Feature Extraction and Modeling of Voice and Text Data for Post‑Loan Management

The article introduces the importance of leveraging unstructured data—voice and text—in post‑loan risk management, outlining how these data sources can be transformed into actionable features for modeling and strategy.

Voice Feature Extraction – Traditional methods include manual rule‑based identification of invalid communication (e.g., low amplitude or AI‑assistant signals) and the use of open‑source packages such as pyaudioanalysis to extract short‑time zero‑crossing rate, energy entropy, and MFCCs. Two case studies demonstrate rule‑based tagging for loss‑contact detection and a promise‑to‑pay (PTP) repayment prediction model built with XGBoost, achieving modest AUC improvements.

Deep‑learning approaches apply short‑time Fourier transform (STFT) to generate mel‑spectrograms, convert amplitudes to decibels, and feed the resulting 2‑D representations into a CRNN (CNN + RNN) architecture. One‑dimensional convolutions focus on temporal patterns, and multi‑segment convolutions capture localized features. The CRNN outputs are passed through LSTM layers and a final fully‑connected layer, yielding AUC around 0.56 on validation and test sets.

Text Feature Extraction – Traditional techniques start with keyword spotting (e.g., gambling‑related terms) and extend to bag‑of‑words and TF‑IDF vectorization using scikit‑learn’s CountVectorizer and TfidfVectorizer . These vectors feed naive Bayes or XGBoost classifiers, achieving AUC ≈ 0.58 for repayment prediction.

Deep‑learning methods encode questions and answers with simple ASCII‑based embeddings, construct association matrices to align query‑answer pairs, and apply transformer models (the basis of BERT) followed by dense layers for classification. Incorporating semantic information improves AUC slightly over the audio‑only model.

The final recommendation is to fuse both audio and text models via stacking or weighted ensembles, enabling risk teams to convert unstructured interactions into structured features for downstream scoring and rule‑based decision making.