Out of vocabulary pattern learning
A method for adapting a speech recognition system for out-of-vocabulary, comprising, decoding by a hybrid speech recognition a speech including out-of-vocabulary terms, thereby generating graphemic transcriptions of the speech with a mixture of recognized in-vocabulary words and unrecognized sub-words, while keeping a track of the decoded segments of the speech, determining in the transcription sequences of sub-words as candidate out-of-vocabulary words based on a first condition with respect to lengths of the sequences of sub-words and a second condition with respect to the number of repetitions of the sequences, audibly presenting to a user the candidate out-of-vocabulary words from the corresponding segments of the speech according to the track, and receiving from the user indications of valid words corresponding to audible presentations of the sequences of sub-words in the candidate out-of-vocabulary words, and training a speech recognition to additionally recognize the candidate out-of-vocabulary words, thereby adapting the speech recognition to recognize out-of-vocabulary words, wherein the method is performed on an at least one computerized apparatus configured to perform the method, and an apparatus for performing the same.
1. A method for adapting a speech recognition system for out-of-vocabulary words, comprising:
decoding by a hybrid speech recognition a speech including out-of-vocabulary terms, thereby generating graphemic transcriptions of the speech with a mixture of recognized in-vocabulary words and unrecognized sub-words, while keeping a time track of the decoded segments of the speech;
converting the sub-words to patterns comprising a set of phoneme sequences by a process of concatenation of a phoneme representation of each sub-word;
subsequently, determining among the patterns, which patterns are candidate to represent out-of-vocabulary words based on a first condition with respect to the lengths of the pattern based on a number of phonemes and a second condition with respect to the number of repetitions of the pattern;
audibly presenting to a user the candidate patterns representing out-of-vocabulary words from the corresponding segments of the speech according to the time track, and receiving from the user indications of valid out-of-vocabulary words responsive to the audible presentations of the candidate patterns; and
training a speech recognition system to additionally recognize the identified out-of-vocabulary words, thereby adapting the speech recognition to recognize out-of-vocabulary words,
wherein the method is performed on an at least one computerized apparatus configured to perform the method.
2. The method according to claim 1 , wherein the first condition with respect to lengths of the sequences measured as the number of phonemes, comprises a first threshold above which the sequences are determined as candidate out-of-vocabulary patterns.
3. The method according to claim 1 , wherein the second condition with respect to the number of repetitions of the sequences comprises a second threshold above which the sequences are determined as candidate out-of-vocabulary patterns.
4. The method according to claim 1 , wherein training the speech recognition system comprises training the hybrid speech recognition system or a word-based speech recognition system.
5. The method according to claim 1 , further comprising validation of the trained speech recognition system by determining an adequate performance of the trained speech recognition system in recognizing of out-of-vocabulary words.
6. The method according to claim 5 , wherein the validation comprises at least one further amendment of the trained speech recognition system to achieve an adequate performance of the trained speech recognition system in recognizing of out-of-vocabulary words.
7. A system for adapting a speech recognition system for out-of-vocabulary words, comprising:
at least one processor;
an audio database for storing and retrieving audio signals responsive to instructions from said processor;
an audio sounder configured to audibly present distinct segments of audio signals from said audio database responsive to instructions from said processor;
at least one program for execution on said processor to perform the following:
decoding by a hybrid speech recognition a speech including out-of-vocabulary terms, thereby generating graphemic transcriptions of the speech with a mixture of recognized in-vocabulary words and unrecognized sub-words, while keeping a time track of the decoded segments of the speech;
converting the sub-words to patterns comprising a set of phoneme sequences by a process of concatenation of a phoneme representation of each sub-word;
subsequently, determining among the patterns, which patterns are candidate to represent out-of-vocabulary words based on a first condition with respect to the lengths of the pattern based on a number of phonemes and a second condition with respect to the number of repetitions of the pattern;
audibly presenting to a user with the audio sounder the candidate patterns representing out-of-vocabulary words from the corresponding segments of the speech according to the time track, and receiving from the user indications of valid out-of-vocabulary words responsive to the audible presentations of the candidate patterns; and
training a speech recognition system to additionally recognize the identified out-of-vocabulary words, thereby adapting the speech recognition to recognize out-of-vocabulary words.