Janne Savela

The mismatch negativity and reaction time as indices of the perceptual distance between the corresponding vowels of two related languages.

Two experiments were conducted to determine whether vowel familiarity affects automatic and conscious vowel discrimination. Familiar (Finnish) and unfamiliar (Komi) vowels were presented to Finnish subjects. The good representatives of Finnish and Komi mid vowels were grouped into three pairs: front /e- epsilon /, central /ø-oe/, and back /o-o/. The acoustic difference for /e- epsilon / and /o-o/ was smaller than that for /ø-oe/. For /e- epsilon /, the Komi vowel / epsilon / was at the boundary between the Finnish /e/ and /ae/. The stimuli were presented in an oddball paradigm. In three different blocks, each Komi vowel in turn served as the standard (probability 0.86) and the corresponding Finnish vowel as the deviant stimulus (probability 0.14), and vice versa. In Experiment 1, subjects were instructed to press a button as soon as they detected a deviant stimulus. In Experiment 2, event-related potentials (ERPs) were recorded to these stimuli in order to use the mismatch negativity (MMN) as an index of the perceptual distance between the members of each vowel pair, while subjects did not attend to the stimuli. There were similar effects of the acoustic distance within a vowel pair for both the reaction time (RT) and the MMN amplitude; the RT decreased and the MMN amplitude increased when the acoustic difference between the stimuli increased. However, the RT was longer when the Komi / epsilon / was the standard and the Finnish /e/ was the deviant than vice versa. No such pattern was found for the MMN. Thus, the phonemic status of the standard stimulus seems to play a role at the attentive but not at the pre-attentive level.

Two dissimilarity measures for HMMS and their application in phoneme model clustering

This paper introduces two approximations of the Kullback-Leibler divergence for hidden Markov models (HMMs). The first one is a generalization of an approximation originally presented for HMMs with discrete observation densities. In that case, the HMMs are assumed to be ergodic and the topologies similar. The second one is a modification of the first one. The topologies of HMMs are assumed to be left-to-right with no skips but the models can have different number of states unlike in the first approximation. Both measures can be presented in a closed form in the case of HMMs with Gaussian (single-mixture) observation densities. The proposed dissimilarity measures were experimented in clustering of acoustic phoneme models for the purposes of multilingual speech recognition. The obtained recognizers were compared to both recognition system based on previously presented dissimilarity measure and one based on phonetic knowledge. The performance of the multilingual recognizers was evaluated in the task of speaker independent isolated word recognition. Small differences were observed in the recognition accuracy of the multilingual recognizers. However, the computational cost of the proposed methods are significantly lower.

Attention modulates the use of spectral attributes in vowel discrimination: Behavioral and event-related potential evidence

Speech contains a variety of acoustic cues to auditory and phonetic contrasts that are exploited by the listener in decoding the acoustic signal. In three experiments, we tried to elucidate whether listeners rely on formant peak frequencies or whole spectrum attributes in vowel discrimination. We created two vowel continua in which the acoustic distance in formant frequencies was constant but the continua differed in spectral moments (i.e., the whole spectrum modeled as a probability density function). In Experiment 1, we measured reaction times and response accuracy while listeners performed a go/no-go discrimination task. The results indicated that the performance of the listeners was based on the spectral moments (especially the first and second moments), and not on formant peaks. Behavioral results in Experiment 2 showed that, when the stimuli were presented in noise eliminating differences in spectral moments between the two continua, listeners employed formant peak frequencies. In Experiment 3, using the same listeners and stimuli as in Experiment 1, we measured an automatic brain potential, the mismatch negativity (MMN), when listeners did not attend to the auditory stimuli. Results showed that the MMN reflects sensitivity only to the formant structure of the vowels. We suggest that the auditory cortex automatically and pre-attentively encodes formant peak frequencies, whereas attention can be deployed for processing additional spectral information, such as spectral moments, to enhance vowel discrimination.

Brain responses reveal hardwired detection of native-language rule violations

Mismatch negativity (MMN) is a neural correlate of the preattentive detection of any change in the acoustic characteristics of sounds. Here we provide evidence that violations of a purely phonological constraint in a listeners native language can also elicit the brains automatic change-detection response. The MMN differed between Finnish and Estonian listeners, conditions being equal except for the native language of the listeners. We used two experimental conditions: synthetic vowels in isolation and the same vowels embedded in a pseudo-word context. MMN responses to isolated vowels were similar for Finns and Estonians, while the same vowels in a pseudoword context elicited different MMN patterns depending on the listeners mother tongue.

Weighted vowel prototypes in Finnish and German

This study explores the perceptual vowel space of the Finnish and German languages, which have a similar vowel system with eight vowels, /ɑ/ /e/ /i/ /o/ /u/ /y/ /æ∼ε/ /ø/. Three different prototypicality measures are used for describing the internal structuring of the vowel categories in terms of the F1 and F2 formant frequencies: The arithmetic mean (centroid) of the F1-F2 space of the category (Pc), the absolute prototype of the category (Pa), and the weighted prototype of the category (Pω), in which the stimulus formant values are weighted by their goodness rating values. The study gave the following main results: (1) in both languages, the inter-subject differences were the smallest in Pω, and on the order of Difference Limen (DL) of F1-F2 frequencies for all of the three measures, (2) the Pa and Pω differed significantly from the centroid, with the absolute prototypes being the most peripheric, (3) the vowel systems of the two languages were similar (Euclidean distances in Pω of Finnish and German 7-34 mels) although minor differences were found in /e/, / ø/, and /u/, and (4) the mean difference of the prototypes from some earlier published production data was 100-150 mels.