System for building a language model network for speech recognition
First Claim
Patent Images
1. A system for recognizing continuous speech configured so as to perform the following steps:
- to acquire an acoustic signal comprising words spoken by a speaker,to process the acoustic signal so as to generate a signal indicative of acoustic parameters present in the acoustic signal, andto decode the signal indicative of acoustic parameters so as to generate an output signal indicative of the words pronounced by the speaker, the decoding step comprising a step of comparing the signal indicative of a language and with a lexicon relating to the words spoken by the speaker, the language model being represented by means of a tree-like probabilistic network of finite states of the lexicon,wherein said network is constructed, in a preliminary stage, with the use of a linear interpolated language model to assign the probabilities to the network,wherein said language model is based on bigrammes, and wherein said system uses the following function to assign the respective probability to each bigramme;
##EQU1## PR(z/y) being the probability of a generic bigramme yz, γ
(y) being the total probability assigned to the bigrammes with zero frequency in the context y, Pr(z), the a priori probability of z, f'"'"'(z/y) being given by;
space="preserve" listing-type="equation">f'"'"'(z/y)=(1-γ
(y))f(z/y) f(z/y) being the relative frequency of the bigramme yz and c(y) being the number of occurrences of y in a sample acoustic signal.
2 Assignments
0 Petitions
Accused Products
Abstract
A system for recognizing continuous speech, for example for automatic dictation applications, uses a bigramme language model organized as a network with finite probability states. The system also uses methods of estimating the probabilities associated with the bigrammes and of representing the model of the language in a tree-like probability network.
50 Citations
17 Claims
-
1. A system for recognizing continuous speech configured so as to perform the following steps:
-
to acquire an acoustic signal comprising words spoken by a speaker, to process the acoustic signal so as to generate a signal indicative of acoustic parameters present in the acoustic signal, and to decode the signal indicative of acoustic parameters so as to generate an output signal indicative of the words pronounced by the speaker, the decoding step comprising a step of comparing the signal indicative of a language and with a lexicon relating to the words spoken by the speaker, the language model being represented by means of a tree-like probabilistic network of finite states of the lexicon, wherein said network is constructed, in a preliminary stage, with the use of a linear interpolated language model to assign the probabilities to the network, wherein said language model is based on bigrammes, and wherein said system uses the following function to assign the respective probability to each bigramme;
##EQU1## PR(z/y) being the probability of a generic bigramme yz, γ
(y) being the total probability assigned to the bigrammes with zero frequency in the context y, Pr(z), the a priori probability of z, f'"'"'(z/y) being given by;
space="preserve" listing-type="equation">f'"'"'(z/y)=(1-γ
(y))f(z/y)f(z/y) being the relative frequency of the bigramme yz and c(y) being the number of occurrences of y in a sample acoustic signal. - View Dependent Claims (2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15)
-
-
16. A system for recognizing continuous speech configured so as to perform the following steps,
acquisition of an acoustic signal comprising words spoken by a speaker in form of numeric samples, transformation of the sampled speech into a sequence of acoustic feature vectors, conveying spectral information; -
decoding of the feature sequence into a word sequence by employing a beam search based algorithm on a network of finite states that represents a linearly interpolated bigramme language model and embodies linguistic and lexical constraints; wherein the language model is estimated by means of a stacked interpolation algorithm and wherein the language model network of finite states is built by optimizing a tree-based network. - View Dependent Claims (17)
-
Specification