Methodology Overview The pipeline consists of four main components:

Audio Preprocessing (preprocess.py):

Background noise removal using spectral gating and high-pass filtering Leading/trailing silence removal Pause detection with configurable duration thresholds

Speech-to-Text Conversion (aud_to_text.py):

Audio transcription using OpenAI's Whisper model (via Hugging Face transformers) Output in structured formats (text and CSV)

Feature Extraction (analysis.py):

Gemini API integration for sophisticated linguistic analysis Extraction of 15 standardized features across four categories:

Fluency and hesitation markers Prosodic and temporal characteristics Lexical retrieval abilities Sentence structure and completion metrics

Machine Learning Analysis (ml.py):

Primary Method: Isolation Forest for anomaly detection

Unsupervised learning approach suitable for detecting deviations from normal speech patterns Contamination parameter set to 0.3 to identify the most anomalous 30% of samples

Feature importance analysis using Spearman correlation Dimensionality reduction using PCA for visualization