Noise-Robust Speaker Identification in Cars
* Presenting author
Speech is one of the important communication tools between the human and the machine within a car. Besides voice recognition, the speaker identity is an important information extractable from the speech signal. By detecting the speakers, the infotainment system may automatically be personalized or utilized for speaker dependent speech recognition. The main focus of this work lies in audio-based speaker identification in cars using the existing hands-free system.Many different features like Gammatone and Mel Frequency Cepstral Coefficients, Linear Predictive Cepstral Coefficients and corresponding delta and delta-delta features are extracted from the speech. Linear Discriminant Analysis is used to reduce the dimensionality of the features. Gaussian Mixture Models are used as the initial classifier. The system is implemented to operate in real-time.The speech data is collected from 12 different speakers at four different positions inside the car, using the built-in hands-free microphones. The actual driving noise, measured at 60 & 120 km/h, is added to the signals. The classification accuracy is measured for respective test cases to assess the performance of the system under realistic acoustic scenarios.