JP2019039946A - Model learning device, voice section detection device, method thereof and program - Google Patents

Abstract

To improve flexibility in voice section detection.SOLUTION: Input voice likelihood series corresponding to voice likelihood of respective time sections of an input sound signal is applied to a state transition model for obtaining an estimation result of state transition in the respective time sections on a voice state and a non-voice state of the input sound signal with the input voice series corresponding to voice likelihood of the respective time sections of the input sound signal as input. The estimation result of the state transition in the respective time sections on the voice state and the non-voice state of the input sound signal is obtained to be outputted.SELECTED DRAWING: Figure 1

Description

  The present invention relates to a speech section detection technique.

  There is a technique called VAD (voice activity detection) as one of the voice segment detection techniques (see, for example, Non-Patent Document 1). In VAD, a voice section is performed using sound intensity, vibration intensity (number of zero crossings), acoustic features, and the like. However, VAD has a problem that a consonant or a short pause between words is determined as a non-speech segment and a segmented speech segment is detected. In order to cope with this, a technique called hangover is used (see, for example, Non-Patent Document 1). This is a method in which, when the number of frames of a non-speech section between two speech sections obtained by VAD is shorter than a threshold, these two speech sections are regarded as a series of speech sections.

ITU, “A silence compression scheme for G.729 optimized for terminals conforming to recommendation V.70,” ITUT / Recommendation G.729-Annex B. 1996.

  However, the conventional method using hangover has low flexibility. That is, the length of the non-speech section between two speech sections obtained by VAD differs depending on scenes such as a speech task and a domain. Therefore, it is necessary to manually set an optimum threshold value for each assumed scene. In the same scene, ideally, an appropriate threshold value should be given for each utterance. However, the conventional method does not have such flexibility.

  The present invention has been made in view of these points, and an object thereof is to improve the flexibility in voice segment detection.

  A state transition model that uses the input speech likelihood sequence corresponding to the speech likelihood of each time interval of the input acoustic signal as input and obtains estimation results of state transitions in each time interval for the speech state and non-speech state of the input acoustic signal Applying the input speech likelihood sequence corresponding to the speech likelihood of each time interval of the input acoustic signal, and obtaining and outputting the estimation result of the state transition in each time interval for the speech state and the non-speech state of the input acoustic signal To do.

  In the present invention, the flexibility in voice segment detection can be improved.

FIG. 1A is a block diagram illustrating a functional configuration of the model learning device according to the embodiment. FIG. 1B is a block diagram illustrating a functional configuration of the voice detection device according to the embodiment. FIG. 2 is a state transition diagram of the state transition model of the embodiment. 3A and 3B are diagrams illustrating acoustic signals. FIG. 3C is an example of a speech segment and a non-speech segment detected by VAD. FIG. 3D is an illustration of a speech segment and a non-speech segment detected by the method of the embodiment. FIG. 4A is a diagram illustrating an acoustic signal having a plurality of speech sections. FIG. 4B is an example of a speech segment and a non-speech segment detected by the method of the embodiment.

Embodiments of the present invention will be described below.
[Overview]
First, the outline of each embodiment will be described. In each embodiment, the input speech likelihood sequence corresponding to the speech likelihood of each time interval of the input acoustic signal is applied to the state transition model, and the speech state and the non-speech state of the input acoustic signal in each time interval are applied. Obtain and output state transition estimation results. The “state transition model” is an input speech likelihood sequence corresponding to speech likelihood in each time interval of the input acoustic signal, and state transition in each time interval for the speech state and non-speech state of the input acoustic signal. This model obtains the estimation result of This state transition model is obtained by modeling the transition of an actual speech segment and a non-speech segment (that is, how the actual speech segment and the non-speech segment appear) with respect to the input speech likelihood sequence. Therefore, by applying the input speech likelihood sequence to this state transition model, even if the input speech likelihood sequence represents a time interval such as a consonant or a short pause, it is not a non-speech interval but a speech interval Can be estimated appropriately. As a result, it is possible to solve the problem that a consonant or a short pause is determined as a non-speech segment and a segmented speech segment is detected. Moreover, since the generation of the state transition model can be applied to various input speech likelihood sequences, it is more flexible than hangover. In particular, the state transition model according to the present embodiment estimates transitions between speech intervals and non-speech intervals using an input speech likelihood sequence as an input. The input speech likelihood sequence is obtained by collecting input acoustic signals whose fluctuations are significant according to the environment into a series corresponding to speech likelihoods having smaller fluctuations. Therefore, the state transition model of this embodiment can flexibly cope with various environments and enables high-precision estimation. As described above, according to the method of each embodiment, it is possible to improve the flexibility in voice section detection as compared with the conventional technique. Generation of the state transition model is possible by machine learning using learning data, and the tuning cost is lower than that of hangover in which a threshold is manually set.

