Cognition of Phones

Abstract

The places of articulation of vowels, plosives, and affricates are generally believed to be perceived on the basis of the formant frequencies, particularly of the first two formats and their dynamic behavior. Bangla, like most of the major Indian languages, have a large number plosive/stop sounds, around 20 in number. These are organized into four groups based on place of articulation, generally named as velar, alveolar (retroflexed in most language), dental, and labial. Each group again has five different manners of production, namely, unaspirated unvoiced, aspirated unvoiced, unaspirated voiced, aspirated voiced, and nasal. There has been a considerable interest in the recognition of place of articulation since early 50s for both human cognition and Automatic Speech Recognition (ASR) in Indian Statistical Institute. The chapter presents details of this endeavor. For this investigation, 400 segments of each of the seven Bangla vowels extracted from most frequently used Bangla words embedded in a neutral carrier sentence and spoken by four male and four female native informants were used. For objective studies, attention was mainly limited to the first two formant frequencies, namely, F₁ and F₂. The distribution of different Bangla vowels in F₁–F₂ plane and their spreads are described. These show considerable overlap. A careful perusal of the distribution reveals that there are regions where more than one vowel coexist. However, the clusters though not disjoint show reasonable power of discrimination (about 85% recognition rate has been reported for Bangla vowels). This chapter also presents a detailed study in the cognitive domain to test whether the aforesaid formant-related parameters generally used for machine recognition are also cognitively relevant. For this purpose, the hypothesis tested is that the first two formants necessary and sufficient for human cognition of the place of articulation of plosives are determined by the transition of the first two formants and the place of articulation of vowels are determined by the steady-state values of the first two formants. The significance of the hypothesis tested was not good enough. Specially prepared signals from actual speech sounds are used for the purpose of cognitive experiments. The technique of preparation is described in detail. Listening tests are conducted with 30 native listeners. Experimental procedure and corresponding results on objective identification of vowels and plosives on a large speech database collected from native informants of both sexes are presented in details. The study on the cognition of place of articulation required fine-scale manipulations at the signal level which are also described in necessary details. The results of cognition with these manipulated signals are discussed. The chapter also includes an in-depth study of the spectral cues for nasal/oral distinction in Bangla vowels. This acquires significance simply for the fact that in Bangla; in contrast with English, nasality is phonemic. A comprehensive study in the cognitive domain identified that only one or two cognitive cues, usually not referred to in literature, are really necessary and sufficient for nasal/oral distinction. The paradigm of analysis through synthesis has been used for this investigation.