JP6645995B2 - Voiced pause detection device, its method, and program - Google Patents

Description

  The present invention relates to an apparatus, a method, and a program for detecting a voiced pause expression frequently appearing in naturally uttered speech.

  In recent years, the application of AI technology is expected in various fields, and “voice”, which is the most natural tool for humans, is regarded as an important input channel to these AIs.

  On the other hand, recognizing natural voices (natural utterances) that humans usually exchange is a very difficult task because of various unclear utterances included in the utterances.

  Indistinct speech is a factor that hinders the understanding of speech for acoustic or linguistic reasons, and the most significant factor is a phenomenon called voiced pause.

  Voiced pause is a word spoken by the speaker to think about the next word to speak or to retain the right to speak, and to use only expressions such as "ah" or "um" called filler and "early spring". And the expression that extends the middle or end of a word.

To detect such a voiced pause, a detection method using acoustic features has been proposed by Goto et al. (See Non-Patent Document 1), and Ogata et al. Also studied to improve the recognition rate. (See Non-Patent Document 2). FIG. 1 shows an outline of the processing according to Non-Patent Documents 1 and 2. In Non-Patent Documents 1 and 2, changes in two types of acoustic features (fundamental frequency (hereinafter, also referred to as F0) and spectrum envelope) are obtained from audio signals (S90, S91, S92), and voiced pause P _f (t) is obtained (S93). As long as the voiced pause-likeness P _f (t) is larger than a certain threshold A _each (S94), P _f (t) is added (S95), and this sum sum becomes larger than another threshold A _sum If the frame has been paused (S97), it is determined that the frame is a voice pause (S98), and if it is small (S97), it is determined that the frame is not a voice pause (S99). This sum sum is reset to 0 when P _f (t) falls below A _each (S94, S96).

  It has become clear that even non-native speakers can identify the occurrence of the lingering phenomenon peculiar to natural language with a higher accuracy than the chance rate (the correct answer rate when choosing the correct answer at random). Reference 3), suggesting that detection is possible with common acoustic features independent of language.

Goto et al., “A Real-Time Detection System for Stagnation in Natural Speech”, IPSJ Speech and Language Division Research Group, 99-SLP-27-2, Vol.99, No.64, July, 1999. Ogata et al., “Improvement Method of Speech Recognition Based on Automatic Detection of Voiced / Unvoiced Pauses”, IPSJ Research Report, 2006-SLP-62, 2006. R. Carlson, J. Hirschberg, “Cross-Cultural Perception of Discource Phenomena”, INTERSPEECH2009, pp. 1723-1726, 2009.

  However, Non-Patent Literatures 1 and 2 have a very low detection rate for conference speech in which spontaneous speech is very likely to occur, such as brainstorming and meetings within a group. This is because when the likelihood of a voiced pause changes improperly due to a sudden factor, it is extremely likely that a section that is not a voiced pause will be erroneously detected rather than detection omission of a correct answer section. Therefore, it is necessary not only to detect a voiced pause section with high accuracy, but also to make sure that a normal speech section is not mistaken for a voiced pause section.

  An object of the present invention is to provide a voiced pause detection device, a method thereof, and a program, which suppress erroneous detection as compared with the related art and have high detection accuracy.

To solve the above problem, according to one embodiment of the present invention, a voiced pause detection device includes a voice speed calculation unit that calculates a voice speed v (m) of voice data, a voice speed v and a voiced pause continuation length. A distribution management unit that manages the probability distribution f _{v, L} with L as a variable and obtains the probability distribution f _{v (m), L} corresponding to the speech speed v (m), and the fundamental frequency of the voice data to be detected. The amount of change, the voiced pause likelihood calculated using the amount of change in the spectral envelope of the audio data to be detected, and the probability distribution _{fv (m), L} or the probability distribution _{fv (m-1), L} And a voiced pause detection unit that detects a voiced pause section using

To solve the above problem, according to another aspect of the present invention, a voiced pause detection method includes a voice speed calculating step of calculating a voice speed v (m) of voice data; and a voice speed v and a voiced pause continuation. A distribution management step of managing the probability distribution f _{v, L} with the length L as a variable and obtaining the probability distribution f _{v (m), L} corresponding to the speech speed v (m); and the fundamental frequency of the audio data to be detected. And the voiced pause likelihood calculated using the change in the spectral envelope of the voice data to be detected, and the probability distribution _{fv (m), L} or the probability distribution _{fv (m-1), L} A voiced pause detection step of detecting a voiced pause section using

  According to the present invention, there is an effect that erroneous detection can be suppressed and detection accuracy can be improved as compared with the related art.

