Apparatus and method for speech recognition using a plurality of confidence score estimation algorithms

A speech recognition method, the method comprising:

• calculating, using at least one processing device, a first confidence score using a ratio between a likelihood of a keyword model for feature vectors per frame of a speech signal and a likelihood of a Filler model for the feature vectors;
• calculating a second confidence score based on a determined similarity between a Gaussian distribution trace of the keyword model per frame of the speech signal and a Gaussian distribution trace sample of a stored corresponding keyword of the keyword model;
• determining a confidence of a result calculated using the keyword model in accordance with a determination of whether a position determined by the first and second confidence scores is included in a keyword area or in a non-keyword area on a confidence coordinate system including an axis of the first confidence score calculated using the likelihood ratio of the first likelihood and the second likelihood and another axis of the second confidence score calculated based on the determined similarity,
• wherein the calculating the second confidence score comprises assigning a weight value which given in accordance with sub-word, a serial number, and an order of the mixed waveform probability distributions, and in accordance with an arrangement order of the Gaussian distribution trace and the Gaussian distribution trace sample;
• and
• identifying whether the first and second confidence scores are to be used to generate a reference function or to be used to recognize the speech signal, and
• wherein at least one of a position and form of the generated reference function is updated based on the speech signal.

A speech recognition method, comprising:

• determining, using at least one processing device, a first likelihood using a keyword model that extracted features of an input speech signal match a stored keyword of a keyword model;
• determining a second likelihood using a filler model that the extracted features of the input speech are noise or do not match the stored keyword;
• calculating a first confidence score based on a likelihood ratio of the first likelihood and the second likelihood;
• calculating a second confidence score based on a determined similarity between a Gaussian distribution trace of the keyword model of the input speech signal and a Gaussian distribution trace sample of a stored corresponding keyword;
• identifying whether the first and second confidence scores are to be used to generate a reference function or to be used to recognize the speech signal;
• and
• setting a standard of judgment as to confidence of the result of speech recognition on a coordinate system defined by axes respectively corresponding to the first confidence score and the second confidence score using the generated reference function, and determining whether the input speech signal corresponds to a word in a keyword database based on a position determined by the first and second confidence scores on the coordinate system including an axis of the first confidence score calculated using the likelihood ratio of the first likelihood and the second likelihood and another axis of the second confidence score calculated based on the determined similarity,
• wherein the calculating the second confidence score comprises assigning a weight value which given in accordance with a sub-word, a serial number, and an order of the mixed waveform probability distributions, and in accordance with an arrangement order of the Gaussian distribution trace and the Gaussian distribution trace sample,
• wherein at least one of a position and form of the generated reference function is updated based on the input speech signal.

A computer-readable non-transitory storage medium encoded with processing instructions to control a processor to execute a speech recognition method, the method comprising:

• calculating a first confidence score using a ratio between a likelihood of a keyword model for feature vectors per frame of a speech signal and a likelihood of a Filler model for the feature vectors;
• calculating a second confidence score based on a determined similarity between a Gaussian distribution trace of the keyword model per frame of the speech signal and a Gaussian distribution trace sample of a stored corresponding keyword of the keyword model;
• determining a confidence of a result calculated using the keyword model in accordance with a determination of whether a position determined by the first and second confidence scores is included in a keyword area or in a non-keyword area on a confidence coordinate system, including an axis of the first confidence score calculated using the likelihood ratio of the first likelihood and the second likelihood and another axis of the second confidence score calculated based on the determined similarity,
• wherein the calculating the second confidence score comprises assigning a weight value which given in accordance with a sub word, a serial number, and an order of the mixed waveform probability distributions in accordance with an arrangement order of the Gaussian distribution trace and the Gaussian distribution trace sample;
• and
• identifying whether the first and second confidence scores are to be used to generate a reference function or to be used to recognize the speech signal,
• wherein at least one of a position and form of the generated reference function is updated based on the speech signal.

A computer-readable non-transitory storage medium encoded with processing instructions to control a processor to execute a speech recognition method, the method comprising:

• determining a first likelihood using a keyword model that extracted features of an input speech signal match a stored keyword of a keyword model;
• determining a second likelihood using a filler model that the extracted features of the input speech are noise or do not match the stored keyword;
• calculating a first confidence score based on a likelihood ratio of the first likelihood and the second likelihood;
• calculating a second confidence score based on a determined similarity between a Gaussian distribution trace of the keyword model of the input speech signal and a Gaussian distribution trace sample of a stored corresponding keyword;
• identifying whether the first and second confidence scores are to be used to generate a reference function or to be used to recognize the speech signal;
• and
• setting a standard of judgment as to confidence of the result of speech recognition on a coordinate system defined by axes respectively corresponding to the first confidence score and the second confidence score using the generated reference function, and determining whether the input speech signal corresponds to a word in a keyword database based on a position determined by the first and second confidence scores on the coordinate system including an axis of the first confidence score calculated using the likelihood ratio of the first likelihood and the second likelihood and another axis of the second confidence score calculated based on the determined similarity,
• wherein the calculating the second confidence score comprises assigning a weight value which given in accordance with a sub-word, a serial number, and an order of the mixed waveform probability distributions, and in accordance with an arrangement order of the Gaussian distribution trace and the Gaussian distribution trace sample, and
• wherein at least one of a position and form of the generated reference function is updated based on the speech signal.