Hidden conditional random field models for phonetic classification and speech recognition
AssigneeMicrosoft Corporation
InventorsAsela J. Gunawardana, Milind Mahajan, Alejandro Acero
Life of this patent
Estimating the term…
Abstract
A method and apparatus are provided for training and using a hidden conditional random field model for speech recognition and phonetic classification. The hidden conditional random field model uses feature functions, at least one of which is based on a hidden state in a phonetic unit. Values for the feature functions are determined from a segment of speech, and these values are used to identify a phonetic unit for the segment of speech.
Claims12
A computer-readable storage medium having computer-executable instructions stored thereon that when executed by a processor cause the processor to perform steps comprising:
- receiving a speech signal;
- determining a sequence of observed values from the speech signal;
- forming a separate constrained trellis for each of a plurality of phonetic units, each constrained trellis comprising hidden states at a plurality of time points;
- computing a recursive score for each constrained trellis, each recursive score based on an exponential function with e raised to a sum of weighted feature function values, wherein at least one of the feature function values is set equal to a value of a formant in the speech signal based on a determination that a current hidden state matches a particular hidden state;
- comparing the recursive scores for the constrained trellises to each other to identify the constrained trellis that provides the greatest score;
- and
- identifying the phonetic unit associated with the constrained trellis that provides the greatest score as the phonetic unit represented by the speech signal.
A method of decoding a speech signal to identify at least one phonetic unit, the method comprising:
- converting the speech signal into a sequence of observed values;
- forming a constrained trellis for each of a set of phonetic units, each constrained trellis comprising a lattice of hidden states extending across time points from an initial time point to a final time point;
- for each phonetic unit:
- setting an initial forward recursion score to one for each hidden state at the initial time point;
- a processor producing a forward recursion score for a final hidden state at the final time point by recursively computing a forward recursion score for each hidden state for each time point after the initial time point, each forward recursion score formed by determining a sum of summands taken over all states at a previous time point, each summand in the sum being a product of a forward recursion score at a hidden state at the previous time point and an exponential function with e raised to a function, where the function is a sum of weighted feature function values, where a feature function value for a squared element feature function at a first state at a time point comprises the square of an element of an observed value for the time point;
- comparing the forward recursion scores for the final hidden states at the final time points of each phonetic unit to identify the phonetic unit with the greatest forward recursion score;
- and
- selecting the phonetic unit with the greatest forward recursion score as the phonetic unit represented by the speech signal.
Dependencies are inferred from the claim text.