Speaker independent isolated word recognition system using neural networks

Claim 1

Speaker independent isolated word recognition system, where the speech signal is digitized and submitted to spectral analysis at constant temporal intervals using fast Fourier transform, the analysis result is submitted to an orthogonal transformation to obtain cepstral parameters and the logarithm of the total energy contained in each temporal interval is calculated , then the numerical derivatives of these values are calculated, obtaining the characteristic parameters of the speech signal for each temporal interval, the word ends are detected through the energy level of the signal and the word is analysed by a recogniser (RNA), in which complete words are modelled with Markov model automata of the left-to-right type with recursion on the states, each of which corresponds to an acoustic portion of the word, and in which the recognition is carried out through a dynamic programming according to Viterbi algorithm on all automata for finding the one with the minimum cost path, which corresponds to the recognized word indicated at output (PR), the emission possibilities being calculated with a neural network with feedback specially trained and the transition probabilities being estimated in a suitable way, the training method is characterized by the fact that it comprises the following operations: initialization: a.initialization of the neural network with small random synaptic weights; b.creation of the first segmentation by segmenting the training set words uniformly. iteration: 1. initialization of the training set with all the segmented words; 2. random choice of a word not already learnt (a word is considered learnt if the mean error for that word is sufficiently low); 3. updating of synaptic weights w ij for the considered word by applying a correlative training; more particularly the neural network input is made to vary according to a window sliding from left to right on the word and for every input window a suitable objective vector is supplied at the output, constructed by setting a 1 on the neuron corresponding to the state to which the input window belongs, according to the current segmentation, and by setting 0 on all the other neurons; 4. segmentation recomputation for the considered word, by using the neural network trained until now, and performing a dynamic programming only with the correct model; 5. updating of the current segmentation S t+1 ; 6. if there still are non considered words in the training set, go to step 2; 7. recomputation of transition probabilities of automata; 8. if the number of iterations on the training set is greater than a maximum preset number NMAX, terminate, or otherwise go to step 1.