  The “input sound signal” is a time-series digital sound signal divided into a plurality of predetermined time intervals (for example, frames, subframes, etc.). The “speech likelihood” of a time interval represents the likelihood (likelihood) that the time interval is a speech interval. The “speech likelihood” may indicate the likelihood that the time interval is a speech interval as it is, or indirectly indicate the likelihood that the time interval is a non-speech interval, thereby indicating the likelihood that the time interval is a non-speech interval. It may be expressed. The “input speech likelihood sequence” is a time series of values corresponding to the “speech likelihood” of each time interval. The “input speech likelihood series” may be a time series of “speech likelihood” of each time interval, or may be a time series of values representing “speech likelihood” of each time interval. The value representing “voice likelihood” may be a continuous value or a binary value. For example, the value representing “speech likelihood” may be a function value of “speech likelihood” or may be a binary value that is threshold-determined using “speech likelihood”. However, from the viewpoint of estimation accuracy, the “input speech likelihood sequence” is preferably a time series of “speech likelihood” in each time interval or a time series of values representing “speech likelihood” that is a continuous value. . The “state transition” in each time interval for the speech state and the non-speech state is, for example, a state transition from the speech state to the speech state, a state transition from the non-speech state to the non-speech state, or from the speech state to the non-speech state. State transition and state transition from a non-voice state to a voice state. However, the state transition from the voice state to the voice state means that the voice state is maintained, that is, the “voice state”. Similarly, the state transition from the non-voice state to the non-voice state means that the non-voice state is maintained, that is, the “non-voice state”. The state transition estimation result obtained by applying the input speech likelihood sequence to the state transition model may be the likelihood of each state transition, or may be a function value of the likelihood of each state transition. The value obtained by comparing the likelihood of each state transition (for example, the state transition with the highest likelihood) may be used. The state transition model learning includes a speech likelihood sequence corresponding to the speech likelihood of each time interval of the acoustic signal, and a sequence of correct values of state transitions in each time interval for the speech state and the non-speech state of the acoustic signal. And learning data including the set of. In this learning data, elements of each time interval of the speech likelihood series (for example, speech likelihood) and the correct value of the state transition in each time interval are associated with each other. For example, an element of each time interval of the speech likelihood series and a correct value of the state transition in each time interval are associated with an identifier representing the time interval. Usually, the learning data includes a set of a speech likelihood sequence for a plurality of acoustic signals and a sequence of correct values of state transitions. That is, the learning data includes a plurality of sets of a speech likelihood sequence and a sequence of correct values of state transitions. However, the learning data may include only a set of a speech likelihood sequence for one acoustic signal and a sequence of correct values of state transitions. The “acoustic signal” is a time-series digital acoustic signal divided into a plurality of predetermined time intervals (for example, frames, subframes, etc.). The speech likelihood sequence of the learning data does not need to be the same as the assumed input speech likelihood sequence, but needs to be of the same type. For example, if the assumed input speech likelihood sequence is a time series of speech likelihoods in each time interval, the speech likelihood sequence of learning data also needs to be a time series of speech likelihoods in each time interval.

[First Embodiment]
A first embodiment will be described.
<Model learning device 11>
First, the model learning device 11 of this embodiment will be described. As illustrated in FIG. 1A, the model learning device 11 of the present embodiment includes a storage unit 111 that stores learning data 111a and a learning unit 112 that learns the state transition model 123a.

  The learning data 111a includes a speech likelihood sequence corresponding to the speech likelihood of each time interval of the acoustic signal, a sequence of correct values of state transitions in each time interval for the speech state and non-speech state of the acoustic signal, A set of The learning unit 112 uses the learning data 111a read out from the storage unit 111 and inputs an input speech likelihood sequence corresponding to the speech likelihood of each time interval of the input acoustic signal, regarding the speech state and non-speech state of the input acoustic signal. The state transition model 123a for obtaining the state transition estimation result in each time interval is obtained and output. State transition of this form is state transition from voice state to voice state, state transition from non-voice state to non-voice state, state transition from voice state to non-voice state, and state from non-voice state to voice state One of the transitions.

  The speech likelihood series of each time interval included in the learning data 111a is obtained by extracting acoustic information and analysis information from the acoustic signal for learning data and applying known VAD to them (for example, Non-Patent Document 1). Etc.). As acoustic information and analysis information used for VAD, for example, a change in sound power, the number of zero crossings per fixed time of a waveform, a change in the characteristics of acoustic features, and combinations thereof can be exemplified. The acoustic signal for learning data is a time-series digital acoustic signal divided every predetermined time interval. The learning data acoustic signal may be an analog acoustic signal observed with a microphone or the like obtained by AD conversion at a predetermined sampling frequency, or may be an arbitrary digital acoustic signal created in advance. An example of the speech likelihood series is as described above, and is a time series of values obtained for each time interval of the acoustic signal for learning data. When the speech likelihood sequence is a speech likelihood of each time interval or a series of continuous values representing the speech likelihood, the speech likelihood sequence is, for example, 0.1, 0.5,. When the speech likelihood sequence is a binary value sequence representing the speech likelihood of each time interval, the speech likelihood sequence is, for example, 0, 0, 0, 1,.

