Juhani Toivanen

Durational and tonal correlates of accent in Finnish

Abstract The paper reports the first study that explicitly distinguishes the phonetic correlates of sentence accents from those of word stress in Finnish (stress acted as the baseline whose correlates as against of stress were not investigated). Sentences were constructed that successfully elicited no accent, moderate accent, and strong accent on target words. The three degrees of prominence were clearly differentiated phonetically. Mere word stress was not signalled tonally, while accents were signalled mainly tonally. Strong accent involved longer segmental durations than the other degrees of prominence. Timing and extent of the accentual tonal movements were highly uniform across different word structures, and not invariably tied to the initial (stressed) syllables. The data were consistent with a view that the timing of f0 movements was dependent on the moraic structure of target words. This finding seems to be connected with the typologically rare combination of stress (or rhythmic) and quantity systems in the language.

Automatic Discrimination of Emotion from Spoken Finnish

In this paper, experiments on the automatic discrimination of basic emotions from spoken Finnish are described. For the purpose of the study, a large emotional speech corpus of Finnish was collected; 14 professional actors acted as speakers, and simulated four primary emotions when reading out a semantically neutral text. More than 40 prosodic features were derived and automatically computed from the speech samples. Two application scenarios were tested: the first scenario was speaker-independent for a small domain of speakers while the second scenario was completely speaker-independent. Human listening experiments were conducted to assess the perceptual adequacy of the emotional speech samples. Statistical classification experiments indicated that, with the optimal combination of prosodic feature vectors, automatic emotion discrimination performance close to human emotion recognition ability was achievable.

Emotions in [a]: a perceptual and acoustic study.

The aim of this investigation is to study how well voice quality conveys emotional content that can be discriminated by human listeners and the computer. The speech data were produced by nine professional actors (four women, five men). The speakers simulated the following basic emotions in a unit consisting of a vowel extracted from running Finnish speech: neutral, sadness, joy, anger, and tenderness. The automatic discrimination was clearly more successful than human emotion recognition. Human listeners thus apparently need longer speech samples than vowel-length units for reliable emotion discrimination than the machine, which utilizes quantitative parameters effectively for short speech samples.

Perception of basic emotions from speech prosody in adolescents with Asperger's syndrome

Abstract Aspergers syndrome (AS) belongs to the group of autism spectrum disorders and is characterized by deficits in social interaction, as manifested e.g. by the lack of social or emotional reciprocity. The disturbance causes clinically significant impairment in social interaction. Abnormal prosody has been frequently identified as a core feature of AS. There are virtually no studies on recognition of basic emotions from speech. This study focuses on how adolescents with AS (n=12) and their typically developed controls (n=15) recognize the basic emotions happy, sad, angry, and ‘neutral’ from speech prosody. Adolescents with AS recognized basic emotions from speech prosody as well as their typically developed controls did. Possibly the recognition of basic emotions develops during the childhood.

Classification of emotion in spoken Finnish using vowel-length segments: Increasing reliability with a fusion technique

Classification of emotional content of short Finnish emotional [a:] vowel speech samples is performed using vocal source parameter and traditional intonation contour parameter derived prosodic features. A multiple kNN classifier based decision level fusion classification architecture is proposed for multimodal speech prosody and vocal source expert fusion. The sum fusion rule and the sequential forward floating search (SFFS) algorithm are used to produce leveraged expert classifiers. Automatic classification tests in five emotional classes demonstrate that significantly higher than random level emotional content classification performance is achievable using both prosodic and vocal source features. The fusion classification approach is further shown to be capable of emotional content classification in the vowel domain approaching the performance level of the human reference.

Tone choice and voice quality of dispreferred turns in the English of Finns

Authentic Finnish-English speech data was collected as part of an English conversation class in a Finnish college. Intonation was coded utilizing the framework involving ‘tone’, ‘key’ and ‘termination’. A categorization of voice quality was chosen (e.g., ‘modal voice’, ‘creak’, ‘breathy’, and ‘tense’). The tempo of speech was transcribed with such descriptors as, e.g., ‘fast’ and ‘slow’. The majority of dispreferred turns in the data represented mitigated disagreement, with structural complexity (involving, e.g., wordiness). The p tone was predominant: a low/mid key often accompanied mitigated disagreement. The r and r+ tones were virtually absent in the mitigated dispreferred turns. Instead, the speakers often used a very lax/breathy voice quality and a slow/decelerating tempo, often resulting in creak near a transition relevant place.