Manner of Articulation based Split Lattices for Phoneme Recognition

Abstract

Phoneme lattices have been shown to be a good choice to encode in a compact way alternative decoding hypotheses from a speech recognition system. However the optimal phoneme sequence is produced by tracing all the phoneme identities in the lattice. This not only makes the search space of the decoder huge but also the final phoneme sequence may be prone to have false substitutions or insertion errors. In this paper, we introduce the split lattice structures that is generated by splitting the speech frames based on the manner of articulation. Spectral flatness measure (SFM) is exploited to detect the two broad manner of articulation sonorants and non-sonorants. The manner of sonorants includes broadly the vowels, the semivowels and the nasals whereas the fricatives, stop consonants and closures belong to non-sonorants. The conventional way of speech decoder produces one lattice for one test utterance. In our work, we split the speech frames into sonorants and non-sonorants based on SFM knowledge and generate split lattices. The split lattice generated are modified according to the manner of articulation in each split so as to remove the irrelevant phoneme identities in the lattice. For instance, the sonorant lattice is forced to exclude the non-sonorant phoneme identities and hence minimizing false substitutions or insertion errors. The proposed split lattice structure based on sonority detection decreased the phone error rates by nearly 0.9 % when evaluated on core TIMIT test corpus as compared to the conventional decoding involved in the state-of-the-art Deep Neural Networks (DNN).