Glanz, M.M.GlanzLinhard, K.K.LinhardKroschel, K.K.Kroschel2022-03-082022-03-081990https://publica.fraunhofer.de/handle/publica/317908When considering impediments to speech recognition systems in cars, only road noise is normally taken into account. This paper not only examines interference due to this noise source, but also that due to the output from a car radio. The speech recognition system, which is a state-of-the-art speaker-dependent isolated-word recognizer, is operated hands-free. Road noise suppression is based on spectral subtraction. A frequency domain LMS algorithm including a modification to control the convergence rate is used to compensate the output signal from the car radio. Recognition tests were performed with several speakers travelling at 100 km/h and 140 km/h with the radio switched on. Applying noise suppression and compensation of the radio output signal, the original recognition error rate of the basic algorithm was reduced from about 50% to 4% at 100km/h and from about 100% to 10% at 140 km/h.enfrequency least mean square algorithmnoise compensationnoise suppressionspectral subtractionspeech recognition004conference paper