Subword unit posterior probability for measuring confidence
AssigneeMicrosoft Corporation
InventorsPeng Liu, Ye Tian, Jian-Lai Zhou, Frank Kao-Ping K. Soong
Life of this patent
Estimating the term…
Abstract
Speech recognition such as command and control speech recognition generally use a context free grammar to constrain the decoding process. Word or subword background model are constructed to repopulate dynamic hypothesis space, especially when word spareness is at issue. The background models can be later used in speech recognition. During speech recognition, background and conventional context free grammar decoding are used to measure confidence. The discussion above is merely provided for general background information and is not intended to be used as an aid in determining the scope of the claimed subject matter.
Claims10
A computer-readable storage medium including instructions readable by a computer which, when implemented comprise:
- receiving a speech-based signal;
- generating a decoder result comprising at least one sequence of linguistic units based on applying a context free grammar to the speech-based signal;
- decomposing the decoder result into a decomposed decoder result comprising at least one sequence of subword elements;
- accessing a background model having probabilities of subword elements in one or more natural language;
- generating competitive sequences using the decomposed decoder result and the background model;
- and
- calculating a confidence measure using the competitive sequences and the decomposed decoding result by calculating a model based posterior probability score using the decomposed decoding result and the competitive sequences.
A method for performing speech recognition comprising:
- receiving a speech based signal based on words spoken in one or more natural languages;
- applying a context free grammar to the speech based signal with a decoder implemented by a processor to generate a decoding result that identifies speech recognized by the decoder in the speech based signal;
- decomposing the decoding result into subword linguistic units consistent with subword units of a background model to obtain a decomposed decoding result;
- generating competitive sequences using the decomposed decoding result and the background model, the competitive sequences being sequences of alternative subword linguistic units that are alternatives to the decomposed decoding result, the background model comprising probabilities of subword units in a dictionary of the one or more natural languages;
- and
- calculating a model based posterior probability score for a given combination of subword linguistic units by considering both the decomposed decoding result and the competitive sequences.
A system for performing speech recognition comprising:
- a processor;
- a decoder or decoding module executable on the processor comprising:
- a decoder or module that generates a decoder result comprising a sequence of words;
- a decomposition module that decomposes the sequence of words into a decomposed decoder result comprising a sequence of subword linguistic units;
- a background decoder or module that uses a background model of subword linguistic units to generate competitive sequences of subword linguistic units, based on the decoder result;
- a module that calculates a model based posterior probability score for at least one given path of subword units through a network of subword units that comprises using the decomposed decoding result and the competitive sequences;
- and
- an output module executable on the processor that provides an indication of correctness of the decoder result, the indication of correctness being based on the model based posterior probability score and being used as a rejection threshold for rejecting hypothesis decoder results to identify a correct decoder result.
Dependencies are inferred from the claim text.