JP4521673B2 - Utterance section detection device, computer program, and computer - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide an utterance section detecting device capable of properly detecting an utterance section without reference to environmental noise. <P>SOLUTION: The utterance section detecting device includes a speech input part 104 which generates speech data in frames, a frame buffer 110 which stores the energy value of the frame-constituted speech in an FIFO basis, an initial environmental noise calculation part 112 which processes energy values of frames in the frame buffer 110 in a specified statistical method to calculate an initial value of an estimated value of environmental noise, a dynamic threshold calculation part 116 which calculates thresholds of energy values for detecting an utterance section by frames according to the energy values stored in the frame buffer 110 to vary following up the environmental noise included in speech data, and a state decision part 118 which decides the states of the frames according to the threshold. <P>COPYRIGHT: (C)2005,JPO&amp;NCIPI

Description

  The present invention relates to an apparatus for detecting an utterance section as preprocessing such as speech recognition processing, and in particular, an utterance section detection apparatus capable of avoiding erroneous detection of an utterance section due to environmental noise in real-time speech recognition processing, The present invention also relates to a speech energy normalization apparatus for calculating speech energy normalized as a feature amount for each frame.

  In processing such as speech recognition, prior to speech recognition, it is necessary to distinguish between an utterance interval in an input signal and other intervals (referred to as silent intervals). Otherwise, speech recognition of a section without speech gives a meaningless result.

  Conventionally, detection of such an utterance section (or silent section) is performed by calculating the power (energy) of an input voice signal, and if the value exceeds a predetermined threshold value, the utterance section and threshold value are calculated. If it is less than, it is performed by the method of setting it as a silence area. At this time, it is usual to detect a speech section or a silent section in consideration of the time during which such a condition is established.

  Such a technique is disclosed in Patent Document 1. Patent Document 1 discloses a technique for extracting a portion to be summarized in order to automatically create a summary from video information with audio. It is known that the magnitude of environmental noise varies depending on the content (genre) of video with audio. For example, news programs have low environmental noise, and programs such as sports broadcasts have high environmental noise. For this reason, there is a problem that when the utterance section is detected using the same threshold value, the result varies depending on the genre of the video information. For this purpose, in the technique disclosed in Patent Document 1, the video information has accompanying information indicating the genre, and a threshold value assigned in advance to each genre is selected according to the accompanying information.

JP 2003-101939 (paragraphs 209 and 210, FIGS. 1 and 7)

  However, with the technique described in Patent Document 1 described above, only one type of threshold value can be used for one piece of video information. Therefore, there is a problem that when the environmental noise changes in the program, there is a problem in detecting the utterance section.

  In particular, when performing real-time speech recognition, it is not considered that the supplementary information as described above can be used. In addition, when voice recognition is used for automatic answering by telephone, it is impossible to predict what environmental noise will exist in the background of the voice signal. For example, when sudden environmental noise occurs, there is a high possibility of erroneous detection of the speech section.

  Further, it is known that in voice recognition, it is effective to use a feature amount obtained by normalizing the voice energy of each frame with the maximum value of voice energy during speech. However, for that purpose, it is necessary to wait until the end of the utterance and calculate the maximum power during the utterance, and then normalize the voice energy of each frame during the utterance using the calculated maximum power. However, there is a problem that real-time speech recognition cannot be performed if waiting for the end of the utterance.

  Accordingly, an object of the present invention is to provide an utterance section detection apparatus capable of appropriately detecting an utterance section regardless of environmental noise.

  Another object of the present invention is to provide an utterance interval detection device that can appropriately detect an utterance interval even if environmental noise changes.

  Still another object of the present invention is to provide an utterance section detection device capable of appropriately detecting an utterance section in real time even if environmental noise changes.

  Still another object of the present invention is to provide an utterance section detection device capable of appropriately detecting an utterance section in real time even if there is a sudden change in environmental noise.

  Another object of the present invention is to provide a speech energy normalization apparatus capable of normalizing speech energy of each frame in real time.

  An utterance section detecting device according to a first aspect of the present invention includes a framing means for sequentially framing speech data, a speech energy value framed by the framing means for each frame, and a FIFO ( Frame energy calculation and storage means for storing energy values of a first number of frames in a First-In First-Out) format, and energy values of a second number of frames are stored in the frame energy calculation and storage means. In response, an initial value calculating means for calculating an initial value of the estimated value of the environmental noise included in the audio data by processing the energy values of the second number of frames according to a predetermined statistical method; Based on the initial value of the estimated value and the energy value of the sound sequentially stored in the frame energy calculation and storage means, Means for sequentially calculating the threshold value of the energy value for detecting the utterance section for each frame so as to change following the change of the ambient noise, and based on the threshold value, the second Utterance interval estimation means for estimating a frame corresponding to the start position or end position of the utterance interval of the voice data among the frames after the number of frames.

  An initial value of the environmental noise estimate is calculated by statistically processing the energy values of the second number of frames. Thereafter, on the basis of the initial value of the estimated value and the energy value of the sound sequentially stored in the frame energy calculation and storage means, the utterance is changed so as to follow the change of the environmental noise included in the sound data. The threshold value of the energy value for detecting the section is sequentially calculated for each frame. A frame corresponding to the start position or end position of the speech section of the voice data is estimated using the threshold value. Since the threshold value changes following the environmental noise change, the start position or the end position of the utterance section can be estimated accurately.

  Preferably, the initial value calculation means sets the second number of frames to a first cluster centered on the first energy value and a larger number than the first energy according to the energy value of each frame. Means for clustering into a second cluster centered on an energy value of 2 and means for outputting the first energy value as an initial value of an estimate of environmental noise.

  The audio signal includes environmental noise and speech. When each frame is clustered, it is considered that the frames are classified into two groups: a frame containing only environmental noise and a frame containing environmental noise and speech. When a frame is clustered into two clusters according to the magnitude of energy, the proportion of frames consisting only of environmental noise increases in the cluster consisting of the first frame with low energy. Therefore, if the average of the energy values of the frames of the first cluster is used as the initial value of the estimated value of the environmental noise, the initial value of the environmental noise can be estimated with high reliability.

  More preferably, the means for clustering comprises means for determining a boundary value for clustering the second number of frames into the first and second clusters, and an energy value less than the boundary value. And means for classifying frames having a first cluster and other frames into a second cluster.

  The means for determining the boundary value is for selecting two frames having a first sort order and a second sort order that are predetermined when sorting using the energy value as a key out of the second number of frames. The first average value calculating means for calculating the average value of the energy values of the two selected frames, and whether the energy value is smaller than the average value calculated by the first average value calculating means. On the basis of whether or not, a means for classifying the second number of frames into the first and second groups, and an average value of energy values of the frames belonging to the first and second groups, respectively. A second average value calculating unit; and a third average value calculating unit for further calculating an average value of the two average values calculated by the second average value calculating unit and outputting the average value as a boundary value. Good

  Preferably, the means for sequentially calculating the threshold value for each frame is based on the frame energy calculation and the energy value of the frame stored in the storage means and the initial value of the estimated value of the environmental noise. Means for estimating the environmental noise energy value of the frame stored in the calculation and storage means for each frame, and stationary background noise among the frame energy values stored in the frame energy calculation and storage means And means for sequentially estimating the maximum value of the total energy value of the uttered speech for each frame, the estimated energy value of the environmental noise, and the estimated total energy value of the background noise and the uttered speech And means for calculating a threshold value of energy for detecting an utterance section for each frame.

