Vocal-Part-Aware Singer Identification with MFCCs and LSTM Classification
Kabita Thaoroijam, Sri Raman Kothuri, L. Bhagyalakshmi, Sanjay Kumar Suman
Abstract
Automatic singer identification benefits greatly from focusing on vocal-only regions and modeling long-range temporal dependencies in timbre and articulation. We propose a vocal-part-aware pipeline that combines precise vocal segment detection, Melfrequency cepstral coefficient (MFCC) feature extraction, and Long Short-Term Memory (LSTM) classifiers for joint singer and gender recognition. Training uses curated vocal segments from the MIR-1K corpus and additional in-house recordings, followed by sequence modeling with stacked bidirectional LSTMs. At test time, incoming audio is first vocal-segmented, transformed to MFCC sequences, and classified by two heads: an LSTM for singer identity and a separate LSTM for gender. The system integrates utterance-level attention pooling to emphasize informative frames and a confidence-weighted decision fusion across segments from the same track. In experiments, the method delivers 97.6% top-1 singer identification accuracy and 99.2 % gender accuracy, outperforming CNN baselines (spectrogram-only) and traditional GMM-UBM approaches. Ablations show vocal-segment selection and attention pooling contribute most to gains, reducing confusions among timbrally similar singers and improving robustness to accompaniment bleed. The approach is lightweight (sub- <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$3 M$</tex> parameters), real-time on commodity GPUs, and applicable to music retrieval, rights management, and content analytics.