Abstract: A method of adapting a neural network of an automatic speech recognition device, includes the steps of: providing a neural network including an input stage, an intermediate stage and an output stage, the output stage outputting phoneme probabilities; providing a linear stage in the neural network; and training the linear stage by means of an adaptation set; wherein the step of providing the linear stage includes the step of providing the linear stage after the intermediate stage.

Abstract: A Chinese to English Translation Tool (CETT) locates pre-existing translations of phrases and increases probabilities for a match by searching independently of proper nouns and Western characters. The CETT comprises an enhancement program, a search program and a display program. The CETT is adapted to access a translation database, a proper noun table, and an unmatched translation table. The translation database is a central repository of an accumulated body of translated literature. The proper noun database is a table showing the associations between English and Chinese for proper nouns.

Abstract: Methods, systems, and apparatus, including computer program products, to identify topic words in a document corpus that includes topic documents related to a topic are disclosed. A reference topic word divergence value based on the document corpus and the topic document corpus is determined. A candidate topic word divergence value for a candidate topic word is determined based on the document corpus and the topic document corpus. The candidate topic word is determined to be a topic word if the candidate topic word divergence value is greater than the reference topic word divergence value.

Abstract: A speech recognition device controller is provided, which is capable of properly determining a user's leaning level in speech so as to provide an efficient interaction. It includes a voice input unit 2 for receiving speech input from a user, a speech recognition unit 3 for recognizing the input speech, a response output unit 10 for outputting a response to the user, and a device control unit 11 for controlling a device based on a recognition result of the speech recognition unit 3.

Abstract: Embodiments of the present invention can gather data from native language sources to produce a valid collation sequence that is appropriate for a particular language and application. Sequences of characters in this data are tested to determine strength levels used by the given language. The data is also recursively probed with other sequences to test for contractions and identify expansions. Sequences in the data may then be compared against a known or predetermined sequence to generate a set of sorting rules that is specific to the language and application. The rules are formatted to replicate the sorting order found in the data.

Abstract: A method and apparatus is disclosed herein for a quantizing parameters using partial information on atypical subsequences. In one embodiment, the method comprises partially classifying a first plurality of subsequences in a target vector into a number of selected groups, creating a refined fidelity criterion for each subsequence of the first plurality of subsequences based on information derived from classification, dividing a target vector into a second plurality of subsequences, and encoding the second plurality of subsequences, including quantizing the second plurality of subsequences given the refined fidelity criterion.

Abstract: A method for automatic speech recognition includes determining for an input signal a plurality scores representative of certainties that the input signal is associated with corresponding states of a speech recognition model, using the speech recognition model and the determined scores to compute an average signal, computing a difference value representative of a difference between the input signal and the average signal, and processing the input signal in accordance with the difference value.

Abstract: A system and method for generating synthetic speech, which operates in a computer implemented Text-To-Speech system. The system comprises at least a speaker database that has been previously created from user recordings, a Front-End system to receive an input text and a Text-To-Speech engine. The Front-End system generates multiple phonetic transcriptions for each word of the input text, and the TTS engine uses a cost function to select which phonetic transcription is the more appropriate for searching the speech segments within the speaker database to be concatenated and synthesized.

Abstract: A method of forming a shareable filler model (shareable model for garbage words) from a word n-gram model is provided. The word n-gram model is converted into a probabilistic context free grammar (PCFG). The PCFG is modified into a substantially application-independent PCFG, which constitutes the shareable filler model.

Abstract: Methods are disclosed for automatic accent labeling without manually labeled data. The methods are designed to exploit accent distribution between function and content words.

Abstract: A method for removing periodic noise pulses from a continuous audio signal generated in a pressurized air delivery system includes the steps of: detecting, in a time-windowed segment of the continuous audio signal generated in the pressurized air delivery system, a plurality of the periodic noise pulses having a pulse period and being representable in the form of a plurality of signal components combined by convolution; deconvolving the plurality of signal components to generate a plurality of deconvolved signal components; and removing at least a portion of the periodic noise pulses from the time-windowed segment of the continuous audio signal using the deconvolved signal components.

Abstract: A method for text alignment of a first document and a second document that is a translation version of the first document. The method first divides paragraphs of the first and second documents into sub-sentential segments according to the punctuations in the language of the first and second documents. Each of sub-sentential segments corresponds to a plurality of words. After the sub-sentential segmenting process, pairs of alignment units are summarized from the first and second documents. The alignment units in the first and second documents are then aligned and scored mainly based on the probability of corresponding punctuations. To increase the alignment accuracy, the pairs of alignment units can also be aligned and scored based on at least one of length corresponding probability, match type probability, and lexical information. The method allows for fast, reliable, and robust alignment of document and translated document in two disparate languages.

