Bettering scores of spoken phrase spotting

1. A method for improving a phrase spotting confidence score, comprising:

providing, to a processor, a test speech and a transcription thereof;

obtaining, by the processor, an input phrase in a textual form for spotting in the provided test speech;

partitioning, by the processor, the input phrase into one or more phoneme sequences;

searching the one or more phoneme sequences in a phoneme lattice, by the processor, to find candidate events of the input phrase in the test speech;

calculating, by the processor, a posterior probability estimation (PPE) score for each of the candidate events;

generating, by the processor, a phonetic transcription for the input phrase;

applying, by the processor, a classifying model to the phonetic transcription of the test speech according to a posterior probability feature based on the calculated PPE scores and to a set of phrase spotting features related to the input phrase, the phrase spotting features selected from features extracted from a decoding process of the phrase, context-based features, or a combination thereof, thereby spotting the given phrase with a confidence score; and

using a trained monotone transformation, by the processor, to simplify the feature normalization and a Support Vector Machine (SVM) model into a single vector in a size of the phrase spotting feature set;

wherein the classifying model utilizes the Support Vector Machine (SVM) with a linear kernel;

wherein the features related to the input phrase are selected from phrase-related features, event-related features, decoding process features, context-based features, or a combination thereof;

wherein at least one feature related to the input phrase is based on a number of long phonemes of the input phrase with phoneme time durations longer than a predefined threshold; and

wherein the classifying model is priorly trained for spotting the input phrase according to the posterior probability feature and to the set of phrase spotting features.

2. The method according to claim 1 , further comprising transforming the confidence score according to a predetermined trained transformation, to obtain a transformed score.

3. The method according to claim 2 , comprising calculating one or more trained transformations to obtain a transformed score that is correlated to a predetermined characteristic.

4. The method according to claim 1 , wherein the classifying model is priorly trained according to a target context, language, accent, or scope.

5. The method according to claim 1 , wherein the training of the classifying model includes adaptation of the classifying model to a specific lexicon.

6. The method according to claim 1 , wherein the classifying model is priorly trained to enhance a confidence level of identifying the input phrase.

7. The method according to claim 1 , wherein the at least one feature related to the input phrase is based on an edit distance score in a lattice of the input phrase.

8. The method according to claim 1 , wherein the at least one feature related to the input phrase is based on a number of possible phonetic transcriptions of the input phrase.

9. The method according to claim 1 , wherein the at least one feature related to the input phrase is based on an edit probability score of the input phrase.

10. The method according to claim 1 , comprising identifying a subset of contributing features from a preliminary set of phrase spotting features.

11. The method according to claim 10 , comprising filtering non-robust or correlative features using a cross validation procedure.

12. The method according to claim 10 , wherein identifying a subset of contributing features from a supported set of phrase spotting features comprises:

determining, an initial subset of selected features; and

for each feature in the supported set of phrase spotting features, adding the feature to or removing the feature front the initial subset of selected features to obtain a temporary subset, training the classifiers using the temporary subset of features, and evaluating the quality of the phrase spotting results obtained after the phrase spotting classification process is performed using the temporary subset of features.

13. A method for improving a phrase spotting confidence score, comprising:

providing, to a processor, a test speech and a transcription thereof,

obtaining, by the processor, an input phrase in a textual form for spotting in the provided test speech;

partitioning, by the processor, the input phrase into one or more phoneme sequences;

searching the one or more phoneme sequences in a phoneme lattice, by the processor, to find candidate events of the input phrase in the test speech;

calculating, by the processor, a posterior probability estimation (PPE) score for each of the candidate, events;

generating, by the processor, a phonetic transcription for the input phrase; and

applying, by the processor, a classifying model to the phonetic transcription of the test speech according to a posterior probability feature based on the calculated PPE scores and to a set of phrase spotting features related to the input phrase, the phrase spotting features selected from features extracted from a decoding process of the phrase, context-based features, or a combination thereof, thereby spotting the given phrase with a confidence score; and

using a trained monotone transformation, by the processor, to simplify the feature normalization and a Support Vector Machine (SVM) model into a single vector in a size of the phrase spotting feature set;

wherein the classifying model is priorly trained for spotting the input phrase according to the posterior probability feature and to the set of phrase spotting features;

wherein the classifying model utilizes the Support Vector Machine (SVM) with a linear kernel.

14. A method for improving a phrase spotting confidence score, comprising:

providing, to a processor, a test speech and a transcription thereof;

obtaining, by the processor, an input phrase in a textual form for spotting in the provided test speech;

partitioning, by the processor, the input phrase into one or more phoneme sequences;

searching the one or more phoneme sequences in a phoneme lattice, by the processor, to find candidate events of the input phrase in the test speech;

calculating, by the processor, a posterior probability estimation (PPE) score for each of the candidate events;

generating, by the processor, a phonetic transcription for the input phrase; and

applying, by the processor, a classifying model to the phonetic transcription of the test speech according to a posterior probability feature based on the calculated PPE scores and to a set of phrase spotting features related to the input phrase, the phrase spotting features selected from features extracted from a decoding process of the phrase, context-based features, or a combination thereof, thereby spotting the given phrase with a confidence score;

using a trained monotone transformation, by the processor, to simplify the feature normalization and a Support Vector Machine (SVM) model into a single vector in a size of the phrase spotting feature set;

identifying, by the processor, a subset of contributing features from a preliminary set of phrase spotting features; wherein identifying the subset of contributing features from a supported set of phrase spotting features further comprises:

determining, by the processor, an initial subset of selected features; and

for each feature in the supported set of phrase spotting features, by the processor, adding the feature to or removing the feature from the initial subset of selected features to obtain a temporary subset, training the classifiers using the temporary subset of features, and evaluating the quality of the phrase spotting results obtained after the phrase spotting classification process is performed using the temporary subset of features;

wherein the classifying model utilizes the Support Vector Machine (SVM) with a linear kernel; and

wherein the classifying model is priorly trained for spotting the input phrase according to the posterior probability feature and to the set of phrase spotting features.