More preferably, the utterance period estimation means includes means for determining a state of frames after the second number of frames based on the threshold, the state includes a non-speech state, and the energy of environmental noise The means for sequentially estimating the value for each frame stores the means for storing the energy value of the environmental noise estimated at the time point one frame before and the time when the initial value of the estimated value of the environmental noise is calculated. Means for storing the initial value of the estimated value of the environmental noise in the means for performing, a value stored in the means for storing, a frame energy calculation and a frame energy value included in the storage means, and a state of the frame Based on the determination result by the means for determining
b (t) = b (t−1) × α + E (t) × (1−α) (When the state is a non-speech state)
b (t) = b (t−1) (When the state is other than non-speech state)
Where α is a predetermined forgetting factor, E (t) is the energy value of the frame at time t, and means for calculating the background noise b (t) at time t, and the means for storing is calculated The background noise b (t) is stored.

  The means for estimating the maximum value of the total energy value for each frame includes means for sorting the frames stored in the frame energy calculation and storage means using the energy value as a key and means for sorting. Means may be included for selecting the energy values of the frames that are in the predetermined order as a result of the sorting as the maximum value Emax (t) of the total energy values.

Preferably, the means for sequentially calculating the threshold value for each frame includes the threshold value Eth ₁ (t) for detecting the utterance start position at time t,
Including means for calculating according to Eth ₁ (t) = b (t) + max (β, Emax (t) −b (t)) × first constant.

More preferably, the means for sequentially calculating the threshold value for each frame further includes a threshold value Eth ₂ (t) for detecting the utterance end position at time t,
It includes means for calculating according to Eth ₂ (t) = b (t) + max (β, Emax (t) −b (t)) × second constant where the second constant <the first constant.

  The utterance section detecting device further normalizes the audio data of each frame using the larger of the maximum energy value of the audio data of each frame from the head of the utterance or a predetermined default reference value, and the audio feature parameter of each frame Voice energy normalization means may be included.

  Since normalization is performed using the maximum energy value of voice data of each frame from the beginning of the utterance or a predetermined default reference value, whichever is greater, without waiting for the end of the utterance, it is simulated in real time. It becomes possible to normalize. Therefore, speech energy can be obtained in real time as one of the speech feature parameters.

  Preferably, the sound energy normalization means includes a reference value storage means for storing a reference value for normalization, and a reference value in which the sound energy calculated by the frame energy calculation and storage means is stored in the reference value storage means. Detecting means for outputting a detection signal and a reference value stored in the reference value storage means in response to the detection signal output from the detection means for calculating and storing the frame energy By dividing the sound energy value calculated by the means calculated by the means and the frame energy calculation and storage means by the reference value stored in the reference value storage means, Dividing means for normalizing.

  More preferably, the utterance section detection device replaces the stored content of the reference value storage means with a predetermined default value in response to the utterance section estimation means estimating the frame corresponding to the end position of the utterance section. Means for further comprising:

  The utterance section detection device may further include means for setting a predetermined default value based on an option value given when the utterance section detection device is activated.

  A computer program according to the second aspect of the present invention is for operating a computer as one of the utterance section detection devices described above.

  A speech energy normalization apparatus according to a third aspect of the present invention is a speech energy normalization apparatus for calculating the normalized speech energy of framed speech data in real time, wherein a normalization reference value is set. Reference value storage means for storing, means for calculating the sound energy of the sound data for each frame, and the sound energy calculated by the sound energy calculation means exceeds the reference value stored in the reference value storage means And a reference value stored in the reference value storage means in response to the detection signal output from the detection means and the detection signal output from the detection means. By dividing the sound energy calculated by the value and the sound energy calculated by the sound energy calculating means by the reference value stored in the reference value storing means, And a dividing means for normalizing speech energy.

  At the beginning of the utterance interval, the voice energy is normalized using the default value as a reference value. If the voice energy of the frame exceeds the reference value in the middle of the speech section, the voice energy is normalized using the voice energy of the frame as a new reference value. Even if it does not reach the end of the utterance section, it is possible to normalize the voice energy in real time although it is pseudo. An error occurs at the beginning of the utterance interval, but when the speech energy actually reaches the maximum value in the utterance interval, accurate normalization can be performed thereafter. Further, by appropriately selecting a default value, an error occurring at the beginning of the utterance interval can be suppressed to a small value.

  Preferably, the speech energy normalization device detects the end of the utterance section and outputs an utterance end detection signal, and in response to the utterance end detection signal, the stored content of the reference value storage means And means for replacing with a default value.

  When the utterance period ends, the reference value can be reset to the default value again. Speech energy can be normalized using an appropriate reference value for each frame.

  More preferably, the speech energy normalization device further includes means for setting the predetermined default value based on an option value provided upon activation of the speech energy normalization device.

  Since the default value can be set according to the option value at the time of startup, the voice energy normalization apparatus can be operated with various option values as default values. As a result, it becomes easy to more appropriately realize the normalization processing of the sound energy.

  A computer program according to the fourth aspect of the present invention is for operating a computer as one of the above-described speech energy normalization apparatuses.

  A computer according to the fifth aspect of the present invention is programmed by the computer program according to the second aspect described above or the computer program according to the fourth aspect, and operates as an utterance section detection device or a speech energy normalization device.

  The speech segment detection apparatus according to the present embodiment changes a threshold value at the time of speech segment detection by a statistical method based on a voice signal input as a frame. At that time, a statistical method is devised so that the delay at the time of startup of the apparatus is minimized as much as possible and the utterance section can be stably detected even if there is sudden noise. In addition, when calculating the normalized speech energy of the frame as the feature amount parameter for speech recognition, a contrivance is made so that pseudo-normalization can be performed by real-time processing.

[Speech interval detection principle]
FIG. 1 shows an audio signal and various parameters used in the technique used for detecting a speech interval in the present embodiment. Referring to FIG. 1, utterance start position 26 and ending position 28 are determined for voice signal 20 using two threshold values, utterance start threshold value 22 and utterance end threshold value 24. The utterance start threshold value 22 and the utterance end threshold value 24 are determined by a statistical method from energy calculated in units of frames from input waveform data. A method for determining these will be described later.

  In FIG. 1, temporal parameters T1 to T6 used in the determination of an utterance section have the following meanings.

  T1: Pre-roll time When it is determined that a certain frame is the start position of an utterance, a mark as an utterance start frame is added to a frame (reference numeral 26 in FIG. 1) that is further back from this frame by this pre-roll time. Is done.

  T2: Utterance start determination time As a first condition for determining that an utterance has started, a time during which the energy value in units of frames must continuously exceed the utterance start threshold.