  A sequence of correct values of state transitions included in the learning data 111a is in each time interval of a speech likelihood sequence (hereinafter referred to as “speech likelihood sequence for learning data”) obtained from the acoustic signal for learning data described above. It is obtained by assigning correct values of state transitions for voice states and non-voice states. The correct value of this state transition will be described in detail. For example, suppose a case where “Today is a good weather” is read out. At this time, even if the main part (today) and the predicate (good weather) are continuously read as shown in FIG. 3A, the main part and the predicate are spaced apart as shown in FIG. 3B. Even if it is read out, it aims to estimate the entire “Today is good weather” as one true speech segment from the input speech likelihood sequence corresponding to the input acoustic signal. That is, when the main part and the predicate are read out at intervals as shown in FIG. 3B, the corresponding input speech likelihood series has a short non-speech interval between “today” and “good weather”. It becomes a series representing this. For example, when the input speech likelihood sequence is a binary sequence representing a speech interval and a non-speech interval, this input speech likelihood sequence is a time interval of a non-speech interval between “Today” and “Good weather”. (FIG. 3C). In this embodiment, a true speech segment that absorbs vibration of speech likelihood due to a short non-speech segment sandwiched between speech segments is appropriately estimated. On the other hand, even when sudden noise or the like is mixed between non-speech intervals, a true non-speech interval in which the vibration of speech likelihood due to the noise is absorbed is appropriately estimated. Note that the “true speech interval” means a time interval from when the first speech is observed in one utterance interval until the last speech is no longer observed. “True non-speech segment” means a time segment other than the true speech segment. The “speech segment” means a time segment in which a coherent utterance was made, such as “Today is a good weather”. “Speech start” means the start of “speech segment”, and “utterance end” means the end of “speech segment”.

In order to perform such estimation, a state transition model 123a is obtained that obtains an estimation result of state transition in each time interval for the speech state and the non-speech state of the input sound signal by using the input speech likelihood sequence as an input. In this embodiment, the state transition from the non-voice state to the sound state (start transition state) and the state transition from the sound state to the non-voice state (terminal transition state) are emphasized. In addition, a special label is assigned to each of the start transition state and the end transition state. For example, the following labels are assigned to each state transition.
Label “0”: State transition from non-voice state to non-voice state (non-voice state)
Label “1”: State transition from voice state to voice state (voice state)
Label “2”: State transition from non-voice state to voice state (start transition state)
Label “3”: State transition from voice state to non-voice state (terminal transition state)
Note that the speech state is a state corresponding to a true speech interval, and the non-speech state is a state corresponding to a true non-speech interval.

FIG. 2 shows a state transition diagram of the state transition model 123a when such a label is given. The state transition model 123a receives the input speech likelihood sequence as an input, loops the non-speech state and the speech state, and corresponds to the labels in the non-speech state, the speech state, the start transition state, and the end transition state, respectively. The likelihood of. In this case, the pattern of the learning data has a time interval corresponding to the speech likelihood sequence for learning data,
(1) whether or not it contains a true speech segment
(2) includes a true speech segment and a true non-speech segment immediately before
(3) include a true speech segment and a true non-speech segment immediately after
Can be subdivided.

Pattern 1: When a time interval corresponding to a speech likelihood sequence for learning data includes a true speech interval and true non-speech intervals immediately before and immediately after this, this is the most common pattern. In this case, the following labels are given to each time interval corresponding to the speech likelihood sequence for learning data.
-Time interval from the start of speech to the end of non-speech state: non-speech state “0”
-Time interval for transition from the end of the non-speech state to the speech state: Start transition state “2”
-Time period from the beginning to the end of the voice state: voice section “1”
-Time period for transition from the end of the voice state to the non-voice state: the end transition state “3”
-Time interval from the beginning (utterance end) to the end (next utterance start) of the non-speech state: non-speech state “0”
That is, the sequence of correct values of state transitions in each time interval corresponding to the speech likelihood sequence is 0,..., 0, 2, 1,..., 1, 3, 0,. For example, even when the main part and the predicate are read out continuously as shown in FIG. 3A, or even when the main part and the predicate are read out at intervals as shown in FIG. A series of correct values of state transitions in a time interval (a series of labels) are 0,..., 0, 2, 1,..., 1, 3, 0,. Normally, two or more time intervals in the start transition state do not continue, and two or more time intervals in the end transition state do not continue. However, two or more of these time intervals may be continuous.

Pattern 2: When the time interval corresponding to the speech likelihood sequence for learning data includes a true speech segment and a true non-speech segment immediately after it, but does not include a non-speech segment immediately before the speech segment. This time interval is regarded as the time interval of the start transition state. That is, the following labels are given to each time interval corresponding to the speech likelihood sequence for learning data.
-First time section: Start transition state “2”
-Time interval from the next time interval to the end of the audio state: audio interval “1”
-Time period for transition from the end of the voice state to the non-voice state: the end transition state “3”
-Time interval from the beginning (utterance end) to the end (next utterance start) of the non-speech state: non-speech state “0”
That is, the sequence of correct values of state transitions in each time interval corresponding to the speech likelihood sequence is 2, 1,..., 1, 3, 0,.

