Universal semi-word model for vocabulary contraction in automatic speech recognition

A non-transitory computer-readable storage medium storing instructions that when executed by a computer processor perform actions comprising:

• accessing a pronunciation dictionary that maps words in an associated vocabulary to corresponding sequences of phonemes, and an acoustic model representing sound characteristics of phonemes in a language associated with the pronunciation dictionary;
• for a given predetermined number of phonemes, segmenting the words of the pronunciation dictionary into (i) an initial sequence of the predetermined number of phonemes, and (ii) a reference to a semi-word model representing a sequence of remaining phonemes in whole words that follow the predetermined number of phonemes;
• creating a restructured pronunciation dictionary based on the segmented initial sequences of phonemes and references to semi-word models representing sequences of remaining phonemes, the restructured pronunciation dictionary containing references to a semi-word model for words having more than the predetermined number of phonemes;
• using a multi-level neural network to train a restructured acoustic model for the acoustic model based on the segmented initial sequences of phonemes and references to semi-word models;
• receiving a natural language utterance;
• segmenting the natural language utterance into words at least in part using the restructured acoustic model;
• and
• for at least one of the words:
• recognizing, using the restructured pronunciation dictionary, a first sequence of the predetermined number of phonemes in the word,
• determining, based on a presence of a reference to a semi-word model in the restructured pronunciation dictionary, that there are additional phonemes to be recognized in the word following the sequence of the predetermined number of phonemes, and
• responsive to determining that there are additional phonemes to be recognized in the word following the sequence of the predetermined number of phonemes, using at least one semi-word model for a remainder of the word comprising the additional phonemes to recognize text of the word.

A speech recognition system comprising:

• a computer processor;
• and
• a non-transitory computer-readable storage medium storing instructions that when executed by the computer processor perform actions comprising:
• accessing a pronunciation dictionary that maps words in an associated vocabulary to corresponding sequences of phonemes, and an acoustic model representing sound characteristics of phonemes in a language associated with the pronunciation dictionary;
• for a given predetermined number of phonemes, segmenting the words of the pronunciation dictionary into (i) an initial sequence of the predetermined number of phonemes, and (ii) a reference to a semi-word model representing a sequence of remaining phonemes in whole words that follow the predetermined number of phonemes;
• creating a restructured pronunciation dictionary based on the segmented initial sequences of phonemes and references to semi-word models representing sequences of remaining phonemes, the restructured pronunciation dictionary containing references to a semi-word model for words having more than the predetermined number of phonemes;
• using a multi-level neural network to train a restructured acoustic model for the acoustic model based on the segmented initial sequences of phonemes and references to semi-word models;
• receiving a natural language utterance;
• segmenting the natural language utterance into words at least in part using the restructured acoustic model;
• and
• for at least one of the words:
• recognizing, using the restructured pronunciation dictionary, a first sequence of the predetermined number of phonemes,
• determining, based on a presence of a reference to a semi-word model in the restructured pronunciation dictionary, that there are additional phonemes to be recognized in the word following the sequence of the predetermined number of phonemes, and
• responsive to determining that there are additional phonemes to be recognized in the word following the sequence of the predetermined number of phonemes, using at least one semi-word model for a part of the word to recognize text of the word.

A computer-implemented method of a speech recognition system for recognizing a natural language utterance, the method comprising:

• accessing a pronunciation dictionary that maps words in an associated vocabulary to corresponding sequences of phonemes, and an acoustic model representing sound characteristics of phonemes in a language associated with the pronunciation dictionary;
• for a given predetermined number of phonemes, segmenting the words of the pronunciation dictionary into (i) an initial sequence of the predetermined number of phonemes, and (ii) a reference to a semi-word model representing a sequence of remaining phonemes in whole words that follow the predetermined number of phonemes;
• creating a restructured pronunciation dictionary based on the segmented initial sequences of phonemes and references to semi-word models representing sequences of remaining phonemes, the restructured pronunciation dictionary containing references to a semi-word model for words having more than the predetermined number of phonemes;
• using a multi-level neural network to train a restructured acoustic model for the acoustic model based on the segmented initial sequences of phonemes and references to semi-word models;
• receiving a natural language utterance;
• segmenting the natural language utterance into words at least in part using the restructured acoustic model;
• and
• for at least one of the words:
• recognizing, using the restructured pronunciation dictionary, a first sequence of the predetermined number of phonemes within the word,
• determining, based on a presence of a reference to a semi-word model in the restructured pronunciation dictionary, that there are additional phonemes to be recognized in the word following the sequence of the predetermined number of phonemes, and
• responsive to determining that there are additional phonemes to be recognized in the word following the sequence of the predetermined number of phonemes, using at least one semi-word model for phonemes of the word, other than the phonemes of the first sequence, to recognize text of the word.