William F. Ganong

A cross‐language study of range of voice onset time in the perception of initial stop voicing

A series of experiments was carried out to compare the extent of range effects in the phonetic categorization of voice onset time (VOT) by speakers of Polish and of English, two languages which contrast different VOT categories. Results indicate that Poles are more prone to range effects than are Americans. For acoustic continua with appreciable numbers of prevoiced stimuli, monolingual Polish speakers’ perceptual boundaries fall in the gap between their production categories. For ranges of VOT which include few prevoiced stimuli, their boundaries are substantially shifted. Americans show no shifts of this type, although they do show some small shifts. It was determined that the much smaller shifts shown by the American subjects were not due to expectations about the test. Results are interpreted in terms of the different VOT contrasts involved: their spacing along the VOT continuum, and their psychophysical basis.

Method of optimizing a composite speech recognition expert

In a continuous speech recognizer which includes at least, one acoustic expert and one linguistic expert which generate respective scores, a method is disclosed for adjusting the relative weighting to be applied to those scores employing training data utilizing the words to be recognized in multiple word phrases. Multiple word test phrases are applied to the acoustic expert to determine, for each phrase, plural multi-word hypotheses each having corresponding cumulative scores. The linguistic expert generates corresponding cumulative linguistic scores. An objective function is calculated for each test phrase having a value which is variable as a function of the difference between the combined score of any correct hypothesis and that of the most easily confused incorrect hypothesis. The objective function values are cumulated and a gradient descent procedure is used to adjust the relative weighting of the acoustic and linguistic scores in obtaining a combined score.

Vocabulary partitioned speech recognition apparatus

The speech recognition system disclosed herein operates to select, from a collection of tokens which represent vocabulary words, those tokens which most closely match an unknown spoken word. The collection of tokens is divided into partitions, each of which is characterized or identified by a representative one of the tokens. Both the tokens and the unknown speech word are represented by a sequence of standard data frames which may, for example, define characteristic spectra. In operation, the system computes the distance from the unknown to each of the representative tokens and then, starting with the partition having the nearest representative token and proceeding through partitions represented by successively more distant tokens, examines the other tokens in that partition while keeping a list of predetermined length identifying the examined tokens which thus far provide the best match. This process is continued until the number of distance calculations performed reaches a preselected level.

Measuring phoneme boundaries four ways

In order to determine whether the concept ’’phoneme boundary’’ has validity independent of the particular experimental paradigm used to measure it, four procedures were used to locate phoneme boundaries on two acoustic continua: a VOT continuum, and a formant transition continuum. The procedures were a standard identification test, a version of the method of adjustment, an adaptive psychophysical procedure (PEST), and a version of the method of limits. The procedures differed in reliability, and there were substantial individual differences in the locations of the phoneme boundaries, but the correlations between boundary locations for each subject found by the different methods were quite high. Thus the ’’locus of the phoneme boundary’’ represents an internal variable which can be measured in a number of different ways. We also examined a subsidiary question. A hardware synthesizer does not produce identical waveforms given the same stimulus parameters twice. Do these different tokens produce the same perceptual responses? In the identification tests, different tokens of the various stimuli elicited different average rating responses. Thus, there are reliable (but quite small) token effects for synthetic speech stimuli which can affect phonetic perception.

Selective adaptation of VOT in Polish

In order to test the hypothesis that selective adaptation and range effects in speech perception are produced by the same proccesses, we compared the size of range effects and adaptation effects for monolingual Polish and English subjects. Previous work has indicated that Polish speakers show much larger range effects in phonetic categorization of VOT than do English speakers. We therefore tested the adaptation effect of a +70 ms VOT stimulus on the perception of two apical VOT continua (ranging from −70 to +30 and −30 to +70 ms VOT). Eight monolingual Polish and English subjects participated. As in our previous work, Poles showed large range effects in categorizing VOT stimuli. (Their mean boundaries were at −3.4 vs 8.4 ms VOT.) They also showed substantial selective adaptation boundary shifts (mean = 18.2 ms), which were reliably larger than the adaptation effects shown by the English speakers (mean = 7.6 ms). Thus, the Polish speakers perception of VOT is more sensitive than English speakers to both ...

A word advantage in phoneme boundary experiments

This study investigated one aspect of the way people integrate acoustic information derived from the speech signal with linguistic knowledge. Acoustic continua between words and nonwords were constructed, and subjects phonetic categorizations of the stimuli were examined for a bias toward phonetic renderings which make words. Seven pairs of voice onset time continua were synthesized. For one continuum of each pair, the voiced, but not the voiceless end of the continuum was a word (e.g., “gift‐kift”), and vice versa for the other continuum (e.g.) “giss‐kiss”). First, these stimuli were presented in random order for phonetic categorization of the first segment of each stimulus. Those subjects who correctly identified the endpoints of the continua in a second whole‐syllable identification condition showed, in the first condition, a clear and consistent bias toward phonetic categorizations which made words. All 15 of the subjects and all seven of the continua showed this bias. More detailed examination of th...

Effect of selective adaptation on the discrimination of small differences in voice onset time (VOT)

The effect of selective adaptation (i.e., listening to a rapidly repeated syllable) on the discrimination of voiceless stimuli which differ only in VOT was investigated. In an AX procedure under conditions of minimal stimulus uncertainty, subjects in the unadapted state were able to discriminate (d′=1)small (5 msec) differences in the VOT of stimuli from the voiceless end of a /ba/—/pa/ continuum. Control experiments showed this discrimination was not based on memorization of the aspiration noise waveform, and that subjects could represent small differences in VOT in long‐term memory. Adaptation depressed discrimination slightly, showing that adaptation can affect a perceptual task which does not depend on phonetic categorization. The data were not precise enough to indicate whether adaptation affected a level sensitive to the acoustic similarity of the adapting and test stimuli or to the relationship of the adapting stimulus to the phonetic category.