Pattern 3: When the time interval corresponding to the speech likelihood sequence for learning data includes a true speech segment and a true non-speech segment immediately before it, but does not include a non-speech segment immediately after the speech segment. Is regarded as the time interval of the terminal transition state. That is, the following labels are given to each time interval corresponding to the speech likelihood sequence for learning data.
-Time interval from the start of speech to the end of non-speech state: non-speech state “0”
-Time interval for transition from the end of the non-speech state to the speech state: Start transition state “2”
-Time interval from the beginning of the audio state to immediately before the last time interval: Audio interval “1”
-Last time interval: terminal transition state “3”
That is, the sequence of correct values of state transitions in each time interval corresponding to the speech likelihood sequence is 0,..., 0, 2, 1,.

Pattern 4: When the time interval corresponding to the speech likelihood sequence for learning data includes a true speech interval but does not include a non-speech interval before and after that, all the time intervals corresponding to the speech likelihood sequence for learning data are all This is the case of a true speech segment. In this case, the following labels are given to each time interval corresponding to the speech likelihood series.
・ Time section from the beginning to the end: Voice section “1”
That is, the sequence of correct values of state transitions in each time interval corresponding to the speech likelihood sequence is 1,.

Pattern 5: When the time interval corresponding to the speech likelihood sequence does not include a true speech interval In this case, the following labels are given to each time interval corresponding to the speech likelihood sequence.
・ Time section from the beginning to the end: Voice section “0”
That is, the sequence of correct values of state transitions in each time interval corresponding to the speech likelihood sequence is 0,.

  Note that the learning data 111a may include a series of correct values of all the state transitions of the above patterns 1 to 5, or may include only a series of correct values of some of the state transitions. . In addition, the time interval corresponding to the speech likelihood sequence may include a plurality of true speech intervals. The learning data 111a in this case consists of a combination of series of correct values of the state transitions of the above patterns 1 to 5. For example, as illustrated in FIGS. 4A and 4B, when a true non-speech segment is included before and after the first true speech segment I, and no non-speech segment exists after the second speech segment II, The learning data 111a includes a speech likelihood sequence of the first true speech interval I and a sequence of correct values 0,..., 0, 2, 1,. ,..., 0, and the second true speech interval I speech likelihood sequence and the corresponding sequence value 0 of the pattern 1 corresponding to state transition 0,..., 0, 2, 1,. 1 and 3 are included.

  The labeling described above is not limited to whether each time section of the acoustic signal for learning data is a speech section or a non-speech section, but also a speech section and a non-speech section as a speech section of the learning data acoustic signal. It is necessary to carry out based on (speech segment and non-speech segment associated with each utterance segment). Various methods are conceivable for applying such a label. In the first method, a person views each utterance included in an acoustic signal for learning data, observes the waveform, and gives an accurate label. The second method is a method of automatically assigning the above-mentioned label by using a known recognition decoder. However, in the second method, an erroneous label may be given to some learning data. Even in such a case, there is no major problem as long as the label is correctly assigned as the entire learning data. The third method is a compromise method in which a label is first assigned by the above-described decoder and then manually checked for errors.

  The learning unit 112 learns the state transition model 123a represented by the state transition diagram of FIG. 2 using the learning data 111a generated as described above. This state transition model 123a is a known method (for example, Reference 1: Hochreiter, S., & Schmidhuber, J., “Long Short-Term Memory,” Neural Computation, 9 (8), 1735-1780, 1997). Can be expressed and learned using The technique of Reference 1 can hold a very long series of information, can construct a state transition model 123a in consideration of the state of a farther position and the speech likelihood, and is very Useful. In addition to the method described in Reference 1, it is possible to use a model (such as RNN or HMM) that can handle time series. However, in the case of the HMM, the future state cannot be taken into consideration due to its structure.

<Voice section detection device 12>
Next, the speech section detection device 12 of this embodiment will be described. As illustrated in FIG. 1B, the speech segment detection device 12 of the present exemplary embodiment includes an input unit 121, a speech segment detection unit 122, a storage unit 123, an estimation unit 124, and an output unit 126.

<Storage unit 123>
The storage unit 123 stores the state transition model 123a output from the model learning device 11 as described above. The state transition model 123a may be stored in the storage unit 123 before the speech section detection by the speech section detection device 12 is started, or each time a new state transition model 123a is output from the model learning device 11. May be stored in the storage unit 123.

<Input unit 121>
The input unit 121 receives an input acoustic signal to be detected as a voice segment. The input acoustic signal to be detected as a voice segment is a time-series digital acoustic signal divided for each predetermined time segment. The input sound signal may be an analog sound signal observed with a microphone or the like obtained by AD conversion at a predetermined sampling frequency, or may be an arbitrary digital sound signal created in advance. Note that the length of the time interval of the input acoustic signal is preferably the same as or approximate to the length of the time interval of the acoustic signal for learning data described above. The input acoustic signal is sent to the voice segment detection unit 122.

