آیا ASR برای راه اندازی بی سیم آماده است: اندازه گیری فن آوری محوری برای برنامه های کاربردی انتخاب شده

It is estimated that by the end of 2001 as many as 500 million people worldwide will use cellular services. The nature of hands-busy and eyes-busy situations inherent in the anywhere and anytime wireless communication paradigm presents exciting marketing opportunities and, at the same time, unique technical challenges to the current-generation ASR technology and their new applications. Current industry trends clearly show that incorporating ASR technology into existing or new wireless services as a replacement for touch-tone input is a natural progression in user interface. But is the current-generation ASR technology ready for prime time over wireless channels? Both qualitative and quantitative assessments for the core technology must be adopted by the industry before answering this question. In this paper, we will describe a set of benchmark tasks designed to evaluate the state-of-the-art ASR technologies from a wireless perspective and present the results of these benchmark tests on two commercially available software-based ASR systems that represent the best core ASR technology on the market.

The rapidly growing wireless telephony marketplace of the past five years creates a compelling need for automatic speech recognition (ASR) enabled voice-activated services. A recent survey indicates that a considerable percentage of wireless users (32%) want some type of hands-free voice activation for any services they might want to access from their cell phones (Rietman, 1997). This is quite understandable, given the fact that 83% of all cellular calls in the US are placed inside a vehicle (Rietman, 1997). For this very reason, basic voice-activated dialing (VAD) has become a de facto service pursuit for all major mobile service operators worldwide who try to apply ASR to wireless telephony applications. More recently, ASR has been applied to other wireless applications such as voice messaging and personal access services. The latest deployment of Network Wildfire by Pacific Bell Wireless at their California markets is just one more witness to this long-term trend. As these new applications begin to impact how the mainstream wireless users would use new ASR-based network technologies, it is important to conduct necessary benchmark tests so that we can better predict their field performance. This is one of the major motivations behind this study.

For simple recognition tasks like the grammar G-1 (VS=15; SL5=1; P 6=0), we expect that the today’s best off-the-shelf ASR technology would be robust enough for most wireless applications. However, the technology must improve its rejection capability for applications involving a large number of out-grammar utterances. For more challenging task grammars like G-2 (VS=11; SL=7–11; P=10), the best commercial system (ASR2) still under-performed by a huge margin the comparable lab prototype ( Chien et al., 1997) which achieved a 9.7% string-level error rate on a 9-digit database recorded over a wireless telephony environment. Generally speaking, the gap between the best R&D prototypes and the off-the-shelf products is still significant although it is narrowing considerably as host-based implementations make technology transfer from laboratories to the marketplace more easily. Specifically, the current technology is particular weak in areas of speech detection (i.e., endpoint algorithms) as our test results indicated that both systems performed consistently better on different tasks without the endpoint feature. However, there are studies ( Mauuary and Monne, 1993 and Karray et al., 1998) indicating that the problems in reliably detecting speech endpoints in noisy wireless environments can be overcome with different statistical modeling techniques.