F ASR TECHNOLOGY
Consider the following four scenarios:
1. A court reporter listens to the opening arguments of the defense and types the words into a steno-machine attached to a word-processor.
2. A medical doctor activates a dictation device and speaks his or her patient's name, date of birth, symptoms, and diagnosis into the computer. He or she then pushes "End input" and "print" to produce a written record of the patient's diagnosis.
3. A mother tells her three-year old, "Hey Jimmy, get me my slippers, will you?" The toddler smiles, goes to the bedroom, and returns with papa's hiking boots.
4. A first-grader reads aloud a sentence displayed by an automated Reading Tutor. When he or she stumbles over a difficult word, the system highlights the word, and a voice reads the word aloud. The student repeats the sentence - this time correctly - and the system responds by displaying the next sentence.
At some level, all four scenarios involve speech recognition. An incoming speech signal elicits a response from a "Listener." In the first two instances, the response consists of a written transcript of the spoken input, whereas in the latter two cases, an action is performed in response to a spoken command. In all four cases, the "Success" of the voice interaction is relative to a given task as embodied in a set of expectations that accompany the input. The interaction succeeds when the response - by a machine or human "listener" - matches these expectations.
Recognizing and understanding human speech requires a considerable amount of linguistic knowledge: a command of the phonological, lexical, semantic, grammatical, and pragmatic conventions that constitute a language. The listener's command of the language must be "up" to the recognition task or else the interaction fails. Jimmy returns with the wrong items, because he cannot yet verbally discriminate between different kinds of shoes. Likewise, the reading tutor would miserably fail in performing the court-reporter's job or transcribing medical patient information, just as the medical dictation device would be a poor choice for diagnosing a student's reading errors. On the other hand, the human court reporter - assuming he or she is an adult native speaker - would have no problem performing any of the tasks mentioned under (1) through (4). The linguistic competence of an adult native speaker covers a broad range of recognition tasks and communicative activities. Computers, on the other hand, perform best when designed to operate in clearly circumscribed linguistic sub-domains.
Humans and machines process speech in fundamentally different ways (Bernstein & Franco, 1996). Complex cognitive processes account for the human ability to associate acoustic signals with meanings and intentions. For a computer, on the other hand, speech is essentially a series of digital values. However, despite these differences, the core problem of speech recognition is the same for both humans and machines: namely, of finding the best match between a given speech sound and its corresponding word string. Automatic speech recognition technology attempts to simulate and optimize this process computationally.
Since the early 1970s, a number of different approaches to ASR have been proposed and implemented, including Dynamic Time Warping, template matching, knowledge-based expert systems, neural nets, and Hidden Markov Modeling (HMM) (Levinson & Liberman, 1981; Weinstein, McCandless, Mondshein, & Zue, 1975; for a review, see Bernstein & Franco, 1996). HMM-based modeling applies sophisticated statistical and probabilistic computations to the problem of pattern matching at the sub-word level. The generalized HMM-based approach to speech recognition has proven an effective, if not the most effective, method for creating high-performance speaker-independent recognition engines that can cope with large vocabularies; the vast majority of today's commercial systems deploy this technique. Therefore, we focus our technical discussion on an explanation of this technique.
An HMM-based speech recognizer consists of five basic components: (a) an acoustic signal analyzer which computes a spectral representation of the incoming speech; (b) a set of phone models (HMMs) trained on large amounts of actual speech data; (c) a lexicon for converting sub-word phone sequences into words; (d) a statistical language model or grammar network that defines the recognition task in terms of legitimate word combinations at the sentence level; (e) a decoder, which is a search algorithm for computing the best match between a spoken utterance and its corresponding word string. Figure 1 shows a schematic representation of the components of a speech recognizer and their functional interaction.
В
Figure 1. Components of a speech recognition device
A. Signal ...