  T3: Shortest utterance time The minimum time during which the energy value of each frame must be exceeded in order to be determined to be the utterance start. It is determined that the utterance starts only after the energy value exceeds the utterance start threshold value continuously for T2 time and continuously for T3 time.

  T4: Longest silence time The longest time during which an utterance is not determined to be finished even if the energy value in units of frames falls below the utterance end threshold.

  T5: Utterance end determination time As a first condition for determining that the utterance has ended, a time during which the energy value in units of frames must continuously fall below the utterance end threshold. When the energy value falls below the utterance end threshold value for T5 hours continuously and falls below T4 time continuously, it is determined that the utterance is finished.

  T6: Afterroll time When it is determined that the utterance has ended in a certain frame, a frame as the utterance end frame is attached to a frame (reference numeral 28 in FIG. 1) at a position further lower than this frame by the afterroll time. .

  Symbols S1 to S4 written in the vicinity of the horizontal axis in FIG. 1 indicate the state of each frame determined by a method described later. FIG. 2 shows frame state transitions.

  Referring to FIG. 2, the frame transitions between four states (non-speech state (S1) 30, utterance start state (S2) 32, utterance state (S3) 34, and utterance end state (S4) 36). . Transitions between states are performed as follows.

  (1) In the non-speech state (S1) 30, when the energy value of the frame exceeds the speech start threshold 22, the state transitions to the speech start state (S2) 32 (arc 42).

  (2) If the utterance start state (S2) 32 continues for a predetermined time T3, the state becomes the utterance state (S3) 34 (arc 48).

  (3) In the utterance start state (S2) 32, when the energy value of the frame falls below the utterance start threshold 22, the state transitions to the non-utterance state (S1) 30 (arc 46).

  (4) In the utterance state (S3) 34, when the energy value of the frame falls below the utterance end threshold 24, the state transitions to the utterance end state (S4) 36 (arc 52).

  (5) If the utterance end state (S4) 36 continues for a certain time T4, the state transitions to the non-utterance state (S1) 30 (arc 58).

  (6) In the utterance end state (S4) 36, when the energy value of the frame exceeds the utterance end threshold 24, the state returns to the utterance state (S3) 34 (arc 54).

  (7) Otherwise, the state maintains the current state (arcs 40, 44, 50 and 56).

  In the apparatus according to the present embodiment, the various parameters described above are set manually when starting the apparatus. If there is no setting, the default value is used. The parameter setting portion has no direct relationship with the present invention and will not be described in detail in the following description.

[Frame structure]
As will be described later, the apparatus according to the present embodiment processes an audio input signal in units of frames. FIG. 3 is a schematic diagram for explaining the concept of frame and frame shift.

  Referring to FIG. 3, each of the frames 70, 72, 74,... Is an audio signal having a frame length Tw = 30 milliseconds. In this embodiment, the audio signal is sequentially framed while moving this frame on the time axis in units of 10 milliseconds. This amount of movement is called a frame shift amount. Accordingly, the audio data to be processed by the apparatus according to the present embodiment has a frame length of 30 milliseconds and a frame shift amount of 10 milliseconds.

  The energy of each frame is obtained by multiplying the data in the frame by the value indicated by the window function 80 (Humming window) and calculating the sum. A method for calculating energy for each frame will be described later.

  In the apparatus of the present embodiment, normally, the utterance start threshold value 22 and the utterance end threshold value 24 are dynamically calculated by statistically processing 100 frames of data. When dynamic processing is performed in this manner, processing cannot be started unless a certain amount of data is accumulated. On the other hand, if statistical processing is performed using too much data, the time delay until the device operates properly becomes long, and the beginning of the utterance may not be detected correctly.

  Therefore, in the apparatus according to the present embodiment, it is assumed that there is no sound for the first 400 milliseconds after the start of processing, and during this time, 40 frames of data are collected in the frame buffer. The initial value of the environmental noise is obtained using the data for 40 frames, and the initial value of the threshold is further determined using the value. Thereafter, the threshold value is dynamically calculated using the collected data while accumulating the frame data in the frame buffer until data for 100 frames is collected. After reaching 100 frames, the threshold value is calculated while maintaining 100 frame data in the FIFO (First-In First-Out) format. This maximum number of frames (maximum number of frames stored and used in the frame buffer) is referred to as a frame buffer size. In addition, the number of frames used for obtaining the initial value of the environmental noise is referred to as an initial buffer size. That is, in the apparatus of the present embodiment, the frame buffer size is 100 and the initial buffer size is 40.

  Note that these frame buffer size and initial buffer size are merely examples, and other values may be used.

  In the following description, the number of the input frame is represented by t (0 ≦ t). Since the frame is input every 10 milliseconds, t also represents the time. Therefore, in the following description, the “t-th frame” may be simply expressed by the expression “frame at time t”.

  By performing such a process, the delay at the start of the process is 400 milliseconds, and there is no practical problem. Usually, since the threshold value is calculated using 100 frame data, the speech section can be detected with high reliability.

[Device configuration]
FIG. 4 is a functional block diagram showing the configuration of the utterance period detection device according to the present embodiment. Referring to FIG. 4, this utterance period detecting device 100 is for detecting an utterance period from a voice signal given from a microphone 102. The utterance section detection apparatus 100 samples the voice signal given from the microphone 102, digitizes it by quantizing it, and outputs it as frame data of the above format every 10 milliseconds, and outputs the frame data. A voice input unit 104 for outputting a frame output signal 124 shown in FIG. 6 and an input buffer 106 for storing a plurality of frame data provided from the voice input unit 104 are included.

  The utterance section detecting apparatus 100 further reads out frame data from the input buffer 106 to calculate frame information such as an energy value, and stores frame information output from the frame information calculating unit 108. Frame buffer 110. The buffer size of the frame buffer 110 is 100 frames as described above. The frame buffer 110 can hold 100 pieces of input frame information in a FIFO format.

  In the present embodiment, frame information calculation section 108 calculates the sound energy E (t) of the frame at time t according to the following equation.

ただし、Ｎは１フレーム中のデータサンプル数、Ｓ_i（ｉ＝１〜Ｎ）はデータの値、Ｈ_i（ｉ＝１〜Ｎ）はハミング窓関数の値を、それぞれ示す。

Here, N represents the number of data samples in one frame, S _i (i = 1 to N) represents a data value, and H _i (i = 1 to N) represents a Hamming window function value.

  The utterance section detection apparatus 100 further normalizes the voice energy of the frame calculated by the frame information calculation unit 108 with respect to the maximum power during the utterance and writes it into the input buffer 106 as one element of the frame feature vector. A frame audio energy normalization processing unit 126 is included. It is recognized that it is effective to normalize the voice energy of a frame with the maximum energy in one utterance and use it as a feature quantity for voice recognition. However, for that purpose, it is necessary to wait until the end of the utterance and calculate the maximum value of the frame energy. However, real time processing cannot be performed.

  Therefore, the frame sound energy normalization processing unit 126 has a function of normalizing sound energy in real time although it is simulated by updating the dynamic range of sound energy in real time. Therefore, the frame audio energy normalization processing unit 126 has a configuration as shown in FIG.

