The Collection of Spoken Language Resources in Car Environments

Over the last two years, we have recorded 400 speakers in five different mid-range and top-range cars using single-channel, fourchannel, or seven-channel recording equipment (see Table 1). This paper documents the acquired knowledge of microphone selection, positioning and fixation, of preamplifying and recording devices, of shielding, grounding and power supply, and of data processing. On the basis of our experience, we would like to make recommendations for the collection of speech data in the car environment, in order to help others avoid the mistakes we made. Furthermore, we would like to define a standard for this specific recording situation so that car speech data recorded by different institutions in various places can nevertheless be uniform. As a result, the exchange of such databases will be considerably more interesting for such institutions in the future. 1 The need for car speech databases For speech recognition and speech synthesis, the automotive market is one of the greatest future fields of application. While acceptance of speech recognition and speech synthesis in the office turns out to be relatively low, in the car this man-machine-interface is not only helpful but even essential in the case of navigation systems because eyes, hands and feet are required for driving (Van Compernolle, 1997 [10]). However, with the attempts to employ previous speech recognition technology in the car it turned out that the error rates increased enormously. First successes were only obtained after training the systems with authentic speech material, recorded in the car (the driver speaks), specific noise removal techniques (Schless & Class, 1997 [8], Wang et al., 1993 [11]), and channel adaption techniques (Shozakai et al., 1997 [9]). It is presumed that the error rates have not achieved telephone speech error rates up to now for the following two reasons (Langmann et al., 1997 [5]): First, the available car speech databases are small. Second, the signal-to-noise ratio is lower and moreover is time-variant. Consequently, when designing a car speech database the collection costs should be as small as possible so that large databases can be collected, and recording CSDC customer sub cars ch 2 Siemens 238 BMW540, Opel Astra 4/7 1 Philips 159 Golf, Fiat, Hyundai 1 4 VW 103 Golf 4 Tab. 1: Brief overview of the car speech databases (CSDC) we collected (for details see Langmann et al, 1998 [4]). quality should not be reduced by inappropriate procedures (Chi & Oh, 1996 [1]). 2 Speech recordings in the car 2.1 Microphone selection Three different types of microphones were used for speech recording in the car: 1. The condenser cardioid microphone beyerdynamic MCE-10 with phantom power was choosen as a reference microphone because of its small size and its studio quality. 2. The AKG Q-400-II mouse microphone had already proved effective for the car environment in the past because of its highpass filter characteristic. 3. With regard to investigations concerning the difference of effect between expensive and cheap microphones we used One-Dollar microphone modules. These three microphone types had cardioid directional characteristics. Since the expenditure for microphone arrays is considered excessive by the car manufacturers, we did not take them into account (Grenier, 1993 [3]). 2.2 Microphone positioning The car industry favours the following three positions for microphones in the car shown in figure 1.