Abstract
Automatic Speaker Recognition (ASkR) is a critical capability for humanoid robots interacting in real-world environments. This study explores how the binaural nature of robotic "hearing"—having two microphones positioned like human ears—can be leveraged to improve recognition accuracy and noise robustness. By comparing two signal processing strategies (concatenation and intercorrelation) and two classification models (Predictive Neural Networks and Gaussian Mixture Models), the researchers demonstrate that cross-correlating signals from both ears significantly outperforms standard monaural approaches, especially in noisy conditions.
Introduction
Most ASkR systems are optimized for single-microphone inputs or large arrays. Humanoid robots, however, typically possess a binaural configuration. This setup introduces complexities such as spatial filtering by the robot's head (Head-Related Transfer Functions) and varying speaker positions. The primary challenge addressed here is maintaining high recognition rates in the presence of background noise and reverberation using only two spatial audio points.
Binaural Processing and Classification Methods
The researchers evaluated two distinct pipelines for handling two-channel audio:
- Concatenation: MFCC features are extracted independently from each ear and then stacked into a single feature vector.
- Intercorrelation: The raw signals from both ears are cross-correlated prior to feature extraction. This method aims to emphasize the common speech signal while suppressing uncorrelated background noise.
For classification, two models were compared:
- Predictive Neural Networks (PNN): A per-speaker network that attempts to predict subsequent audio frames based on previous inputs.
- Gaussian Mixture Models (GMM): A statistical model representing speaker identities as clusters in the feature space, which is a standard technique in speech processing.
Experimental Setup
The study utilized a database of French radiophonic monologues featuring 10 speakers. To simulate a humanoid's auditory experience, clean recordings were processed using KEMAR dummy-head Head-Related Transfer Functions (HRTFs). This process simulates how sound is filtered by the human head and torso from various spatial directions, making the training data more representative of real-world robotic perception.
Key Results and Discussion
The experimental results highlight several critical findings:
- The Intercorrelation Advantage: Intercorrelating the two ear signals acts as a natural noise filter. Because noise is often spatially distinct or random, it does not correlate across channels as strongly as the primary speaker's voice. This pre-filtering significantly improves resilience at low Signal-to-Noise Ratios (SNR), such as 0dB and -3dB.
- Model Performance: GMMs consistently outperformed PNNs. The researchers attribute this to the effectiveness of clustering MFCC features compared to the difficulty of predicting highly variable speech sequences.
- Binaural vs. Monaural: The binaural GMM approach achieved at least a 10% improvement in recognition rates over traditional monaural methods.
- Temporal Integration: Recognition accuracy increases sharply with the duration of the test sample. Listening for 3 to 15 seconds allows the system to aggregate information across many frames, leading to near-perfect recognition in high-SNR environments.
Conclusion
Leveraging binaural audio through intercorrelation provides a robust framework for speaker recognition in humanoid robots. While GMMs currently offer superior performance, future work will focus on integrating more sophisticated Voice Activity Detection (VAD) and adapting the system for dynamic environments where both the robot and the speaker are in motion.
Summary updated