  Referring to FIG. 5, frame speech energy normalization processing unit 126 stores a default maximum value used as a default value of maximum speech energy when there is no sufficiently large speech energy frame at the beginning of the speech. The default maximum value storage unit 132 for storing the default maximum value given from the default maximum value storage unit 132 in the first part of the utterance, and a frame having voice energy larger than the default maximum value is detected during the utterance In this case, the maximum value storage unit 134 for storing the value of the sound energy and the sound energy 128 from the frame information calculation unit 108 are divided by the maximum value stored in the maximum value storage unit 134 and the result is input. A division unit 136 for writing as one of the feature quantities of the corresponding frame in the buffer 106, and a maximum value storage unit 134 Comparing the value of both receiving and sound energy 128 from the output and the frame information calculation unit 108, and a comparing unit 138 for giving the maximum value storing unit 134 a comparison result signal 139. The comparison result signal 139 becomes the H (high) level when the value indicated by the sound energy 128 exceeds the maximum value stored in the maximum value storage unit 134, and becomes the L (low) level otherwise. Note that if there is an optional value given when this device (program) is activated as an option, it is rewritten with that value.

  When the signal 200 given from the state determination unit 118 indicates that the utterance has ended, the maximum value storage unit 134 stores the value of the default maximum value storage unit 132 as a new maximum value and compares the value from the comparison unit 138. When the result signal 139 becomes H level, the value indicated by the sound energy 128 is stored as a new maximum value. Therefore, the value stored in the maximum value storage unit 134 becomes the default value stored in the default maximum value storage unit 132 at the start of utterance, and when the voice energy exceeds the default value with the progress of the utterance, the voice energy is displayed. It becomes. Thereafter, the same processing is repeated while the utterance is in progress. By normalizing the voice energy of each frame using this value as the maximum value of the voice energy during speech, the voice energy can be normalized in real time although pseudo.

  It is desirable that the default value is determined in advance by an experiment.

  Further, the utterance section detecting device 100 receives the frame output signal 124 from the voice input unit 104, and the reading points and writing points of the input buffer 106, the frame information calculation unit 108 and the frame buffer 110, and the writing / reading of them. The frame data 160 stored in the frame buffer 110 is read up to 400 milliseconds after the start of processing by the input / output / address management unit 114 for managing the timing and the speech section detection device 100, and the initial environmental noise is calculated. Initial environmental noise calculation unit 112, frame data 192 from frame buffer 110, initial environmental noise estimation value 194 from initial environmental noise calculation unit 112, and current state is non-speech state (S1) 30 (FIG. 2). Signal) indicating whether or not a speech is started, and an utterance start threshold value is received therefrom. 2 and utterance termination threshold 24 dynamically calculates, and a dynamic threshold calculation unit 116 for outputting a signal 198 indicative of the value of the threshold.

  The input buffer 106, the frame buffer 110, and the like are realized by a semiconductor memory device or the like. The input / output / address management unit 114 is equipped with a timer, and in synchronization with the digitization of the audio data by the audio input unit 104, a pointer for writing to the input buffer 106, the frame buffer 110, etc., and a read pointer from them are provided. to manage. The input / output / address management unit 114 also provides the dynamic threshold value calculation unit 116 with an initial flag 196 that takes a value of H level when processing a frame up to 400 milliseconds after activation and thereafter takes a value of L level. It also has a function. The processing of the dynamic threshold value calculation unit 116 is controlled by the values of the initial flag 196 and the signal 190.

  The utterance section detection apparatus 100 further determines the frame state from the signal 198 indicating the threshold value output from the dynamic threshold value calculation unit 116 and the frame data 192 from the frame buffer 110 according to a method described later. In response to the state determination unit 118 for outputting the signal 200 indicating the state and the signal 200 indicating the state output from the state determination unit 118, the input data corresponding to the frame whose state has been determined is read from the input buffer 106. A feature vector for calculating the speech feature vector of this frame by a predetermined calculation method, and in the case of the start or end frame of the speech section, a mark indicating the same is added to the feature vector 122 and output. Output unit 120. The state determination unit 118 also has a function of generating a signal 190 indicating whether or not the current state is the non-speech state (S1) 30 and giving it to the dynamic threshold value calculation unit 116.

  FIG. 6 is a block diagram of the initial environmental noise calculation unit 112. The initial environmental noise calculation unit 112 sorts the energy values for each frame out of the frame information provided from the frame buffer 110 in ascending order, and sorts the frame energy. The sort processing unit 140 for storing in the storage unit 142, and the energy values of the frames corresponding to the magnitudes of 25% and 75% from the lower order of the energy values for each frame stored in the sort processing unit 140 are calculated, Each of them includes a seed calculation unit 144 for outputting values em1 and em2 which are seeds for clustering processing to be described later, and a storage unit 146 for storing these values em1 and em2.

Initial Environmental noise calculating unit 112 further reads the value em1 and em2 from the storage unit 146, a first average value calculating section 148 for calculating the average value e _average, the first average value calculation unit 148 outputs A frame classifying unit 150 for classifying each frame in the post-sort frame energy storage unit 142 into two clusters C1 and C2 based on whether or not the average value to be used is a boundary value and has a larger energy value, A second average value calculation unit 152 for calculating average values Em1 and Em2 of the energy values of the frames belonging to each of the two clusters C1 and C2 obtained by the frame classification unit 150 according to the following equation.

ただし、Ｎはフレームバッファ１１０内のフレーム数、Ｉ１はｅ_averageより小さいエネルギ値を持ち、クラスタＣ１に属するフレームの数、Ｉ２はｅ_averageより大きいエネルギ値を持ち、クラスタＣ２に属するフレームの数を、それぞれ表す。

Where N is the number of frames in the frame buffer 110, I1 has an energy value smaller than e _{average and} the number of frames belonging to cluster C1, I2 has an energy value larger than e _{average and} the number of frames belonging to cluster C2. , Respectively.

  The initial environmental noise calculation unit 112 further stores the two average values Em1 and Em2 calculated by the second average value calculation unit 152 in the storage unit 146 as new values em1 and em2, respectively, and further stores the first average value. The calculation unit 148, the frame classification unit 150, and the second average value calculation unit 152 repeatedly execute the above processing, and the average value Em1 obtained as a result is shown in FIG. 4 as an estimated value (em1) 194 of initial environmental noise. And a determination unit 154 for giving to the dynamic threshold value calculation unit 116 shown.

  Hereinafter, processing performed by the first average value calculation unit 148, the frame classification unit 150, and the second average value calculation unit 152 will be described with reference to FIGS. 4 and 6 to 9. FIG. In general, the energy value of each frame stored in the frame buffer 110 shown in FIG. 4 varies according to the magnitude of the energy of the input audio signal, as shown in FIG. When this is sorted in ascending order according to the magnitude of energy, it is assumed that the result is as shown in FIG. The sort processing performed by the sort processing unit 140 is such processing, and the frame information stored in the post-sort frame energy storage unit 142 corresponds to that shown in FIG.

