Abstract: Disclosed is a system and method of decomposing a lattice transition matrix into a block diagonal matrix. The method is applicable to automatic speech recognition but can be used in other contexts as well, such as parsing, named entity extraction and any other methods. The method normalizes the topology of any input graph according to a canonical form.

Abstract: The present invention successfully combines neural-net discriminative feature processing with Gaussian-mixture distribution modeling (GMM). By training one or more neural networks to generate subword probability posteriors, then using transformations of these estimates as the base features for a conventionally-trained Gaussian-mixture based system, substantial error rate reductions may be achieved. The present invention effectively has two acoustic models in tandem—first a neural net and then a GMM. By using a variety of combination schemes available for connectionist models, various systems based upon multiple features streams can be constructed with even greater error rate reductions.

Abstract: Predetermined macrosegments of the fundamental frequency are determined by a neural network, and these predefined macrosegments are reproduced by fundamental-frequency sequences stored in a database. The fundamental frequency is generated on the basis of a relatively large text section which is analyzed by the neural network. Microstructures from the database are received in the fundamental frequency. The fundamental frequency thus formed is thus optimized both with regard to its macrostructure and to its microstructure. As a result, an extremely natural sound is achieved.

Abstract: The invention relates to a method for selecting a sound algorithm for processing an audio signal. The audio signal is analyzed and the type of audio signal is ascertained based on the analysis. The audio signal is classified as a music signal or another signal, and different sound algorithms are used for the further processing and subsequent output of the audio signal.

Abstract: Methods for processing speech data are described herein. In one aspect of the invention, an exemplary method includes receiving a text sentence comprising a plurality of words, each of the plurality of words having a part of speech (POS) tag, generating a POS sequence based on the POS tag of each of the plurality of words, detecting a prosodic phrase break through a recurrent neural network (RNN), based on the POS sequence, and generating a prosodic phrases boundary based on the prosodic phrase break. Other methods and apparatuses are also described.

Abstract: A method and apparatus for efficient implementation and evaluation of state machines and programmable finite state automata is described. In one embodiment, a state machine architecture comprises a plurality of node elements, wherein each of the plurality of node elements represents a node of a control flow graph. The state machine architecture also comprises a plurality of interconnections to connect node elements, a plurality of state transition connectivity control logic to enable and disable connections within the plurality of interconnections to form the control flow graph with the plurality of node elements, and a plurality of state transition evaluation logic coupled to the interconnections and operable to evaluate input data against criteria, the plurality of state transition evaluation logic to control one or more state transitions between node elements in the control flow graph.

Abstract: A distributed voice recognition system and method for obtaining acoustic features and speech activity at multiple frequencies by extracting high frequency components thereof on a device, such as a subscriber station and transmitting them to a network server having multiple stream processing capability, including cepstral feature processing, MLP nonlinear transformation processing, and multiband temporal pattern architecture processing. The features received at the network server are processed using all three streams, wherein each of the three streams provide benefits not available in the other two, thereby enhancing feature interpretation. Feature extraction and feature interpretation may operate at multiple frequencies, including but not limited to 8 kHz, 11 kHz, and 16 kHz.

Abstract: Embodiments of the invention provide a programmable FSA building block, having a number of programmable registers and associated logic implemented therein, that provide the capability of contextually evaluating complex REs of arbitrary size against multiple data streams. Embodiments of the invention provide fully programmable hardware in which all of the states of an RE are instantiated and all of the states are fully connected. For one embodiment, the building blocks have a fixed number of states to facilitate implementation on a chip. For such an embodiment, an RE having an excessive number of states is implemented on two or more FSA building blocks and the FSA building blocks are then stitched together to effect evaluation of the RE. For one embodiment, two or more REs having a number of states less than the fixed number of states of a building block may be implemented with a single building block.

Abstract: A selection module allows a user to specify at least one measure to be monitored in at least one dimension of a dimensional hierarchy. A control limit calculator extracts, for each specified measure and for each specified dimension, a time series from a multidimensional database for the specified measure in the specified dimension and automatically calculates one or more control limits for the specified measure in the specified dimension based on the extracted time series using a Statistical Process Control (SPC) technique. Thereafter, a monitoring module monitors newly acquired data including each specified measure in each specified dimension for an out-of-limits condition based on one or more automatically-calculated control limits. An alert module triggers an alert in response to an out-of-limits condition being detected.

