Local Rhyme-based Acoustic Features for Mandarin Tone Recognition

We investigate the use in Mandarin tone recognition of over two hundred possible local acoustic features based on pitch, overall intensity, and band-passed intensity in the rhyme of a syllable. Features involving pitch height are not as useful as one might expect, showing the need for phrase-level pitch height correction. The intensity contour is useful, particularly when rhyme-initial intensity is subtracted. Intensity in certain medium and high-frequency bands also provides useful information. Unsurprisingly, contour tones are better recognized than level tones using only local features. In tonal languages, lexical information is carried both by phonemes and by syllable-specific intonation called tones. In the tonal language Mandarin, the five possible tones (high, rising, low, falling, neutral) carry as much information as vowels [1] [2]. Mandarin Tone Recognition is the problem of determining the tone of a syllable. Here, we assume that we know the syllable boundaries. Acoustic features for Mandarin Tone recognition can be found using duration, pitch, overall intensity, and intensities in various high-frequency bands [3]. However, there are several possible such features. Here we determine a useful subset of them that we can use in further experiments. For now, we deliberately stick to local features. Other than speaker-normalization, we will not consider features that use information outside the syllable boundary. Furthermore, we will limit ourselves to features computed on the rhyme of a syllable [4] to avoid the effect of syllable-initial consonants. Pitch and overall intensity measurements were found using Praat [5]. Band energy measurements were found using multi-taper spectral analysis [6] by considering overlapping 20ms chunks of speech every 5ms. 1 Features Considered The 221 local features we considered were the following. • duration : Duration of the rhyme in milliseconds. • # voiced : Number of voiced samples in the rhyme. • int(F) : mean, gradient, and intercept (all across the rhyme) of the contour energy between F−250 Hz and F+250 Hz, for F = 250, 500, . . ., 7500, 7750 Hz. There were 3× 31 = 93 such features. • We considered three acoustic measures, that we shall refer to as ‘cues’. Each was z-normalized by story before computing any features based on it.