  By sorting as shown in FIG. 8, a histogram of energy values can be easily obtained. An example is shown in FIG. If the speech signal contains environmental noise and speech components, the energy value of the frame containing only the environmental noise and the energy value of the frame containing the speech component are distributed around different values. It seems to be. Then, they will form two peaks in a histogram as shown in FIG. 9, a peak of a relatively low energy part and a peak of a relatively high energy part.

  The first average value calculation unit 148, the frame classification unit 150, and the second average value calculation unit 152 shown in FIG. 6 perform the peak of the 25% and 75% energy values first. As the initial position, the above-described two peaks are obtained by subsequent calculation, and the frames stored in the sorted frame energy storage unit 142 are divided into a frame close to the peak on the environmental noise side and a frame close to the peak on the utterance part side. And clustering into two clusters.

FIG. 10 is a functional block diagram of the dynamic threshold value calculation unit 116 shown in FIG. Referring to FIG. 10, dynamic threshold value calculation unit 116 receives frame data 192, and among the sorted frame information stored in frame buffer 110, the dynamic threshold value calculation unit 116 calculates the frame at the position 90% from the smallest. The maximum energy calculation unit 176 for outputting the energy as the maximum energy e _max (t) (maximum energy signal 182) in the number of frames corresponding to the frame buffer size up to the t-th frame, and the frame data 192 are received. An environmental noise calculation unit 170 for calculating an estimated value of environmental noise according to the above, and a storage unit 174 for storing the estimated value b (t−1) of the environmental noise calculated in the previous process for one frame. Including.

  The dynamic threshold value calculation unit 116 further includes an estimated value b (t−1) that is one frame earlier stored in the storage unit 174, an estimated value of the environmental noise given from the environmental noise calculation unit 170, In response to the initial environmental noise estimated value (em1) 194, if the initial flag 196 is at the H level, the initial environmental noise estimated value (em1) 194 is displayed, and the initial flag 196 is at the L level and the signal 190 indicating the state is displayed. If the value indicates a non-speech state, the output of the environmental noise calculation unit 170 is selected, and if the initial flag 196 is L level and the state signal 190 does not indicate a non-speech state, the output of the storage unit 174 is selected. And a selection unit 172 for outputting as environmental noise b (t) for the t-th frame. The output of the selection unit 172 is given to the storage unit 174 and stored.

  The dynamic threshold value calculation unit 116 further includes the utterance start threshold value 22 based on the maximum energy value from the maximum energy calculation unit 176 and the environmental noise b (t) in the t-th frame from the selection unit 172. A threshold value calculation unit 178 for dynamically calculating the utterance end threshold value 24 is included. A signal 198 representing the threshold value output from the threshold value calculation unit 178 is given to the state determination unit 118 and used for state determination.

The environmental noise calculation unit 170 includes the energy E (t) of the t-th frame in the frame data 192 stored in the frame buffer 110 and the environmental noise b () for the t−1th frame stored in the storage unit 174. The estimated value b ′ (t) of the environmental noise is calculated from t−1) according to the following formula 1.
[Formula 1]
b ′ (t) = b (t−1) × α + E (t) × (1−α)
Here, α represents a predetermined forgetting factor, and E (t) represents the energy of the t-th frame. The forgetting factor is a value between 0 and 1, but 0.8 is used in the present embodiment.

  If the state is other than the non-speech state, the selection unit 172 selects the environmental noise b (t−1) for the t−1th frame output from the storage unit 174. Therefore, the environmental noise does not change in this case. If the state is a non-speech state, the selection unit 172 selects the environmental noise estimated value b ′ (t) output from the environmental noise calculation unit 170.

Therefore, the environmental noise b (t) at time t output from the environmental noise calculation unit 170 is expressed by the following equation. Where E (t) is the energy value of the frame at time t, and α is the forgetting factor described above.
[Formula 2]
b (t) = b (t−1) × α + E (t) × (1−α) (When the state is a non-speech state)
b (t) = b (t−1) (When the state is other than non-speech state)
The threshold calculation unit 178 dynamically calculates the utterance start threshold Eth ₁ and the utterance end threshold c according to the following equations.
[Formula 3]
For 0 ≦ t <400 milliseconds
Eth1 (t) = b (t) + β × γ ₁
Eth2 (t) = b (t) + β × γ ₂ ,
400 milliseconds ≤ t
Eth ₁ (t) = b (t) + max (β, Emax (t) −b (t)) × γ ₁
Eth ₂ (t) = b (t) + max (β, Emax (t) −b (t)) × γ ₂
However, β is the lowest dynamic range of the utterance, and is 20 dB in the present embodiment. Further, γ ₁ and γ ₂ are an utterance start threshold ratio and an utterance end threshold ratio, respectively, and are constants of 0 or more and 1 or less respectively determined experimentally. In the present embodiment, γ ₁ = 0.25 and γ ₂ = 0.20 are used.

The utterance start threshold value Eth ₁ and the utterance end threshold value Eth ₂ calculated in this way are used as the utterance start threshold value 22 and the utterance end threshold value 24 when the utterance section described with reference to FIG. 1 is detected. It is done.

[Device operation]
The apparatus described above operates as follows.

-At startup-
At the time of startup, an area for storing a buffer and option values necessary for processing is secured in the storage device. The option value given at startup is checked, and if the option value is correct, the given value is set for the option. If no option value is given, set the default value. If there is an error in the value of the given option, a message to that effect is displayed and the process is terminated. As for the default maximum value storage unit 132 of the frame audio energy normalization processing unit 126 shown in FIG. 5, if an option value is given at the time of activation, that value is stored as a default value, and further stored in the maximum value storage unit 134. To do. If no option value is given, the default value on the program is stored in the default maximum value storage unit 132 and further stored in the maximum value storage unit 134.

  The write point and read point of each buffer are set to initial values.

  Note that the time (frame number) at which the actual processing is started after activation is set to t = 0. The frame state at this time is set to a non-speech state. Thereafter, the audio input unit 104 shown in FIG. 4 digitizes the electrical signal from the microphone 102 every 10 milliseconds with a frame length of 30 milliseconds.

-From 0 ms to 400 ms-
The initial flag 196 from the input / output / address management unit 114 is at the H level. When the number of data necessary for speech determination is collected, the voice input unit 104 writes a predetermined number of data as the number to be delivered in one process to the address specified by the buffer write pointer in the input buffer 106.

  The frame information calculation unit 108 reads data for one frame from the address specified by the read pointer in the input buffer 106, calculates frame energy, and writes it in the area corresponding to the frame in the frame buffer 110. The frame information calculation unit 108 also supplies the calculated frame energy to the division unit 136, the comparison unit 138, and the maximum value storage unit 134 shown in FIG. The comparison unit 138 compares the value stored in the maximum value storage unit 134 with the value indicated by the sound energy 128 and provides a comparison result signal 139 to the maximum value storage unit 134. When it is detected that the value indicated by the sound energy 128 exceeds the value stored in the maximum value storage unit 134, the comparison result signal 139 becomes H level, and the maximum value storage unit 134 indicates that the comparison result signal 139 is H. In response to reaching the level, the value represented by the voice energy 128 is stored in place of the previously stored value.