Abstract: A method for automatic code generation that includes obtaining a parse tree, wherein the parse tree represents at least one instruction and wherein the instruction is in a natural language, and wherein the at least one instruction is free-form, determining a plurality of sentence subparts from the parse tree, creating at least one instantiated concept for each of the plurality of sentence subparts, inserting the instantiated concept in an instantiated concept map, traversing the instantiated concept map, wherein traversing the instantiated concept map includes for each instantiated concept in the instantiated concept map obtaining a code output rule according to the instantiated concept, and outputting code according to the code output rule be completed upon finalization of claims.

Abstract: A system for multi-lingual speech recognition. The inventive system includes a speech modeling engine, a speech search engine, and a decision reaction engine. The speech modeling engine receives and transfers a mixed multi-lingual speech signal into speech features. The speech search engine locates and compares candidate data sets. The decision reaction engine selects resulting speech models from the candidate speech models and generates a speech command.

Abstract: A system for implementing voice services can include at least one virtual machine, such as a Java 2 Enterprise Edition (J2EE) virtual machine. The virtual machine can include a bean container for handling software beans, such as Enterprise Java Beans. The bean container can include a voice browser bean. The voice browser bean can include a VoiceXML browser.

Abstract: The present invention discloses a method for a voice-enabled computing environment. The method includes a step of receiving a series of voice commands related to an abstraction, wherein the voice command specifies an abstraction type. The method further includes responding to the voice command to replay the commands, executing at least one programmatic action related to the abstraction, where the programmatic action is executed demonstrating characteristics specific to the abstraction type associated with the voice command. The present method applies, but is not limited to, the execution of collections of commands that may have diverging timing characteristics.

Abstract: A method of reducing noise in an audio signal, comprising the steps of: using a furrow filter to select spectral components that are narrow in frequency but relatively broad in time; using a bar filter to select spectral components that are broad in frequency but relatively narrow in time; analyzing the relative energy distribution between the output of the furrow and bar filters to determine the optimal proportion of spectral components for the output signal; and reconstructing the audio signal to generate the output signal. A second pair of time-frequency filters may be used to further improve intelligibility of the output signal. The temporal relationship between the furrow filter output and the bar filter output may be monitored so that the fricative components are allowed primarily at boundaries between intervals with no voiced signal present and intervals with voice components. A noise reduction system for an audio signal.

Abstract: A method for evaluating the processing delay of a speech signal contained in data packets received in a receiver terminal having a telephony module during a voice call to a terminal sending the data packets over a packet-switched network. The method includes the step of obtaining from the received data packets a stream of audio packets containing the speech signal. Within a predetermined decoding time, the stream of obtained audio packets is decoded and a first reconstituted speech signal is created. At least a portion of the speech reconstituted by the telephony module is duplicated to create a second reconstituted speech signal. The time difference between the first and the second reconstituted speech signals is determined. The processing delay of the speech signal in the receiver terminal is calculated from at least the determined time difference between the reconstituted first and second speech signals and the predetermined decoding time.

Abstract: Provides subtitle generation methods and apparatus which recognizes voice in a presentation to generate subtitles thereof, and retrieval apparatus for retrieving character strings by use of the subtitles. An apparatus of the present invention includes: a extraction unit for extracting text from presentation documents; an analysis unit for morphologically analyzing text to decompose it into words; a generation unit for generating common keywords by assigning weights to words; a registration unit for adding common keywords to a voice recognition dictionary; a recognition unit for recognizing voice in a presentation; a record unit for recording the correspondence between page and time by detecting page switching events; a regeneration unit for regenerating common keywords by further referring to the correspondence between page and time; a control unit for controlling the display of subtitles, common keywords, text and master subtitles; and a note generation unit for generating speaker notes from subtitles.

Abstract: A method is provided for identifying phrase alignment pairs between a source sentence and a target sentence. Boundaries for a phrase in the source sentence are identified by requiring that a source word be aligned with at least one target word in a target sentence in order to form a boundary for the source phrase. Boundaries for a phrase in the target sentence are identified based on alignments between words in the source phrase and words in the target sentence. The words in the target phrase are examined to determine if any of the words are aligned with source words outside of the source phrase. If they are not aligned with source words outside of the source phrase, the source phrase and target phrase are determined to form an alignment pair and are stored as a phrase alignment pair.