Method of and a device for speech recognition employing neural network and markov model recognition techniques
AssigneeCSELT - Centro Studi e Laboratori Telecomunicazioni S.p.A.
InventorsLuciano Fissore, Roberto Gemello, Franco Ravera
Life of this patent
Estimating the term…
Abstract
A method and a device for recognition of isolated words in large vocabularies are described, wherein recognition is performed through two sequential steps using neural networks and Markov models techniques, respectively, and the results of both techniques are adequately combined so as to improve recognition accuracy. The devices performing the combination also provide an evaluation of recognition reliability.
Claims16
A method for speech recognition wherein:
- two recognition Steps are performed, one based on the use of neural networks, and the other on the use of hidden Markov models, each step supplying a respective list of recognition hypothesis, in which each hypothesis is associated with a respective acoustic likelihood score;
- the likelihood scores of each list are processed;
- and a single list reordered on the basis of the processed scores is provided, the two recognition steps operating sequentially on a common utterance to be recognized, in such a way that the neural network step analyses the entire active vocabulary and the Markov step analyses only a partial vocabulary represented by the list of hypotheses obtained as a result of the neural network step, and in that additionally an evaluation of recognition reliability in performed for the best hypothesis of the re-ordered list, on the basis of the scores resulting from said combination and associated with such best hypothesis and to one or more hypotheses which seize subsequent positions in the re-ordered list, thereby originating a reliability index that can take up at least two values, corresponding to the cases of recognition rated as “Certain” or “not certain”, respectively.
A speech recognizer, including:
- a pair of cascade-connected recognition units that make use of a recognition technique based on neural networks and a recognition technique based on hidden Markov models, respectively, and provide respective lists of recognition hypotheses, in which each hypothesis is associated with a respective acoustic likelihood score, and
- a processing unit (UE1), including combination means (UE1) for combining the scores determined by both recognition units (NN, MA) and for providing a re-ordered list based on the combined scores,
- characterised in that the neural network recognition unit located before the recognition unit based on hidden Markov models and is designed to perform recognition by operating on the entire active vocabulary, and the recognition unit based on hidden Markov models is designed to perform recognition independently of the neural network recognition unit by acting on a partial vocabulary formed by the hypotheses present in the list supplied by the neural network unit;
- and in that the processing unit includes evaluation means for evaluating recognition reliability of the hypothesis that has the best likelihood score within the re-ordered list, by using the combined scores associated with the hypotheses present in the re-ordered list, said evaluation means being capable of providing a reliability index that can take at least two values, corresponding to a recognition rated as “certain” or as “not certain”, respectively, for such hypothesis.
The recognizer as claimed in 13, wherein such combination means are enabled to carry out the mean and the variance computation and the normalisation of the scores of the lists provided by each recognition unit (NN, MA), only if such lists include a number of hypotheses not lower than a minimum.
Dependencies are inferred from the claim text.