Keyword detector and keyword detection method
A keyword detector includes a processor configured to calculate a feature vector for each frame from a speech signal, input the feature vector for each frame to a DNN to calculate a first output probability for each triphone according to a sequence of phonemes contained in a predetermined keyword and a second output probability for each monophone, for each of at least one state of an HMM, calculate a first likelihood representing the probability that the predetermined keyword is uttered in the speech signal by applying the first output probability to the HMM, calculate a second likelihood for the most probable phoneme string in the speech signal by applying the second output probability to the HMM, and determine whether the keyword is to be detected on the basis of the first likelihood and the second likelihood.
1. A keyword detector comprising:
a processor configured to:
divide a speech signal into frames each with a predetermined time length;
calculate a feature vector including a plurality of features representing characteristics of a human voice, for each frame;
input the feature vector for each of the frames to a deep neural network to calculate a first output probability for each of a plurality of triphones according to a sequence of phonemes contained in a predetermined keyword, for each of at least one state of a Hidden Markov Model and calculate a second output probability for each of a plurality of monophones, for each of at least one state of the Hidden Markov Model;
calculate a first likelihood representing a probability that the predetermined keyword is uttered in the speech signal by applying the first output probability to the Hidden Markov Model;
calculate a second likelihood for a most probable phoneme string in the speech signal by applying the second output probability to the Hidden Markov Model;
determine whether the keyword is detected on the basis of the first likelihood and the second likelihood; wherein
the deep neural network is common to the plurality of triphones and the plurality of monophones, and comprises an input layer to which the feature vectors are input, a plurality of hidden layers common to the plurality of triphones and the plurality of monophones, and an output layer including a plurality of first output neurons corresponding to the plurality of triphones, respectively, and a plurality of second output neurons corresponding to the plurality of monophones, respectively, for each of the at least one state of the Hidden Markov Model, and
when the feature vectors are input to the input layer of the deep neural network, calculation of the first output probability includes calculating the first output probability on the basis of values respectively output from the plurality of first output neurons, and calculation of the second output probability includes calculating the second output probability on the basis of values respectively output from the plurality of second output neurons, and
learn the deep neural network by designating the values output from the output neurons such that when the feature vector calculated from a speech signal of a sample corresponding to a predetermined triphone among the plurality of triphones is input to the deep neural network, the value output from an output neuron corresponding to the predetermined triphone among the plurality of first output neurons, and the value output from an output neuron corresponding to a monophone identical to a central phoneme of the predetermined triphone among the plurality of second output neurons are larger than the values output from the other output neurons.
2. A keyword detection method comprising:
dividing a speech signal into frames each with a predetermined time length;
calculating a feature vector including a plurality of features representing characteristics of a human voice, for each frame;
inputting the feature vector for each of the frames to a deep neural network to calculate a first output probability for each of a plurality of triphones according to a sequence of phonemes contained in a predetermined keyword and a second output probability for each of a plurality of monophones, for each of at least one state of a Hidden Markov Model;
calculating a first likelihood representing a probability that the predetermined keyword is uttered in the speech signal by applying the first output probability to the Hidden Markov Model;
calculating a second likelihood for a most probable phoneme string in the speech signal by applying the second output probability to the Hidden Markov Model;
determining whether the keyword is detected on the basis of the first likelihood and the second likelihood; wherein
the deep neural network is common to the plurality of triphones and the plurality of monophones, and comprises an input layer to which the feature vectors are input, a plurality of hidden layers common to the plurality of triphones and the plurality of monophones, and an output layer including a plurality of first output neurons corresponding to the plurality of triphones, respectively, and a plurality of second output neurons corresponding to the plurality of monophones, respectively, for each of the at least one state of the Hidden Markov Model, and
when the feature vectors are input to the input layer of the deep neural network, calculation of the first output probability includes calculating the first output probability on the basis of values respectively output from the plurality of first output neurons, and calculation of the second output probability includes calculating the second output probability on the basis of values respectively output from the plurality of second output neurons, and
learning the deep neural network by designating the values output from the output neurons such that when the feature vector calculated from a speech signal of a sample corresponding to a predetermined triphone among the plurality of triphones is input to the deep neural network, the value output from an output neuron corresponding to the predetermined triphone among the plurality of first output neurons, and the value output from an output neuron corresponding to a monophone identical to a central phoneme of the predetermined triphone among the plurality of second output neurons are larger than the values output from the other output neurons.
3. A non-transitory computer-readable recording medium having recorded thereon a computer program for keyword detection that causes a computer to execute a process comprising:
dividing a speech signal into frames each with a predetermined time length;
calculating a feature vector including a plurality of features representing characteristics of a human voice, for each frame;
inputting the feature vector for each of the frames to a deep neural network to calculate a first output probability for each of a plurality of triphones according to a sequence of phonemes contained in a predetermined keyword and a second output probability for each of a plurality of monophones, for each of at least one state of a Hidden Markov Model;
calculating a first likelihood representing a probability that the predetermined keyword is uttered in the speech signal by applying the first output probability to the Hidden Markov Model;
calculating a second likelihood for a most probable phoneme string in the speech signal by applying the second output probability to the Hidden Markov Model;
determining whether the keyword is detected on the basis of the first likelihood and the second likelihood; wherein
the deep neural network is common to the plurality of triphones and the plurality of monophones, and comprises an input layer to which the feature vectors are input, a plurality of hidden layers common to the plurality of triphones and the plurality of monophones, and an output layer including a plurality of first output neurons corresponding to the plurality of triphones, respectively, and a plurality of second output neurons corresponding to the plurality of monophones, respectively, for each of the at least one state of the Hidden Markov Model, and
when the feature vectors are input to the input layer of the deep neural network, calculation of the first output probability includes calculating the first output probability on the basis of values respectively output from the plurality of first output neurons, and calculation of the second output probability includes calculating the second output probability on the basis of values respectively output from the plurality of second output neurons, and
learning the deep neural network by designating the values output from the output neurons such that when the feature vector calculated from a speech signal of a sample corresponding to a predetermined triphone among the plurality of triphones is input to the deep neural network, the value output from an output neuron corresponding to the predetermined triphone among the plurality of first output neurons, and the value output from an output neuron corresponding to a monophone identical to a central phoneme of the predetermined triphone among the plurality of second output neurons are larger than the values output from the other output neurons.