How Should a Speech Recognizer Work?

Although researchers studying human speech recognition (HSR) and automatic speech recognition (ASR) share a common interest in how information processing systems (human or machine) recognize spoken language, there is little communication between the two disciplines. We suggest that this lack of communication follows largely from the fact that research in these related fields has focused on the mechanics of how speech can be recognized. In Marr's (1982) terms, emphasis has been on the algorithmic and implementational levels rather than on the computational level. In this article, we provide a computational-level analysis of the task of speech recognition, which reveals the close parallels between research concerned with HSR and ASR. We illustrate this relation by presenting a new computational model of human spoken-word recognition, built using techniques from the field of ASR that, in contrast to current existing models of HSR, recognizes words from real speech input.     

Hemispheric differences in the effects of context on vowel perception

Listeners perceive speech sounds relative to context. Contextual influences might differ over hemispheres if different types of auditory processing are lateralized. Hemispheric differences in contextual influences on vowel perception were investigated by presenting speech targets and both speech and non-speech contexts to listeners’ right or left ears (contexts and targets either to the same or to opposite ears). Listeners performed a discrimination task. Vowel perception was influenced by acoustic properties of the context signals. The strength of this influence depended on laterality of target presentation, and on the speech/non-speech status of the context signal. We conclude that contrastive contextual influences on vowel perception are stronger when targets are processed predominately by the right hemisphere. In the left hemisphere, contrastive effects are smaller and largely restricted to speech contexts.     

Interference of spoken word recognition through phonological priming from visual objects and printed words

Three cross-modal priming experiments examined the influence of preexposure to pictures and printed words on the speed of spoken word recognition. Targets for auditory lexical decision were spoken Dutch words and nonwords, presented in isolation (Experiments 1 and 2) or after a short phrase (Experiment 3). Auditory stimuli were preceded by primes, which were pictures (Experiments 1 and 3) or those pictures’ printed names (Experiment 2). Prime–target pairs were phonologically onset related (e.g., pijl–pijn, arrow–pain), were from the same semantic category (e.g., pijl–zwaard, arrow–sword), or were unrelated on both dimensions. Phonological interference and semantic facilitation were observed in all experiments. Priming magnitude was similar for pictures and printed words and did not vary with picture viewing time or number of pictures in the display (either one or four). These effects arose even though participants were not explicitly instructed to name the pictures and where strategic naming would interfere with lexical decision making. This suggests that, by default, processing of related pictures and printed words influences how quickly we recognize spoken words.     

Foreign accent strength and listener familiarity with an accent codetermine speed of perceptual adaptation

We investigated how the strength of a foreign accent and varying types of experience with foreign-accented speech influence the recognition of accented words. In Experiment 1, native Dutch listeners with limited or extensive prior experience with German-accented Dutch completed a cross-modal priming experiment with strongly, medium, and weakly accented words. Participants with limited experience were primed by the medium and weakly accented words, but not by the strongly accented words. Participants with extensive experience were primed by all accent types. In Experiments 2 and 3, Dutch listeners with limited experience listened to a short story before doing the cross-modal priming task. In Experiment 2, the story was spoken by the priming task speaker and either contained strongly accented words or did not. Strongly accented exposure led to immediate priming by novel strongly accented words, while exposure to the speaker without strongly accented tokens led to priming only in the experiment’s second half. In Experiment 3, listeners listened to the story with strongly accented words spoken by a different German-accented speaker. Listeners were primed by the strongly accented words, but again only in the experiment’s second half. Together, these results show that adaptation to foreign-accented speech is rapid but depends on accent strength and on listener familiarity with those strongly accented words.     

Constraints on the processes responsible for the extrinsic normalization of vowels

Listeners tune in to talkers’ vowels through extrinsic normalization. We asked here whether this process could be based on compensation for the long-term average spectrum (LTAS) of preceding sounds and whether the mechanisms responsible for normalization are indifferent to the nature of those sounds. If so, normalization should apply to nonspeech stimuli. Previous findings were replicated with first-formant (F1) manipulations of speech. Targets on a [pt]–[p?t] (low–high F1) continuum were labeled as [pt] more after high-F1 than after low-F1 precursors. Spectrally rotated nonspeech versions of these materials produced similar normalization. None occurred, however, with nonspeech stimuli that were less speechlike, even though precursor–target LTAS relations were equivalent to those used earlier. Additional experiments investigated the roles of pitch movement, amplitude variation, formant location, and the stimuli's perceived similarity to speech. It appears that normalization is not restricted to speech but that the nature of the preceding sounds does matter. Extrinsic normalization of vowels is due, at least in part, to an auditory process that may require familiarity with the spectrotemporal characteristics of speech.     

Shortlist B: a Bayesian model of continuous speech recognition

A Bayesian model of continuous speech recognition is presented. It is based on Shortlist (D. Norris, 1994; D. Norris, J. M. McQueen, A. Cutler, & S. Butterfield, 1997) and shares many of its key assumptions: parallel competitive evaluation of multiple lexical hypotheses, phonologically abstract prelexical and lexical representations, a feedforward architecture with no online feedback, and a lexical segmentation algorithm based on the viability of chunks of the input as possible words. Shortlist B is radically different from its predecessor in two respects. First, whereas Shortlist was a connectionist model based on interactive-activation principles, Shortlist B is based on Bayesian principles. Second, the input to Shortlist B is no longer a sequence of discrete phonemes; it is a sequence of multiple phoneme probabilities over 3 time slices per segment, derived from the performance of listeners in a large-scale gating study. Simulations are presented showing that the model can account for key findings: data on the segmentation of continuous speech, word frequency effects, the effects of mispronunciations on word recognition, and evidence on lexical involvement in phonemic decision making. The success of Shortlist B suggests that listeners make optimal Bayesian decisions during spoken-word recognition.