A Two-Level Speaker Identification System via Fusion of Heterogeneous Classifiers and Complementary Feature Cooperation

Abstract

We present a new architecture to address the challenges of speaker identification that arise in interaction of humans with social robots. Though deep learning systems have led to impressive performance in many speech applications, limited speech data at training stage and short utterances with background noise at test stage present challenges and are still open problems as no optimum solution has been reported to date. The proposed design employs a generative model namely the Gaussian mixture model (GMM) and a discriminative model—support vector machine (SVM) classifiers as well as prosodic features and short-term spectral features to concurrently classify a speaker’s gender and his/her identity. The proposed architecture works in a semi-sequential manner consisting of two stages: the first classifier exploits the prosodic features to determine the speaker’s gender which in turn is used with the short-term spectral features as inputs to the second classifier system in order to identify the speaker. The second classifier system employs two types of short-term spectral features; namely mel-frequency cepstral coefficients (MFCC) and gammatone frequency cepstral coefficients (GFCC) as well as gender information as inputs to two different classifiers (GMM and GMM supervector-based SVM) which in total leads to construction of four classifiers. The outputs from the second stage classifiers; namely GMM-MFCC maximum likelihood classifier (MLC), GMM-GFCC MLC, GMM-MFCC supervector SVM, and GMM-GFCC supervector SVM are fused at score level by the weighted Borda count approach. The weight factors are computed on the fly via Mamdani fuzzy inference system that its inputs are the signal to noise ratio and the length of utterance. Experimental evaluations suggest that the proposed architecture and the fusion framework are promising and can improve the recognition performance of the system in challenging environments where the signal-to-noise ratio is low, and the length of utterance is short; such scenarios often arise in social robot interactions with humans.

Improved Feature Parameter Extraction from Speech Signals Using Machine Learning Algorithm

“…Feature extraction is accomplished by changing the speech waveform to a form of parametric representation at a relatively lesser data rate for subsequent processing and analysis [ 11 , 12 , 13 , 14 ]. Feature extraction approaches usually yield a multidimensional feature vector for every speech signal.…”

Fusion of Cochleogram and Mel Spectrogram Features for Deep Learning Based Speaker Recognition

“…It does not correspond linearly to the physical frequency of the tone, as the human auditory system apparently does not perceive pitch linearly. The Mel Filter is approximately a linear frequency spacing below 1 kHz and a logarithmic spacing above 1 kHz [4].…”

“…In addition to fusion of feature extraction techniques, fusion of different types of classi ers are applied to improve the performance of speaker recognition. In [4] fusion of GMM and SVM is used to develop a speaker recognition system with MFCC and GFCC feature extraction techniques. In the study [9] speaker veri cation is developed by using fusion of GMM and ANN models with GFCC features.…”

“…It has important applications in the areas like: Access control, forensic science, surveillance, law enforcement and nancial areas for the purpose of identi cation, detection, segmenting and clustering [2]. Previously, most of the speaker recognition systems are developed using a machine learning classi ers like Gaussian Mixture Model (GMM) [3], Support Vector Machine (SVM) [4], i-vector [5] and Hidden Markov Model (HMM). These classi ers use handcrafted features mainly Mel Frequency Cepstral Coe cient (MFCC) [3], Gammatone Frequency Cepstral Coe cient (GFCC) [6] and Linear Predictive Cepstral Coe cient (LPCC) [7] for speaker enrollment and recognition.…”

Attention-Based Temporal-Frequency Aggregation for Speaker Verification

“…Some well-designed metric learning losses have been exploited to train the entire SV system in an end-to-end fashion, such as triplet loss [ 9 , 11 ], generalized end-to-end (GE2E) loss [ 18 ], and cluster-range loss [ 4 ]. Besides, many studies on robust features [ 20 , 21 ] and hybrid models [ 21 , 22 ] have been conducted to further improve the performance of traditional and DNN-based speaker recognition systems. In recent years, CNN has drawn much attention in this research field.…”