Abstract: A method and apparatus are disclosed for transcribing speech when a number of speakers are participating. A number of different speech recognition systems, each with a different speaker model, are executed in parallel. When the identity of all of the participating speakers is known and a speaker model is available for each participant, each speech recognition system employs a different speaker model suitable for a corresponding participant. Each speech recognition system decodes the speech and generates a corresponding confidence score. The decoded output having the highest confidence score is selected for presentation to a user. When all participating speakers are not known, or when there are too many participants to implement a unique speaker model for each participant, a speaker independent speech recognition system is employed together with a speaker specific speech recognition system.

Abstract: A system, method and computer readable medium for quantizing class information and pitch information of audio is disclosed. The method on an information processing system includes receiving audio and capturing a frame of the audio. The method further includes determining a pitch of the frame and calculating a codeword representing the pitch of the frame, wherein a first codeword value indicates an indefinite pitch. The method further includes determining a class of the frame, wherein the class is any one of at least two classes indicating an indefinite pitch and at least one class indicating a definite pitch. The method further includes calculating a codeword representing the class of the frame, wherein the codeword length is the maximum of the minimum number of bits required to represent the at least two classes and the minimum number of bits required to represent the at least one class.

Abstract: A method and system are provided for speech recognition. The speech recognition method includes the steps of preparing training data representing acoustic parameters of each of phonemes at each time frame; receiving an input signal representing a sound to be recognized and converting the input signal to input data; comparing the input data at each frame with the training data of each of the phonemes to derive a similarity measure of the input data with respect to each of the phonemes; and processing the similarity measures obtained in the comparing step using a neural net model governing development of activities of plural cells to conduct speech recognition of the input signal.

Abstract: A speech recognition system that is insensitive to external noise and applicable to actual life includes an A/D converter that converts analog voice signals to digital signals. An FIR filtering section employs powers-of-two conversion to filter the digital signals converted at the A/D converter into numbers of channels. A characteristic extraction section immediately extracts speech characteristics having strong noise-resistance from the output signals of the FIR filtering section without using additional memories. A word boundary detection section discriminates the information of the start-point and the end-point of a voice signal on the basis of the characteristics extracted by the characteristic extraction section.

Abstract: A voice-enabled system for online shopping provides a voice and telephony interface, as well a text and graphic interface, for shopping over the Internet using a browser or a telephone. The system allows customers to access an online shop, search for desired database items, select items, and finally pay for selected items using a credit card, over a phone line or the Internet. A telephony-Internet interface converts spoken queries into electronic commands for transmission to an online shop or database. Markup language-type pages transmitted to callers from the online-shop or database are parsed to extract selected information. The selected information is then reported to the callers via audio messaging.

Abstract: A method is developed which includes 1) defining a switching state space model for a continuous valued hidden production-related parameter and the observed speech acoustics, and 2) approximating a posterior probability that provides the likelihood of a sequence of the hidden production-related parameters and a sequence of speech units based on a sequence of observed input values. In approximating the posterior probability, the boundaries of the speech units are not fixed but are optimally determined. Under one embodiment, a mixture of Gaussian approximation is used. In another embodiment, an HMM posterior approximation is used.

Abstract: The invention relates to a method for speech processing in which input variables containing speech features are mapped onto output variables. In the mapping process, the input variables are weighted and/or identical maps are produced for different sets of input variables and at least one output variable.

Abstract: A method is described for compressing the storage space required by HMM prototypes in an electronic memory. For this purpose prescribed HMM prototypes are mapped onto compressed HMM prototypes with the aid of a neural network (encoder). These can be stored with a smaller storage space than the uncompressed HMM prototypes. A second neural network (decoder) serves to reconstruct the HMM prototypes.

Abstract: An apparatus and method for selecting the length of a variable length code bitstream by using a neural network are provided. The apparatus for selecting the length of a variable length code bitstream includes a bitstream estimation length receiving unit which inputs a predetermined quantization DCT coefficient block to a neural network whose training is finished, and receives the estimation length of a bitstream corresponding to the quantization DCT coefficient block from the neural network; and a bitstream estimation length selection unit which receives user selection about an estimation length received by the bitstream estimation length receiving unit. According to the method and apparatus the length of a variable length code bit stream can be estimated such that a user can select a desired length of a bitstream in advance without performing variable length coding.

Abstract: Method of suppressing audible noise in speech transmission by means of a multi-layer self-organizing fed-back neural network comprising a minima detection layer, a reaction layer, a diffusion layer and an integration layer, said layers defining a filter function F(f,T) for noise filtering.

Abstract: The invention relates to a method for recognizing a phonetic sound sequence or a character sequence, e.g.