  The division unit 136 calculates the normalized speech energy by dividing the value represented by the speech energy 128 by the value stored in the maximum value storage unit 134. The normalized speech energy 130 is written into the normalized speech energy field of the corresponding frame in the input buffer 106. Thereafter, the frame information calculation unit 108 and the frame sound energy normalization processing unit 126 repeat the same operation for each frame.

  The initial environmental noise calculation unit 112 reads the frame energy written in the frame buffer 110 by the frame information calculation unit 108 and calculates the initial environmental noise. The state is not determined between time 0 milliseconds and 400 milliseconds.

  Next, the operation of the initial environmental noise calculation unit 112 will be described with reference to FIG. The sort processing unit 140 sorts the frame energy values 160 read from the frame buffer 110 and stores them in the post-sort frame energy storage unit 142. At t = 0, only one frame energy value is read (E (0)), so that value is written in the first area of the sorted frame energy storage unit 142. After the second time, the result of the previous sorting has already been written in the post-sorting frame energy storage unit 142, and it is only necessary to newly add one frame energy to the position according to the size (heap sorting). ). Therefore, the sort process can be executed with a small amount of calculation.

  After the activation, the processing units subsequent to the seeds calculation unit 144 do not operate for 0 to 400 milliseconds.

-When 400 milliseconds have passed-
When 400 milliseconds elapses after activation, the frame buffer 110 stores energy values of 40 frame data (E (0) to E (39)). This state corresponds to FIG. The post-sort frame energy storage unit 142 stores the energy values of these 40 frames sorted in ascending order. This state corresponds to FIG.

  The frame information calculation unit 108 and the frame audio energy normalization processing unit 126 operate in the same manner up to 400 milliseconds.

  The division unit 136 calculates the normalized speech energy by dividing the value represented by the speech energy 128 by the value stored in the maximum value storage unit 134. The normalized speech energy 130 is written into the normalized speech energy field of the corresponding frame in the input buffer 106.

  The seeds calculation unit 144 calculates values corresponding to 25% and 75% from the smallest of the 40 frame energies stored in the post-sort frame energy storage unit 142. This value is stored in the storage unit 146 and becomes a seed for clustering performed by the first average value calculation unit 148, the frame classification unit 150, and the second average value calculation unit 152.

  The first average value calculation unit 148 calculates the average value of the seeds em1 and em2 from the storage unit 146 and supplies the average value to the frame classification unit 150. The frame classification unit 150 classifies the 40 frames into two clusters based on whether the energy value of each frame is close to the seed em1 or em2, and calculates a second average value of the classified results. To unit 152.

  The second average value calculation unit 152 calculates, for each of the two clusters, average values Em1 and Em2 of the energy values of the frames belonging to the cluster, and gives them to the determination unit 154.

  The determination unit 154 stores Em1 and Em2 given from the second average value calculation unit 152 as new em1 and em2 in the storage unit 146, and performs the same processing as the first average value calculation unit 148, frame classification. The unit 150 and the second average value calculation unit 152 are executed. Of Em1 and Em2 obtained again in this way, Em1 is given to the dynamic threshold value calculation unit 116 as the initial environmental noise estimated value 194 (em1).

  The operation of the dynamic threshold value calculation unit 116 will be described with reference to FIG. The selection unit 172 of the dynamic threshold value calculation unit 116 selects em1 that is the estimated value 194 of the initial environmental noise as the initial value of b (t), and provides it to the storage unit 174 and the threshold value calculation unit 178. The storage unit 174 stores this value.

  On the other hand, the maximum energy calculation unit 176 calculates an energy value corresponding to 90% from the smallest of the sorted frame energy values stored in the post-sort frame energy storage unit 142, and calculates the maximum energy value ( Emax) 182 is given to the threshold value calculation unit 178.

  The threshold value calculation unit 178 is based on the estimated value em1 of the environmental noise given from the selection unit 172 and the maximum energy value (Emax) 182 from the maximum energy calculation unit 176, and the utterance start threshold value according to the above-described equation 3. 22 and the utterance end threshold 24 are calculated (198) and provided to the state determination unit 118 shown in FIG.

  The state determination unit 118 determines the state of the frame according to the determination method shown in FIGS. 1 and 2 based on the utterance start threshold value 22 and the utterance end threshold value 24 given from the dynamic threshold value calculation unit 116. A signal 200 representing the result is supplied to the feature vector output unit 120 and the frame audio energy normalization processing unit 126. The state determination unit 118 also provides the dynamic threshold value calculation unit 116 with a signal 190 indicating whether or not the frame state is a non-speech state.

  If the maximum value storage unit 134 (see FIG. 5) of the frame audio energy normalization processing unit 126 indicates that the utterance period is ended by the signal 200 indicating the state, the maximum maximum storage unit 134 replaces the previously stored value. The value of the value storage unit 132 is stored. By this process, at the start of the normalization process of the voice energy for the next utterance, the default value (or a value given as an option) is used again as the maximum power.

  The feature vector output unit 120 reads out the data of the frame whose state is determined by the processing of the state determination unit 118 from the input buffer 106, calculates the feature vector of the frame, and outputs (122). At this time, if the frame is an utterance start frame or an utterance end frame, the feature vector output unit 120 attaches a mark indicating the frame to the feature vector and outputs it.

-From 400 milliseconds to 1 second-
The initial flag 196 from the input / output / address management unit 114 is turned off. The processing up to the 100th frame after the 40th frame is almost the same as the processing for the 40th frame. In the processing during this time, data for one frame is added to the frame buffer 110 every 10 milliseconds. As a result, the state determination is executed using all the frame information stored in the frame buffer 110.

  In addition, in the dynamic threshold value calculation unit 116 shown in FIG. 10, the storage unit 174 has already stored the estimated value b (t−1) of the environmental noise calculated in the process for the previous frame. The environmental noise calculation unit 170 uses the environmental noise estimated value b (t−1) stored in the storage unit 174 and the energy E (t) of the t-th frame obtained from the frame data 192 according to Equation 1. An estimated noise value b ′ (t) is calculated and given to the selection unit 172.

  Since the value of the initial flag 196 is off, the selection unit 172 selects either the output of the storage unit 174 or the output of the environmental noise calculation unit 170 according to the value of the signal 190 indicating the state. That is, if the state represented by the signal 190 is a non-speech state, the selection unit 172 selects the output of the environmental noise calculation unit 170, and otherwise selects the output of the storage unit 174. The selection unit 172 gives a signal indicating the selected value to the storage unit 174 and the threshold value calculation unit 178.

  In other respects, the dynamic threshold value calculation unit 116 performs the same process as the process for the 40th frame. The operations of the state determination unit 118, the feature vector output unit 120, and the frame audio energy normalization processing unit 126 are the same.