<Audio section detection unit 122>
The voice section detection unit 122 extracts acoustic information and analysis information from the input acoustic signal, and applies a known VAD to the extracted information (for example, see Non-Patent Document 1 etc.), thereby each time section of the input acoustic signal. An input speech likelihood sequence corresponding to the speech likelihood is obtained and output. Examples of acoustic information and analysis information used for VAD are as described above. An example of the input speech likelihood sequence is also as described above, but the type of the input speech likelihood sequence is the same as the type of the speech likelihood sequence for learning data. For example, when the speech likelihood sequence for learning data is a speech likelihood sequence for each time interval, the input speech likelihood sequence is also a speech likelihood sequence for each time interval. The input speech likelihood sequence is sent to the estimation unit 124.

<Estimation unit 124>
The estimation unit 124 applies the transmitted input speech likelihood sequence to the state transition model 123a read from the storage unit 123, and estimates state transitions in each time interval for the speech state and the non-speech state of the input acoustic signal. Obtain and output the result (post-filter processing). Examples of state transition estimation results include the likelihood of the non-speech state (label “0”), the likelihood of the speech state (label “1”), and the likelihood of the start transition state (label “2”) in each time interval. It is a series of four likelihoods consisting of the likelihood and the likelihood of the terminal transition state (label “3”). Alternatively, only the likelihood sequence of the non-speech state (label “0”) in each time interval may be the estimation result of the state transition, or only the likelihood sequence of the speech state (label “1”) in each time interval. May be the state transition estimation result. Alternatively, a sequence of the likelihood of the non-speech state (label “0”) and the likelihood of the speech state (label “1”) in each time interval may be used as the state transition estimation result. In addition, a series of values representing a state having the greatest likelihood in each time interval may be used as the state transition estimation result.

<Output unit 126>
The state transition estimation result output from the estimation unit 124 is sent to the output unit 126. The output unit 126 outputs the state transition estimation result as a speech segment estimation result.

[Second Embodiment]
The second embodiment is a modification of the first embodiment. In this embodiment, the result obtained by performing post-processing on the estimation result of the state transition output from the estimation unit 124 is used as the speech segment estimation result. Below, it demonstrates centering around difference with 1st Embodiment, and it cites the same reference number about the already demonstrated matter, and simplifies description.

<Model learning device 11>
The same as in the first embodiment.

<Audio section detection device 22>
As illustrated in FIG. 1B, the speech segment detection device 22 of this embodiment includes an input unit 121, a speech segment detection unit 122, a storage unit 123, an estimation unit 124, a post-processing unit 225, and an output unit 226. Details of the post-processing unit 225 and the output unit 226, which are the differences from the first embodiment, will be described below.

<Post-processing unit 225>
Unlike the first embodiment, the estimation result of the state transition output from the estimation unit 124 is sent to the post-processing unit 225. The post-processing unit 225 performs predetermined post-processing on the sent state transition estimation result to obtain and output a speech section estimation result. For example, as the estimation result of the state transition, the likelihood of the non-speech state (label “0”), the likelihood of the speech state (label “1”), and the likelihood of the start transition state (label “2”) in each time interval , And the likelihood sequence of the terminal transition state (label “3”), the post-processing unit 225 uses these four likelihood sequences to represent a representative state (non-voice state, (Speech state, start transition state, or end transition state) may be selected, and a series of representative states of each time section obtained thereby may be output as a speech section estimation result. For example, the post-processing unit 225 may select and output a state having the greatest likelihood among the four transmitted states for each time interval. That is, the post-processing unit 225 may convert the four likelihood sequences sent to the maximum likelihood state sequence and output the result. For example, n = 0,..., N−1 is an identifier corresponding to a time interval, N is a positive integer, the likelihood of a non-speech state in time interval n is P _{n, 0,} and the likelihood of a speech state is _Let P _{n, 1} be the likelihood of the start transition state, P _{n, 2,} and let the likelihood of the end transition state be P _{n, 3} . In this case, the post-processing unit 225 uses the four state likelihood sequences ( _{Pn, 0} , _{Pn, 1} , _{Pn, 2} , _{Pn, 3} ) sent to the maximum likelihood state for each time interval n. it may be output by converting the sequence s _n.

ただし、α_ｉは尤度Ｐ_ｎ，ｉに与えられる重みである。例えば、α_ｉは０よりも大きな正値である。あるいは、特定の尤度Ｐ_ｎ，ｉに与えられる重みを０にしてもよい。例えば、α_２＝α_３＝０とすれば、始端遷移状態や終端遷移状態が音声区間推定結果として選択されることを避けることができる。 Alternatively, the post-processing unit 225 emphasizes or weakens the likelihood of specific states of the four state likelihood sequences ( _{Pn, 0} , _{Pn, 1} , _{Pn, 2} , _{Pn, 3} ). the likelihood sequence obtained by may be converted and output to the maximum likelihood state sequence s _n. For example, the post-processing unit 225 may convert the likelihood sequence (P _{n, 0} , P _{n, 1} , P _{n, 2} , P _{n, 3} ) into the following maximum likelihood state sequence s _n and output it. .