Abstract: A spatial image processor neural network for processing image data to discriminate between first and second spatial configurations of component objects includes a photo transducer input array for converting an input image to pixel data and sending the data to a localized gain network (LGN) module, a parallel memory processor and neuron array for receiving the pixel data and processing the pixel data into component recognition vectors and chaotic oscillators for receiving the recognition vectors and sending feedback data to the LGN module as attention activations. The network further includes a temporal spatial retina for receiving both the pixel data and temporal feedback activations and generating temporal spatial vectors, which are processed by a temporal parallel processor into temporal component recognition vectors. A spatial recognition vector array receives the temporal component recognition vectors and forms an object representation of the first configuration of component objects.

Abstract: A method of determining the topology of a network comprising: transmitting a signal comprised of a sequence of bursts of packets formed of orthogonal signals, monitoring devices in the network including the destination device for reception of the signal, and defining a sequence of devices within the network by sensing a sequence of reception of the signal in the devices from the source device toward the destination device.

Abstract: The present invention provides for a method and system of voice recognition, in particular for navigation in a hypertext navigation system. For each new word, a language identification stage, in particular embodied as a neural network, is used to determine the inclusion of the word in a language or a dialect with a given probability factor and the grapheme/phoneme relationship corresponding to the word with the greatest probability coefficient in the phonetic lexicon, or in at least one of the several phonetic lexica, is updated.

Abstract: A method and apparatus are described that allow inexpensive speech recognition in applications where this capability is not otherwise feasible because of cost or technical reasons, or because of inconvenience to the user. A relatively simple speaker independent recognition algorithm, capable of recognizing a limited number of utterances at any one time, is associated with the base unit of an electronics product. To function, the product requires information from an external medium and this medium also provides the data required to recognize several sets of utterances pertinent to other information provided by the external medium.

Abstract: The object of the present invention is to keep a high success rate in recognition with a low-volume of sound signal, without being affected by noise.

Abstract: The model adaptation system of the present invention is a speaker verification system that embodies the capability to adapt models learned during the enrollment component to track aging of a user's voice. The system has the advantage of only requiring a single enrollment for the user. The model adaptation system and methods can be applied to several types of speaker recognition models including neural tree networks (NTN), Gaussian Mixture Models (GMMs), and dynamic time warping (DTW) or to multiple models (i.e., combinations of NTNs, GMMs and DTW). Moreover, the present invention can be applied to text-dependent or text-independent systems.

Abstract: The present invention is embodied in a system and method for recognizing speech and transcribing speech in real time. The system includes a computer, which could be in a LAN or WAN linked to other computer systems through the Internet. The computer has a controller, or similar device, to filter background noise and convert incoming signals to digital format. The digital signals are transcribed to a word list, which is processed by an automatic speech recognition system. This system synchronizes and compares the lists and forwards the list to a speech recognition learning system, which stores the data on-site. The stored data is forwarded to an off-site storage system, and an off-site large scale learning system that processes the data from all sites on the wide area network system.

Abstract: A method and system for objectively evaluating the quality of speech in a voice communication system. A plurality of speech reference vectors is first obtained based on a plurality of clean speech samples. A corrupted speech signal is received and processed to determine a plurality of distortions derived from a plurality of distortion measures based on the plurality of speech reference vectors. The plurality of distortions are processed by a non-linear neural network model to generate a subjective score representing user acceptance of the corrupted speech signal. The non-linear neural network model is first trained on clean speech samples as well as corrupted speech samples through the use of backpropagation to obtain the weights and bias terms necessary to predict subjective scores from several objective measures.

Abstract: The invention relates to a method of controlling function units of a motorcar or of devices (1a, 1b) installed in a motorcar, via speech signals, in which

Abstract: A method and system for recognizing user input information including cursive handwriting and spoken words. A time-delayed neural network having an improved architecture is trained at the word level with an improved method, which, along with preprocessing improvements, results in a recognizer with greater recognition accuracy. Preprocessing is performed on the input data and, for example, may include resampling the data with sample points based on the second derivative to focus the recognizer on areas of the input data where the slope change per time is greatest. The input data is segmented, featurized and fed to the time-delayed neural network which outputs a matrix of character scores per segment. The neural network architecture outputs a separate score for the start and the continuation of a character.

Abstract: A speech recognition system utilizes both matrix and vector quantizers as front ends to a second stage speech classifier such as hidden Markov models (HMMs) and utilizes neural network postprocessing to, for example, improve speech recognition performance. Matrix quantization exploits the “evolution” of the speech short-term spectral envelopes as well as frequency domain information, and vector quantization (VQ) primarily operates on frequency domain information. Time domain information may be substantially limited which may introduce error into the matrix quantization, and the VQ may provide error compensation. The matrix and vector quantizers may split spectral subbands to target selected frequencies for enhanced processing and may use fuzzy associations to develop fuzzy observation sequence data. A mixer provides a variety of input data to the neural network for classification determination. The neural network's ability to analyze the input data generally enhances recognition accuracy.