-After 1 second-
The processing after the 101st frame is almost the same as the processing from 400 milliseconds to 1 second. However, in this process, when new frame information is added to the frame information stored in the frame buffer 110, the oldest frame information is deleted. That is, the frame buffer 110 stores data in the FIFO format. As a result, frame information for 100 frames is always maintained in the frame buffer 110. The same applies to the sort processing by the sort processing unit 140. In the post-sort frame energy storage unit 142, the energy value of the oldest frame is deleted, and the energy value of a new frame is written at a position determined according to the size.

  The initial environmental noise calculation unit 112, the dynamic threshold value calculation unit 116, the state determination unit 118, and the feature vector output unit 120 all estimate background noise based on 100 frames of data stored in the frame buffer 110. , Threshold value calculation, state determination, and feature vector creation are repeatedly executed.

  Thus, the feature vector 122 for each frame output from the feature vector output unit 120 is provided with an utterance start marker if the frame is an utterance start position, and an utterance end marker if the frame is an utterance end position. . With this marker, it is possible to detect the utterance section (from the utterance start position to the utterance end position) of the first voice data.

  Further, the feature vector 122 includes a value obtained by normalizing the sound energy for each frame, and this can be used as a feature amount for speech recognition. This voice energy is not the maximum value calculated by examining the entire utterance, but is pseudo-normalized with the maximum value updated in real time by the maximum value from the beginning of the utterance. This is shown in FIG.

  With reference to FIG. 11, the transition of the maximum value of the sound energy determined by the normalization process will be described. Referring to FIG. 11, according to the conventional method, the maximum value of the speech energy of the utterance is examined at the time when the speech is completed, and the speech energy is normalized by the value. In FIG. 11, the maximum value of the sound energy is represented by a dotted line 212 followed by a thick solid line 218.

  On the other hand, in the above-described embodiment, the maximum value of the voice energy indicated by the dotted line 212 is approximated by a constant default value (or option value) 214 at the start time of the utterance. Further, when the value of the voice energy becomes larger than this default value (the portion of the thick solid curve 216 in FIG. 11), the approximate value of the maximum value of the voice energy is replaced with that value. After reaching the maximum position of the actual voice energy during speech, this approximate value becomes equal to the actual maximum value (the thick solid line 218).

  By this normalization processing, the sound energy can be normalized in real time. Since the default value is used as the maximum value at the beginning of each utterance, there will be some error, but by setting the default value to an appropriate size, even pseudo-normalization is sufficient Effects can be obtained.

-Effect of the embodiment-
According to the apparatus of the present embodiment described above, the utterance start threshold value and the utterance end threshold value for the start and end of the utterance are obtained by statistically processing the actual voice data, thereby It is dynamically changed according to the data. It is possible to detect an utterance interval using a threshold value that changes following environmental noise. As a result, it is possible to correctly detect an utterance section while minimizing the influence of environmental noise.

  In the apparatus according to the above-described embodiment, the maximum energy value of the frame used for calculating the threshold value is 90% of the actual maximum value. Therefore, it is possible to suppress a large change in threshold due to a sudden change in environmental noise. In addition, since the threshold value is calculated by statistical processing for the amount of frames corresponding to the frame buffer size, even if there is a change in the energy value that protrudes in some frames, the effect of the change on the threshold value Is relatively small. As a result, the threshold value can be calculated stably.

  Furthermore, in the apparatus of the present embodiment, the state determination is started when the number of frame data becomes 40. Since a certain number is required for the statistical processing, the reliability of the state determination result is low in the threshold value calculation using a very small number of frame data. Therefore, the threshold value calculation may be started based on the voice data of about 300 milliseconds at the minimum, preferably about 400 milliseconds as in the apparatus of the present embodiment. Moreover, since the state determination is started when the number of frames to be processed reaches 40, the delay time from the start to the start of the state determination is about 400 milliseconds. This magnitude of the delay is not a problem in practical use. If the delay is too large, detection of the start of the utterance interval may fail. In the above embodiment, the delay is 400 milliseconds, but 1000 milliseconds of data is used for threshold determination, so that a highly reliable threshold calculation can be performed with a small delay. .

[Modification]
In the above-described embodiment, the Hamming window is used as the window function for calculating the frame energy. However, usable window functions are not limited to this. Of various window functions, such as Hanning window, Blackman, Kaiser, and Blackman-Harris, those that are considered appropriate may be used.

  In the above embodiment, the frame buffer size is 100 and the initial buffer size is 40. These values are merely examples, and other combinations can be taken. However, if the frame buffer size is too large, it becomes difficult to change the threshold value following the change in environmental noise. If the frame buffer size is too small, the threshold value changes in response to a slight change in environmental noise, and the speech section cannot be detected stably. In addition, if the initial buffer size is too large, the delay until the environmental noise is estimated increases, and the possibility of failing to detect the head of the speech segment increases. Of course, the initial buffer size must be less than or equal to the frame buffer size. Accordingly, the frame buffer size is preferably about 300 to 2000 milliseconds, and the initial buffer size is preferably about 200 to 500 milliseconds. In particular, a frame buffer size of about 600 to 1000 milliseconds and an initial buffer size of about 300 to 450 milliseconds are suitable.

  In the above-described embodiment, a fixed value calculated in advance is used as a default value for normalization of voice energy. However, the present invention is not limited to such an embodiment. For example, this default value can be updated with the maximum power of the previous utterance at the end of the utterance. At this time, the maximum energy may be multiplied by a predetermined coefficient a (0 <a ≦ 1, preferably 0.7 <a <0.9, more preferably about a = 0.8). Further, the default value may be updated as a function of the maximum energy in the past utterances as well as the immediately preceding utterance.

  In the embodiment described above, the logarithmic speech energy obtained by multiplying the absolute value of each speech data in the frame by the average value obtained by multiplying the value of the window function and multiplying by the coefficient 20 is normalized. This is used as a characteristic parameter of voice energy. However, the present invention is not limited to such an embodiment. For example, a logarithm of an average value obtained by multiplying the square of each voice data by the value of the window function is taken, and the logarithmic voice energy is multiplied by a coefficient of 10. The present invention can be similarly applied to the case where the calculation is performed.

  The apparatus of the above-described embodiment is assumed to be realized by a processor such as a DSP (Digital Signal Processor) and a program (including a microprogram) executed on the processor. From the above description, it will be easy for those skilled in the art to create such a program.

  The embodiment disclosed herein is merely an example, and the present invention is not limited to the above-described embodiment. The scope of the present invention is indicated by each claim in the claims after taking into account the description of the detailed description of the invention, and all modifications within the meaning and scope equivalent to the wording described therein are intended. Including.

It is a figure for demonstrating the method of the speech section determination in this invention, and the parameter for it. It is a state transition diagram in the utterance section process in the present invention. It is a figure for demonstrating frame length and frame shift amount. It is a functional block diagram of the utterance area detection apparatus which concerns on one embodiment of this invention. It is a block diagram of the sound energy normalization processing part of the apparatus shown in FIG. FIG. 5 is a functional block diagram of an initial environmental noise calculation unit of the apparatus shown in FIG. 4. It is a figure which shows the example of a change of a frame energy. It is a figure which shows the result of having sorted frame energy into ascending order. It is a histogram of frame energy. It is a functional block diagram of the dynamic threshold value calculation part of the apparatus shown in FIG. It is a figure for demonstrating the audio | voice energy normalization process in one embodiment of this invention.

