Sven Nordholm
Ingvar Claesson
Sven Nordebo
Mattias Dahl
Dept. of Signal Process., Univ. of Karlskrona/Ronneby S-372 25 Ronneby, Sweden
By employing speech generation models and new algorithms more and more a priori information about speech signals is utilized in speech recognition and speech coding. A fair signal-to-noise ratio is therefore required to ensure that the a priori information is correct. This implies a need for noise reduction under adverse conditions, such as hands-free operation of telephones in the car compartment or speech recognition in cars [S. Nordholm et al., ``Adaptive Array Noise Suppression of Handsfree Speaker Input in Cars,'' IEEE Trans. Veh. Tech. 42, 514--518 (1993)]. The paper presents two adaptive microphone array schemes, aimed for this situation. The first, denoted spatial filtering generalized sidelobe canceller (SFGSC), gives good noise suppression with little distortion of the speech but requires careful calibration. The second, denoted adaptive microphone array employing calibration signals recorded on-site (AMAEC), facilitates a simple built-in calibration. It is beneficial from a user point of view to use a calibration signal recorded on site eliminating amplifier tuning and microphone selection. The calibration can be done within 60 s. The AMAEC calibrates the array to the speakers' location, microphone positions and lobe gains, amplifiers, and to the acoustic environment in the car. No a priori information about signal statistics or array geometry is utilized. [Work supported by Nutek.]