Abstract: A method of organizing an acoustic model for speech recognition is comprised of the steps of calculating a measure of acoustic dissimilarity of subphonetic units. A clustering technique is recursively applied to the subphonetic units based on the calculated measure of acoustic dissimilarity to automatically generate a hierarchically arranged model. Each application of the clustering technique produces another level of the hierarchy with the levels progressing from the least specific to the most specific. A technique for adapting the structure and size of a trained acoustic model to an unseen domain using only a small amount of adaptation data is also disclosed.

Abstract: The presence of a voice in an audio signal is detected by sampling frequency components of the audio signal during a window that starts when a power of the audio signal reaches a predetermined threshold and stops when the audio signal's power drops below the predetermined threshold. An array of elements is generated based on the sampled frequency components. Each element in the array corresponds to a time-based sum of frequency components. Whether the audio signal corresponds to a voice is determined using one or values calculated from the generated array. The value may correspond either to a frequency-based sum of array elements or to the window. The calculated values are analyzed using fuzzy logic which generates a measure of a likelihood that the audio signal is a voice.

Abstract: A method for testing an audio device with a trained neural network includes a loopback connector connecting the output port of the audio device to the input port of the audio device. A test signal is transmitted through the audio port and received at the input port. The test signal is converted into a frequency spectrum for analysis. The frequency spectrum is provided as input to a trained neural network, the neural network being previously trained to recognize the frequency spectrum pattern created by a properly working, or ideal, audio device. The neural network is trained by connecting the input port to the output port of an audio device from which the training is to occur. Prior to converting signals to a frequency spectrum, the waveform characteristics of the signal may be iteratively evaluated and recording levels adjusted so that the signal received has characteristics that can be tested by the neural network.

Abstract: A method for recognizing user specified pen-based gestures uses Hidden Markov Models. A gesture recognizer is implemented which includes a fast pruning procedure. In addition, an incremental training method is utilized.

Abstract: A method for recognizing a speaker in which a voice signal is spoken into a computer by a speaker and a feature vector is formed for the voice signal. The feature vector is compared to at least one stored reference feature vector and to at least one anti-feature vector. The reference feature vector is formed from a speech sample of a speaker to be verified. The anti-feature vector was formed from a speech sample that was spoken in by another speaker who is not the speaker to be verified. A 2-class classification is resolved by forming a similarity value and evaluating the similarity value on the basis of a predetermined range within which the similarity value must deviate from a predetermined value so that the voice signal can be classified as deriving from the speaker to be verified.

Abstract: An approximate weighted finite-state automaton can be constructed in place of a weighted finite-state automaton so long as the approximate weighted finite-state automaton maintains a sufficient portion of the original best strings in the weighted finite-state automaton and sufficiently few spurious strings are introduced into the approximate weighted finite-state automaton compared to the weighted finite-state automaton. An approximate weighted finite-state automaton can be created from a non-deterministic weighted finite-state automaton during determinization by discarding the requirement that old states be used in place of new states only when an old state is identical to a new state. Instead, in an approximate weighted finite-state automaton, old states will be used in place of new states when each of the remainders of the new state is sufficiently close to the corresponding remainder of the old state.

Abstract: A device for analysis and filtration of sound which comprises at least one frequency-linear filter, at least one frequency-logarithmic filter and a weighting means for the combining and non-linear weighting of the output signals from the frequency-linear filter and the frequency-logarithmic filter. The sound is fed parallel as an input signal to the two sets of filters. In turn, the output signals from the filters are fed to the weighting means where they are combined and weighted non-linearly on the basis of magnitudes relevant to the sound. A decision with respect to the identity of the sound is made in a decision means. The invention also relates to a method for analyzing and filtering sound with the aid of the above-mentioned device.

Abstract: In an information processing system such as a navigation system or a portable information terminal unit, information output mode of a display unit to a user is switchable to a plurality of output modes. The navigation system switches the language to be used for information output to any one of Japanese language, English language, and German language. Not the input speech itself is used as a command, but whether the input speech is a Japanese language input, an English language input, or a German language input is determined. The language to be used for information output is switched depending on the determination result.

Abstract: A method and apparatus is provided for matching a first sequence of patterns representative of a first signal with a second sequence of patterns representative of a second signal. The system uses a plurality of different pruning thresholds (th) to control the propagation of paths which represent possible matchings between a sequence of second signal patterns and a sequence of first signal patterns ending at the current first signal pattern. In particular, the pruning threshold used for a given path during the processing of a current first signal pattern depends upon the position, within the sequence of patterns representing the second signal, of the second signal pattern which is at the end of the given path.