Here, α _i is a weight given to the likelihood P _{n, i} . For example, α _i is a positive value larger than 0. Alternatively, the weight given to the specific likelihood P _{n, i} may be zero. For example, if α ₂ = α ₃ = 0, it is possible to avoid that the start transition state or the end transition state is selected as the speech section estimation result.

  In addition, even if the post-processing unit 225 applies a known correction method generally performed in VAD to the sent state transition estimation result, the result obtained thereby is output as a speech section estimation result. Good. For example, when only the likelihood sequence of the speech state (label “1”) in each time interval is sent as the estimation result of the state transition, the post-processing unit 225 determines the likelihood of the speech state in each time interval and a predetermined threshold value. May be output, and speech segment estimation results corresponding to the series of comparison results may be output. For example, the post-processing unit 225 may output a speech section estimation result in which a time section having a speech state likelihood equal to or greater than a threshold is set as a speech section and the other time sections are set as non-speech sections. Even when only the likelihood sequence of the non-speech state (label “0”) in each time interval is sent as the estimation result of the state transition, the post-processing unit 225 determines the likelihood of the non-speech state in each time interval and a predetermined value. You may compare with a threshold value and output the audio | voice area estimation result corresponding to the series of the comparison result. In addition, the post-processing unit 225 is a publicly known method that is generally performed in the VAD on the likelihood sequence obtained by emphasizing or weakening the likelihood of a specific state in the state transition estimation result sent. A correction method may be applied and a result obtained thereby may be output as a speech section estimation result. The post-processing unit 225 may use other known techniques for improving the accuracy of voice segment detection.

<Output unit 226>
The speech section estimation result output from the estimation unit 224 is sent to the output unit 226. The output unit 226 outputs this speech segment estimation result.

[Third Embodiment]
The third embodiment is a modification of the first and second embodiments. In the present embodiment, the model learning device generates a plurality of state transition models, the speech section detection device applies the input speech likelihood sequence to the plurality of state transition models, and the input sound for each of the plurality of state transition models. The estimation result of the state transition in each time interval for the speech state and the non-speech state of the signal is obtained, and the most probable estimation result in each time interval is selected from the obtained state transition estimation results. In the following description, differences from the first and second embodiments will be mainly described, and the description will be simplified by citing the same reference numerals for the matters already described.

<Model learning device 31>
As illustrated in FIG. 1A, the model learning device 31 of the present embodiment includes a storage unit 311 that stores learning data 311a and a learning unit 312 that learns the state transition model 323a.

  The learning data 311a of the present embodiment includes a speech likelihood sequence corresponding to the speech likelihood of each time interval of the acoustic signal for a plurality of types of learning data, and each time interval for the speech state and the non-speech state of the acoustic signal. A plurality of sets of correct value series of state transitions are included. These sets are classified according to the type of acoustic signal for learning data. For example, the learning data 311a includes a plurality of sets of speech likelihood sequences corresponding to acoustic signals for learning data obtained in a plurality of types of environments and sequences of correct values of state transitions, and these sets correspond to each other. Each environment is classified. This class is c = 0,..., C-1 (where C is an integer equal to or greater than 2), and a set of a speech likelihood sequence belonging to class c included in the learning data 311a and a sequence of correct values of state transitions Is denoted as D (c). That is, the learning data 311a includes a set D (0), ..., D (C-1).

  The learning unit 312 uses the learning data 311 a read from the storage unit 311, and inputs an input speech likelihood sequence corresponding to the speech likelihood of each time interval of the input acoustic signal as an input for the speech state and non-speech state of the input acoustic signal. A plurality of state transition models 323a for obtaining estimation results of state transitions in each time interval are obtained and output. That is, the learning unit 312 uses the set D (c) included in the learning data 311a to generate a state transition model M (c) corresponding to the class c in the same manner as in the first embodiment, and a plurality of state transitions The models M (0),..., M (C-1) are output as the state transition model 323a. Others are the same as the first embodiment.

<Voice section detection device 32>
As illustrated in FIG. 1B, the speech segment detection device 32 of this embodiment includes an input unit 121, a speech segment detection unit 122, a storage unit 323, an estimation unit 324, a post-processing unit 325, and an output unit 226. Details of the storage unit 323, the estimation unit 324, and the post-processing unit 325, which are different from the first and second embodiments, will be described below.

<Storage unit 323>
The storage unit 323 stores a plurality of state transition models 323a output from the learning unit 312. The state transition model 323a may be stored in the storage unit 323 before the speech section detection by the speech section detection device 32 is started, or each time a new state transition model 323a is output from the model learning device 31. May be stored in the storage unit 323.

