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

Area of Science:

• Speech processing
• Human-robot interaction
• Machine learning

Background:

• Speaker identification is crucial for human-robot interaction but challenged by limited training data and noisy, short test utterances.
• Existing deep learning methods struggle with these real-world conditions, lacking optimal solutions.

Purpose of the Study:

• To develop a robust speaker identification architecture for social robots operating in challenging acoustic environments.
• To improve the accuracy of identifying speakers despite low signal-to-noise ratios and short utterance lengths.

Main Methods:

• A two-stage architecture combining prosodic and spectral features (Mel-frequency cepstral coefficients and Gammatone frequency cepstral coefficients).
• Utilized Gaussian mixture models (GMM) and Support Vector Machines (SVM) for gender and speaker classification.
• Employed a weighted Borda count fusion strategy with Mamdani fuzzy inference system for dynamic weight adjustment based on signal quality and utterance length.

Main Results:

• The proposed architecture demonstrated improved speaker recognition performance in low signal-to-noise ratio and short utterance conditions.
• The fusion framework effectively integrated multiple classifiers, enhancing overall system robustness.
• The system showed promise for real-world applications in social robotics.

Conclusions:

• The novel architecture and fusion method offer a promising solution for speaker identification in challenging social robot interaction scenarios.
• This approach effectively addresses limitations of current systems when dealing with noisy and brief speech samples.
• Further development could enhance human-robot communication through more reliable speaker recognition.