AdaptoNet Achieves High Accuracy in Multi-Speaker Recognition Using Hybrid Optimization

AdaptoNet, a multi-speaker recognition framework, achieved an accuracy of 99% on a primary speech dataset with 101 speakers, according to experimental evaluations conducted by Aniruddha Mohanty and Ravindranath C. Cherukuri. When evaluated on the AMI Meeting Corpus benchmark to assess generalizability, the framework achieved 98.1% accuracy, 96.3% precision, 95.4% recall, and an F1 score of 95.7%, demonstrating robust and generalizable identification performance.

The system utilizes a structured pre-processing pipeline that applies noise reduction through a Wiener filter, resampling via Windowed Sinc Interpolation, and Root Mean Square Normalization. Following pre-processing, the framework extracts multiple acoustic feature sets, including Mel-Frequency Discrete Cosine Coefficients, spectrogram-based characteristics, and chroma representations.

The extracted spectrogram features encompass Spectral Energy, Spectral Power, Spectral Flux, Spectral Centroid, and Spectral Spread. In addition, the pipeline derives chroma metrics consisting of Chroma Energy Normalized, Chroma Cross-Correlation, and Chroma Deviation to capture distinct spectral profiles.

To select the most discriminative features for classification, the researchers introduced a hybrid model called the Spotted Hyena Customized Flow Regime Optimizer, which integrates the Spotted Hyena Optimizer with the Flow Regime Algorithm. AdaptoNet then merges three neural models using a weighted mean technique to perform multi-speaker recognition.

Subject of Research: Multi-speaker recognition using neural model fusion and hybrid feature optimization

Article Title: Multi-speaker recognition with AdaptoNet and hybrid optimization approach

Article References: Mohanty, A., & Cherukuri, R. C. (2026). Multi-speaker recognition with AdaptoNet and hybrid optimization approach. Multimedia Tools and Applications, 85(11), Article 835. https://doi.org/10.1007/s11042-026-21837-y

Image Credits: AI Generated

DOI: 10.1007/s11042-026-21837-y

Keywords: multi-speaker recognition, AdaptoNet, feature extraction, Wiener filter, Spotted Hyena Optimizer, AMI Meeting Corpus, speech processing, acoustic features, Multi-speaker, recognition, hybrid, optimization