<Estimation unit 324>
The estimation unit 324 receives the input speech likelihood sequence output from the speech segment detection unit 122. The estimation unit 324 applies this input speech likelihood sequence to the plurality of state transition models 323a read out from the storage unit 323, and for each of the plurality of state transition models 323a, the speech state and non-speech state of the input acoustic signal. Obtain and output estimation results of state transitions in each time interval. That is, the estimation unit 324 applies the input speech likelihood sequence to each state transition model M (c) (where c = 0,..., C−1), and performs state transition in each time interval for each class c. An estimation result R (c) is obtained and output. The estimation result R (c) for each class c is sent to the post-processing unit 325. Others are the same as the first and second embodiments.

<Post-processing unit 325>
The post-processing unit 325 selects the most probable result in each time interval among the results corresponding to the estimation result R (c) sent for each class c output from the estimation unit 324. For this selection, a known technique for improving the accuracy of voice segment voice detection can be used. For example, when the voice section detection unit 122 performs VAD on the input acoustic signals obtained in a plurality of very different environments, the post-processing unit 325 is referred to the reference document 2 (RA Jacobs, MI Jordan, SJ Nowlan, and GE). Hinton. “Adaptive mixtures of local experts.” Neural Computation, 3: 79-87, 1991.) The most probable estimation result may be selected. That is, the post-processing unit 325 estimates the estimation result R (c ′) corresponding to the environment in which the input acoustic signal is obtained from the results corresponding to the estimation results R (0),. ) Is selected as the most probable result. Note that the result corresponding to the estimation result R (c) may be the estimation result R (c) itself, or the processing of the post-processing unit 225 of the second embodiment is performed on the estimation result R (c). It may be obtained. The post-processing unit 325 may output the selected most probable result as a speech segment estimation result, or the speech obtained by performing the processing of the post-processing unit 225 of the second embodiment on the most probable selected result. The section estimation result may be output. The speech section estimation result is sent to the output unit 226. The subsequent processing is the same as in the second embodiment.

[Characteristics of Embodiment Method]
As described above, in each embodiment, the accuracy of the VAD can be improved without performing manual tuning in various environments, and both cost reduction and accuracy improvement can be realized. In addition, even in a single environment, it is possible to absorb utterance blurring and perform speech segment detection with higher accuracy than existing methods. Furthermore, since the configuration of the estimation unit using the state transition model is a post-filter cascaded to the VAD, it can be used in combination with other VAD accuracy improvement methods. For example, further performance improvement can be expected by performing this post-filter processing after applying the conventional hangover method.

  Conventionally, there has been a VAD using a state transition model that outputs a voice state / non-voice state with an acoustic feature amount as an input. However, the acoustic feature amount varies greatly depending on the environment, the model becomes enormous, and a large amount of learning data is required. On the other hand, the state transition model of each embodiment obtains a state transition estimation result using the speech likelihood sequence after VAD processing as an input feature amount. The speech likelihood sequence is less subject to environmental variation than the acoustic feature amount. Therefore, the state transition model of each embodiment can be created with a small amount of learning data, and operates robustly as a post-filter of various VAD techniques. Furthermore, in each embodiment, a special label is given to the start end (start transition state) and the end (end transition state) of the speech section. As a result, information as “section” can be learned more strongly, and it is difficult to be influenced by the appearance of a temporary voice state frame due to sudden noise or the appearance of a temporary non-voice state due to consonant, breathing, etc. .

[Other variations]
In addition, this invention is not limited to the above-mentioned embodiment. For example, in the first, second, and third embodiments, an input acoustic signal is input to the speech segment detection devices 12, 22, and 32, and the speech segment detection unit 122 corresponds to the speech likelihood of each time segment of the input acoustic signal. The input speech likelihood sequence was obtained. However, an input speech likelihood sequence may be generated from the input acoustic signal outside the speech segment detection devices 12, 22, and 32, and this input speech likelihood sequence may be input to the speech segment detection devices 12, 22, and 32. In this case, the voice section detection unit 122 may be omitted from the voice section detection devices 12, 22, and 32.

  The various processes described above are not only executed in time series according to the description, but may also be executed in parallel or individually as required by the processing capability of the apparatus that executes the processes. Needless to say, other modifications are possible without departing from the spirit of the present invention.

  Each of the above devices is a general-purpose or dedicated computer including, for example, a processor (hardware processor) such as a CPU (central processing unit) and a memory such as random-access memory (RAM) and read-only memory (ROM). Is configured by executing a predetermined program. The computer may include a single processor and memory, or may include a plurality of processors and memory. This program may be installed in a computer, or may be recorded in a ROM or the like in advance. In addition, some or all of the processing units are configured using an electronic circuit that realizes a processing function without using a program, instead of an electronic circuit (circuitry) that realizes a functional configuration by reading a program like a CPU. May be. An electronic circuit constituting one device may include a plurality of CPUs.

