2.wav
        voice3.wav
        voice4.wav

Full python script

from vocalid.trainer import VoiceTrainer
from vocalid.verifier import VoiceVerifier
from vocalid.audio_utils import load_audio
import glob

# 1. TRAINING THE MODEL

pos_files = glob.glob("dataset/my_voice/*.wav")
neg_files = glob.glob("dataset/other_voices/*.wav")

trainer = VoiceTrainer()
trainer.train(pos_files, neg_files, save_path="my_voice_model.pkl")

# (Optional) Check metrics printed by evaluate() in train()
print("Training complete. Model saved.")


# 2. EVALUATING THE TRAINED MODEL (Manually)

# This is useful if you want to evaluate after loading the model.
# Or you want to compute new metrics on a different test set.

# Example test data (can be same folders or separate ones)
test_pos = glob.glob("dataset/my_voice_test/*.wav")
test_neg = glob.glob("dataset/other_voices_test/*.wav")

metrics = trainer.evaluate(test_pos, test_neg)

print("Accuracy:", metrics["accuracy"])
print("Report:\n", metrics["report"])

# Example output:
# Classification report text
# Accuracy: 0.91


# 3. VERIFY A FILE

verifier = VoiceVerifier("my_voice_model.pkl")

to_verify = "verify_samples/unknown_voice.wav"
ok, score = verifier.verify_file(to_verify)

print(f"\nVerification result: {ok}, Score: {score:.3f}")
# ok = True means it matches your voice
# score is probability from the classifier


# 4. VERIFY LIVE MICROPHONE AUDIO (Windows supported)

# Record a short clip and verify
audio_tensor = trainer.record_audio(seconds=4)
ok, score = verifier.verify_array(audio_tensor)

print(f"Live verification: {ok}, Score: {score:.3f}")

Example Evaluate-Only Script

If someone just wants to evaluate the model later:

from vocalid.trainer import VoiceTrainer
import glob

trainer = VoiceTrainer()
trainer.load("my_voice_model.pkl")

test_pos = glob.glob("dataset/my_voice_test/*.wav")
test_neg = glob.glob("dataset/other_voices_test/*.wav")

metrics = trainer.evaluate(test_pos, test_neg)

print("Accuracy:", metrics["accuracy"])
print("Report:\n", metrics["report"])

CLI Commands

    vocalid train --positive my_voice --negative others --output model.pkl
    vocalid evaluate --model model.pkl
    vocalid verify audio.wav --model model.pkl
    vocalid live --model model.pkl --seconds 4

---

Use Cases

• Personal voice unlock systems
• Lightweight identity verification
• Speaker recognition prototypes
• Research experiments in speaker embeddings
• Security analyses for spoof detection

---

Why It Matters

This toolkit allows developers and researchers to:

• Build practical speaker authentication systems quickly
• Learn how ECAPA embeddings work
• Train custom voiceprints without heavy dependencies
• Extend or plug into larger voice systems

---

Contributing

Pull requests are welcome.
Tests can be run with:

    pytest -v

Author

Muhammad Khubaib Ahmad\

AI/ML Engineer, Data Scientist and Voice Intelligence Researcher

Portfolio and Links

• Portfolio:
• Gmail:
• GitHub:
• LinkedIn:
• Kaggle:
• HuggingFace:

---

License

MIT License