How nagmdrs/speaker_recognition works
Project type and purpose
• Type: End-to-end machine-learning pipeline for speaker recognition
• Purpose: Ingest raw audio, compute mel spectrograms, feed into ECAPA-TDNN model, train/evaluate, and produce embeddings + visualizations (attention maps, t-SNE plots).
File structure
Top-level components:
• Configuration: cfg.py, config.yaml
• Core model: base_model.py (ECAPA-TDNN + mel frontend)
• Orchestration wrappers: train_model.py (training + t-SNE), main_model.py (train/eval/inference), plot_attention.py (attention viz)
• Utilities: helperFiles/ (dataLoader.py, losses.py, tools.py, list generators)
• Artifacts: exps/ (checkpoints), Results/ (scores), params/ (speaker lists)
README for additional insights
• Architecture: SE-Res2Blocks, attentive pooling, channel attention
• Outputs: 192-dim embeddings
• Metrics: EER, minDCF
• Dataset: VoxCeleb (external)
Main components & interactions
a. Data Ingestion & Preprocessing
– Scripts: helperFiles/create_train_list.py, create_eval_list.py
– Inputs: raw audio files (external VoxCeleb dataset)
– Outputs: params/train_list.txt, params/eval_list.txt
– DataLoader (helperFiles/dataLoader.py) reads lists → loads waveforms → computes mel spectrograms via frontend in base_model.py
b. Model (ECAPA-TDNN)
– Module: base_model.py
– Receives mel spectrograms → SE-Res2Blocks → attentive pooling → 192-dim embeddings
c. Training & Evaluation
– Script: train_model.py
• Uses config.yaml/cfg.py parameters
• Feeds embeddings + labels into specified loss functions (helperFiles/losses.py)
• Logs metrics (EER, minDCF)
• Saves checkpoints under exps/expX/
• Generates t-SNE plots
– Alternative orchestration: main_model.py (wraps train/eval/inference routines)
d. Visualization & Post-processing
– Script: plot_attention.py
• Loads trained checkpoint
• Runs attention maps over mel spectrograms
• Outputs plots to Images/ or Results/
e. Configuration & Utilities
– cfg.py & config.yaml hold hyperparameters (learning rate, batch size, model depths, mel settings)
– helperFiles/tools.py provides logging, metric computation, etc.
Architectural patterns & design principles
• Separation of concerns: distinct modules for config, data loading, model definition, training loops, visualization
• Modular layering:
– Data layer → Model layer → Training/Eval layer → Visualization layer
• Reusability: helperFiles utilities used across scripts
• Checkpointing & experiment tracking: exps/ directory per experiment