  When the above configuration is realized by a computer, the processing contents of the functions that each device should have are described by a program. By executing this program on a computer, the above processing functions are realized on the computer. The program describing the processing contents can be recorded on a computer-readable recording medium. An example of a computer-readable recording medium is a non-transitory recording medium. Examples of such a recording medium are a magnetic recording device, an optical disk, a magneto-optical recording medium, a semiconductor memory, and the like.

  This program is distributed, for example, by selling, transferring, or lending a portable recording medium such as a DVD or CD-ROM in which the program is recorded. Furthermore, the program may be distributed by storing the program in a storage device of the server computer and transferring the program from the server computer to another computer via a network.

  A computer that executes such a program first stores, for example, a program recorded on a portable recording medium or a program transferred from a server computer in its own storage device. When executing the process, the computer reads a program stored in its own storage device, and executes a process according to the read program. As another execution form of the program, the computer may read the program directly from the portable recording medium and execute processing according to the program, and each time the program is transferred from the server computer to the computer. The processing according to the received program may be executed sequentially. The above-described processing may be executed by a so-called ASP (Application Service Provider) type service that realizes a processing function only by an execution instruction and result acquisition without transferring a program from the server computer to the computer. Good.

  The processing functions of the apparatus are not realized by executing a predetermined program on a computer, but at least a part of these processing functions may be realized by hardware.

  The present invention can be used, for example, for voice segment detection in the previous stage of voice recognition processing or voice dialogue processing. When the present invention is applied to these, it is possible to improve the detection accuracy of the voice section and perform the subsequent voice recognition and voice dialogue with higher accuracy.

11, 31 Model learning device 12, 22, 32 Voice section detection device

Claims (8)

  Learning including a set of a speech likelihood sequence corresponding to the speech likelihood of each time interval of the acoustic signal and a sequence of correct values of state transitions in each time interval for the speech state and the non-speech state of the acoustic signal Using the data, the input speech likelihood sequence corresponding to the speech likelihood of each time interval of the input acoustic signal is input, and the state transition estimation result in each time interval for the speech state and non-speech state of the input acoustic signal A model learning device for learning a state transition model to be obtained. The model learning device according to claim 1,
The state transition includes a state transition from a speech state to a speech state, a state transition from a non-speech state to a non-speech state, a state transition from a speech state to a non-speech state, and a state transition from a non-speech state to a speech state. One of the model learning devices. A state transition model that obtains an estimation result of state transition in each time interval for a speech state and a non-speech state of the input sound signal by using an input speech likelihood sequence corresponding to the speech likelihood of each time interval of the input acoustic signal as an input A storage unit for storing
Applying the input speech likelihood sequence corresponding to the speech likelihood in each time interval of the input acoustic signal to the state transition model, the state transition in each time interval for the speech state and the non-speech state of the input acoustic signal An estimation unit that obtains and outputs an estimation result;
A speech section detecting device having A plurality of states for obtaining an estimation result of state transition in each time interval for the speech state and the non-speech state of the input acoustic signal by using an input speech likelihood sequence corresponding to the speech likelihood of each time interval of the input acoustic signal as an input A storage unit for storing the transition model;
The input speech likelihood sequence corresponding to the speech likelihood of each time interval of the input acoustic signal is applied to a plurality of the state transition models, and the speech state of the input acoustic signal and the non-state for each of the plurality of state transition models. An estimation unit that obtains and outputs an estimation result of the state transition in each time interval for the speech state;
Out of the results corresponding to the state transition estimation results output from the estimation unit, a post-processing unit that selects the most probable result in each time interval;
A speech section detecting device having   Learning including a set of a speech likelihood sequence corresponding to the speech likelihood of each time interval of the acoustic signal and a sequence of correct values of state transitions in each time interval for the speech state and the non-speech state of the acoustic signal Using the data, the input speech likelihood sequence corresponding to the speech likelihood of each time interval of the input acoustic signal is input, and the state transition estimation result in each time interval for the speech state and non-speech state of the input acoustic signal A model learning method for learning a state transition model to be obtained.   A state transition model that obtains an estimation result of state transition in each time interval for a speech state and a non-speech state of the input sound signal by using an input speech likelihood sequence corresponding to the speech likelihood of each time interval of the input acoustic signal as an input The input speech likelihood sequence corresponding to the speech likelihood of each time interval of the input acoustic signal is applied to the estimation result of the state transition in each time interval for the speech state and the non-speech state of the input acoustic signal. A method for detecting and outputting speech segments.   A plurality of states for obtaining an estimation result of state transition in each time interval for the speech state and the non-speech state of the input acoustic signal by using an input speech likelihood sequence corresponding to the speech likelihood of each time interval of the input acoustic signal as an input The input speech likelihood sequence corresponding to the speech likelihood of each time interval of the input acoustic signal is applied to the transition model, and the speech state and the non-speech state of the input acoustic signal for each of the plurality of state transition models A speech interval detection method for obtaining an estimation result of a state transition in each time interval and selecting a most probable result in each time interval from results corresponding to the estimation result of the state transition.   A program for causing a computer to function as the model learning device according to claim 1 or 2 or the speech section detection device according to claim 3 or 4.

Images