Sixth International Conference on Spoken Language Processing

This paper uses the method of mutual information to estimate the distribution of partial phonetic information in the timefrequency plane relative to an acoustic landmark. TIMIT transcriptions are parsed to estimate the locations of consonant closure landmarks, consonant release landmarks, manner change landmarks, and vowel or glide pivot landmarks. A melscale spectrogram is computed over the 250ms centered at each landmark, and the logarithmic energy of each point in timefrequency space is linearly quantized. The phoneme label associated with a landmark determines the values of 25 binary distinctive features. Finally, coincidences between feature and spectral energy values are counted, and the average log probabilities are calculated in order to produce an \infogram" of each distinctive feature: a measurement of the mutual information between the value of the feature and the energy of each point in the timefrequency plane.
Bibliographic reference. HasegawaJohnson, Mark (2000): "Timefrequency distribution of partial phonetic information measured using mutual information", In ICSLP2000, vol.4, 133136.