Explanation of symbols

  20 speech signal, 22 utterance start threshold, 24 utterance end threshold, 30 non-utterance state (S1), 32 utterance start state (S2), 34 utterance state (S3), 36 utterance end state (S4), 100 Speaking section detection device, 102 microphone, 104 voice input unit, 106 input buffer, 108 frame information calculation unit, 110 frame buffer, 112 initial environmental noise calculation unit, 114 input / output / address management unit, 116 dynamic threshold calculation unit 118, state determination unit, 120 feature vector output unit, 122 feature vector, 124 frame output signal, 126 frame speech energy normalization processing unit, 140 sort processing unit, 142 post-sort frame energy storage unit, 144 seed calculation unit, 146, 174 Storage unit, 148 First average value calculation unit, 150 frames Classification unit, 152 second average value calculating unit, 154 determination unit, 160 frame data 170 environmental noise calculating unit, 172 selector, 176 maximum energy calculator, 178 threshold value calculation unit

Claims (13)

Framing means for sequentially framing audio data;
Frame energy calculation and storage for calculating the energy value of the voice framed by the framing means for each frame and storing the energy values of the first number of frames in a FIFO (First-In First-Out) format. Means,
In response to storing the energy values of the second number of frames in the frame energy calculation and storage means, processing the energy values of the second number of frames according to a predetermined statistical technique, An initial value calculating means for calculating an initial value of an estimated value of environmental noise included in the audio data;
Based on the initial value of the estimated value and the energy value of the sound sequentially stored in the frame energy calculation and storage means, the utterance section changes so as to follow the change of the environmental noise included in the sound data. Means for sequentially calculating a threshold value of energy value for detecting each frame,
Utterance section estimation means for estimating a frame corresponding to a start position or an end position of the utterance section of the voice data among the frames after the second number of frames based on the threshold value. An utterance section detecting device ,
The initial value calculating means includes
Depending on the magnitude of the energy value of each frame, the second number of frames is centered on a first cluster centered on the first energy value and a second energy value greater than the first energy. Means for clustering with a second cluster to:
Means for outputting the first energy value as an initial value of the estimated value of the environmental noise . The means for clustering is:
Means for determining a boundary value for clustering the second number of frames into the first and second clusters;
The speech section detection according to claim 1 , further comprising means for classifying a frame having an energy value smaller than the boundary value into the first cluster and a frame other than the frame into the second cluster. apparatus. The means for determining the boundary value is:
Means for selecting two frames having a first sort order and a second sort order that are predetermined when sorting using the energy value as a key out of the second number of frames;
First average value calculating means for calculating an average value of energy values of the two selected frames;
Means for classifying the second number of frames into first and second groups based on whether the energy value is smaller than the average value calculated by the first average value calculating means;
Second average value calculating means for calculating average values of energy values of frames belonging to the first and second groups, respectively.
The utterance according to claim 2 , further comprising: a third average value calculating means for further calculating an average value of the two average values calculated by the second average value calculating means and outputting the average value as the boundary value. Section detection device. Means for sequentially calculating the threshold value for each frame,
Based on the energy value of the frame stored in the frame energy calculation and storage means and the initial value of the estimated value of the environmental noise, the environmental noise energy of the frame stored in the frame energy calculation and storage means Means for estimating the value for each frame;
Means for sequentially estimating, for each frame, a maximum value of the total energy value of stationary background noise and speech among the energy values of the frames stored in the frame energy calculation and storage means;
To calculate an energy threshold value for detecting the utterance period for each frame based on the estimated energy value of the environmental noise and the estimated energy value of the background noise and the uttered speech. The utterance section detecting device according to claim 1, comprising: The utterance interval estimation means includes means for determining a state of a frame after the second number of frames based on the threshold;
The state includes a non-speaking state,
Means for sequentially estimating the energy value of the environmental noise for each frame,
Means for storing an energy value of the environmental noise estimated at a time point one frame before;
Means for storing the initial value of the estimated value of the environmental noise in the means for storing at the time when the initial value of the estimated value of the environmental noise is calculated;
Based on the value stored in the means for storing, the energy value of the frame included in the frame energy calculation and storage means, and the determination result by the means for determining the state of the frame, the following formula
b (t) = b (t−1) × α + E (t) × (1−α) (When the state is a non-speech state)
b (t) = b (t−1) (When the state is other than non-speech state)
Where α is a predetermined forgetting factor, E (t) is the energy value of the frame at time t, and means for calculating the background noise b (t) at time t,
The utterance section detection device according to claim 4 , wherein the storing means stores the calculated background noise b (t). Means for estimating the maximum value of the total energy value for each frame,
Means for sorting the frames stored in the frame energy calculation and storage means using energy values as keys;
6. The speech section according to claim 5 , comprising means for selecting, as the maximum value Emax (t) of the total energy values, the energy values of frames that are in a predetermined order as a result of sorting by the means for sorting. Detection device. Means for sequentially calculating the threshold value for each frame,
The threshold value Eth1 (t) for detecting the utterance start position at time t is
Eth1 (t) = b (t) + max (β, Emax (t) −b (t)) × first constant ,
Where β is a constant predetermined as the minimum dynamic range of the audio data signal,
The utterance section detection device according to claim 6 , comprising means for calculating according to The means for sequentially calculating the threshold value for each frame further includes:
The threshold value Eth2 (t) for detecting the utterance end position at time t is
Eth2 (t) = b (t) + max (β, Emax (t) −b (t)) × second constant ,
Where the second constant <the first constant,
β is the constant that is predetermined as the minimum dynamic range of the audio data signal,
The utterance section detection device according to claim 7 , comprising means for calculating according to Further, the maximum energy value of the audio data of each frame from the head of the utterance or the predetermined default reference value, whichever is greater, is used to normalize the audio data of each frame and output it as the audio feature parameter of each frame including speech energy normalization means, voice activity detection device according to any one of claims 1 to 8. The voice energy normalization means includes
A reference value storage means for storing a normalization reference value;
Detecting means for detecting that the sound energy calculated by the frame energy calculating and storing means exceeds a reference value stored in the reference value storing means, and outputting a detection signal;
Means for replacing a reference value stored in the reference value storage means with a value calculated by the frame energy calculation and storage means in response to the detection signal output by the detection means;
Dividing means for normalizing the voice energy of the frame by dividing the voice energy value calculated by the frame energy calculation and storage means by the reference value stored in the reference value storage means; The utterance section detection apparatus according to claim 9 . In response to the estimation of the frame corresponding to the end position of the utterance interval by the utterance interval estimation means, the information further includes means for replacing the stored content of the reference value storage means with a predetermined default value. voice activity detection apparatus according to claim 1 0. The predetermined default value, further comprising means for setting, based on the option value given at the start of the voice activity detection device, voice activity detection apparatus according to claim 1 0 or claim 1 1. When executed by a computer, to operate as voice activity detection apparatus according to the computer in any one of claims 1 to 1 2, a computer program for voice activity detection.

Images