Refining of segmental boundaries in speech waveforms using contextual-dependent models
AssigneeMicrosoft Corporation
InventorsYong Zhao, Min Chu, Jian-lai Zhou, Lijuan Wang
Life of this patent
Estimating the term…
Abstract
A method and apparatus are provided for refining segmental boundaries in speech waveforms. Contextual acoustic feature similarities are used as a basis for clustering adjacent phoneme speech units, where each adjacent pair phoneme speech units include a segmental boundary. A refining model is trained for each cluster and used to refine boundaries of contextual phoneme speech units forming the clusters.
Claims34
A method of ascertaining phoneme speech unit boundaries of adjacent speech units in speech data, the method comprising:
- receiving training data of speech waveforms with known boundary locations of phoneme speech units contained therein;
- processing the speech waveforms to obtain multi-frame acoustic feature pseudo-triphone representations of a plurality of pseudo-triphones in the speech data, each pseudo-triphone comprising a boundary location, a first phoneme speech unit preceding the boundary location and a second phoneme speech unit following the boundary location;
- clustering the multi-frame acoustic feature pseudo-triphone representations as a function of acoustic similarity in a plurality of clusters;
- training a refining model for each cluster;
- receiving a second set of data of speech waveforms with initial boundary locations of adjacent phoneme speech units contained therein;
- identifying pseudo-triphones in the second set of data and corresponding refining models for each of the pseudo-triphones;
- and
- using the refining model for each corresponding pseudo-triphone for the second set of data to locate a new boundary location different than the initial boundary and provide output indicating the new boundary locations.
A computer-readable storage medium having computer-executable instructions for processing speech data, the computer-readable medium comprising:
- an acoustic feature generator adapted to receive training data of speech waveforms with known boundary locations of phoneme speech units contained therein and generate multi-frame acoustic feature pseudo-triphone representations of a plurality of pseudo-triphones in the training data, each pseudo-triphone comprising a boundary location, a first phoneme speech unit preceding the boundary location and a second phoneme speech unit following the boundary location;
- a clustering module adapted to receive the multi-frame acoustic feature pseudo-triphone representations of the plurality of pseudo-triphones and cluster the representations based on acoustic similarity;
- and
- a refining module generator adapted to operate on each cluster of representations and generate a statistical model therefor indicative of the location of the boundary for each cluster.
Dependencies are inferred from the claim text.