Abstract: A method for automatically generating a pronunciation dictionary in a speech recognition system is disclosed. Pronunciation patterns of a large scale pronunciation dictionary are learned through a neural network without resorting to a phonetic knowledge, and the pronunciation sequences for input words are accurately formed by utilizing an exception grapheme pronunciation dictionary, an exception word pronunciation dictionary for graphemes and words prohibiting the formation of an accurate pronunciation dictionary through the learning neural network, thereby reducing the size of the memory and the amount of calculations. A multi-layer perceptron for directly mapping phonemes relevant to respective graphemes is taught by utilizing a neural network, so as to form an exception word pronunciation dictionary data base, an exception grapheme pronunciation dictionary data base, and a phoneme output multi-layer perceptron parameter data base for respective graphemes.

Abstract: A method and apparatus for signal classification using a multilayer temporal relaxation network involves receiving an input signal feature vector, classifying a first signal feature, and classifying a second signal feature using contextual information. The multilayer temporal relaxation network applies a relaxation process that updates an activation value of a node in a first layer and updates an activation value of a node in a second layer. The multilayer network then generates a signal classification according to an activation value of a node in the multilayer network.

Abstract: An improved method and system for training and classifying using a low complexity and high accuracy multiresolutional polynomial classifier (412) is presented. A method of training an multiresolutional polynomial classifier which reduces the complexity of existing classifiers allows models representing subgroups of classes to easily be created. The models which represent subgroups of classes are applied to an unidentified input to produce a coarse classification of the unidentified input using a low order classifier. Once a coarse classification of the unidentified input is performed, a more detailed classification is performed using another low complexity classifier.

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.

Abstract: A method (900), device (200) and system (100) provide, in response to text/linguistic input, one of a set of pre-determined meanings which is the most likely intended meaning of that input. A trained meaning discriminator is generated from an annotated training corpus and a meaning discriminator trainer. The trained meaning discriminator generates a meaning vector from an input utterance. The intended meaning encoder analyzes the meaning vector to determine the most likely intended meaning and confidence measures.

Abstract: A signal processing arrangement for a band-limited input signal, comprising a plurality N of signal comparators. Each signal comparator is adapted to compare the input signal with a plurality of different exemplar signals and to generate an output indicative of which of the exemplar signals corresponds most closely to the input signal. Each of the exemplar signals is arbitrarily derived independent of any expected input signal. The arrangement provides an N-part output signal which is indicative of the input signal, such that each part of the N-part output signal is derived from the output signal of a respective one of said N signal comparators.

Abstract: A recognition apparatus and method using a neural network is provided. A neuron-like element stores a value of its inner condition. The neuron-like element also updates a values of its internal status on the basis of an output from the neuron-like element itself, outputs from other neuron-like elements and an external input, and an output value generator a value of its internal status into an external output. Accordingly, the neuron-like element itself can retain the history of input data. This enables the time series data, such as speech, to be processed without providing any special devices in the neural network.

Abstract: Classifiers (110) and a comparator (112) perform an identification method (400) to identify a class as one of a predetermined set of classes. The identification method is based on determining the observation costs associated with the unidentified class. The identification method includes combining models representing the predetermined set of classes and the unidentified vectors representing the class. The predetermined class associated with the largest observation cost is identified as the class. Additionally, a unique, low-complexity training method (300) includes creating the models which represent the predetermined set of classes.

Abstract: A multiple confidence measures subsystem of an automated speech recognition system allows otherwise independent confidence measures to be integrated and used for both training and testing on a consistent basis. Speech to be recognized is input to a speech recognizer and a recognition verifier of the multiple confidence measures subsystem. The speech recognizer generates one or more confidence measures. The speech recognizer preferably generates a misclassification error (MCE) distance as one of the confidence measures. The recognized speech output by the speech recognizer is input to the recognition verifier, which outputs one or more confidence measures. The recognition verifier preferably outputs a misverification error (MVE) distance as one of the confidence measures. The confidence measures output by the speech recognizer and the recognition verifier are normalized and then input to an integrator.

Abstract: Training apparatus and method for establishing the network definition function of a trainable processing apparatus for analyzing a signal, includes providing a training sequence having a first signal and a distorted version of the first signal, receiving the training sequence and generating a distortion perception measure for indicating the extent to which the distortion would be perceptible to a human observer, and applying the distortion perception measure to the trainable processing apparatus to determine the network definition function.