The figure which shows the processing flow of the voiced pause detection apparatus of a prior art. FIG. 2 is a functional block diagram of the voiced pause detection device according to the first embodiment. FIG. 4 is a diagram illustrating an example of a processing flow of the voiced pause detection device according to the first embodiment. FIG. 4 is a diagram for explaining a probability distribution according to the first embodiment. FIG. 4 is a diagram illustrating an example of a functional block diagram of a feature amount extraction unit according to the first embodiment. FIG. 4 is a diagram for explaining a likelihood calculation unit according to the first embodiment. FIG. 5 is a diagram illustrating an example of a functional block diagram of a voiced pause detection unit according to the first embodiment. FIG. 4 is a diagram for explaining a voiced pause correction unit according to the first embodiment. FIG. 7 is a diagram for explaining processing content of a speech speed calculation unit according to the first embodiment. FIG. 4 is a diagram for explaining a distribution management unit according to the first embodiment.

Hereinafter, embodiments of the present invention will be described. In the drawings used in the following description, components having the same functions and steps for performing the same processing are denoted by the same reference numerals, and redundant description will be omitted. In the following description, the symbol “ ^→ ” used in the text should be written immediately above the immediately preceding character, but is written immediately after the character due to restrictions on the text notation. In the formula, these symbols are described in their original positions. The processing performed for each element of a vector or matrix is applied to all elements of the vector or matrix unless otherwise specified.

<Points of the first embodiment>
(1) By changing the voiced pause duration based on the voice speed, voiced pause detection can be performed in accordance with the voice speed of the speaker.
(2) When the number of frames in which the likelihood of voiced pause P _f (t) is larger than the threshold value A _each continues for a predetermined threshold A or more, the section can be determined as a voiced pause section. At this time, the higher the speech speed (faster), the shorter the voiced pause continuation length tends to be (short). Therefore, the threshold value A can be set so as to decrease as the speech speed increases. However, even when the number of frames is smaller than the threshold A, there may be a voiced pause. Therefore, in the present embodiment, instead of simply comparing the threshold A with the number of frames whose voiced pause likelihood P _f (t) is larger than the threshold A _each , the number of frames increases as the voice pause likelihood P _f (t) increases. A probability distribution using the average of voiced pause duration and its variance is introduced so that it is easy to determine that the pause is voiced even if it is small. By comparing the product of the likelihood of voiced pause P _f (t) and the probability density with a threshold, it is possible to make a determination in consideration of the variance.

<First embodiment>
FIG. 2 is a functional block diagram of the voiced pause detection device according to the first embodiment, and FIG. 3 shows an example of the processing flow.

  The voiced pause detection device 100 receives the digital voice data p (τ) as input, and performs voiced pause section information of the digital voice data p (τ) (for example, information including the disclosure time and end time of the voiced pause section) B (m , n), performs a predetermined process on the detected voiced pause section, and outputs a processing result. Here, τ is an index indicating the sample time, m is an index indicating the m-th utterance included in the digital voice data p (τ), and (m, n) is the n-th voiced voice of the m-th utterance. This is an index indicating a pause section.

  The voiced pause detection device 100 is configured by a computer including a CPU, a RAM, and a ROM recording a program for executing the following processing, and is functionally configured as follows.

  The voiced pause detection device 100 includes a feature amount extraction unit 101, a likelihood calculation unit 102, a voiced pause detection unit 103, a voiced pause correction unit 104, a speech speed calculation unit 105, and a distribution management unit 106. Hereinafter, an outline of the processing of each unit will be described.

<Feature extraction unit 101>
The feature amount extraction unit 101 receives digital voice data (voice data for which voiced pause is to be detected) p (τ) as an input, extracts voice feature amounts from the digital voice data p (τ) (S101), and outputs it. In the present embodiment, a pitch of a sound (hereinafter, also referred to as a fundamental frequency or F0) and a general feature of a sound (hereinafter, also referred to as a spectrum envelope) are extracted as the speech feature amount. There are various prior examples of methods for calculating F0 and the spectral envelope, and any method may be used. After extracting F0, a change amount A _f (t) for each frame is obtained and output. Here, t indicates a frame time. After extracting the spectrum envelope, a variation A _s (t) for each frame is obtained and output. Thus, finally, F0 and spectral envelope of the variation _{A f (t), A s} (t) is obtained by the number of frames.

<Likelihood calculation unit 102>
The likelihood calculation unit 102 receives F0 and the change amounts A _f (t) and A _s (t) of the spectrum envelope as inputs, and uses these values to determine the likelihood of a voiced pause of each frame (hereinafter, also referred to as voiced pause likelihood say) calculates _{P f (t) (S102)} , and outputs. For example, filled pauses likelihood P _f (t) is, F0 and spectral envelope of the variation A _f (t), represented by the weighted sum of A _s (t). In addition, the expression for calculating the voiced pause likelihood includes a constant for value control.

<Voice pause detector 103>
The voiced pause detection unit 103 receives the voiced pause likelihood P _f (t) and a probability distribution f _{v (m−1), L} described later, and uses these values to generate digital voice data p (τ). Is detected (S103). The obtained voiced pause section information (for example, the start time and end time of the voiced pause section) B (m, n) is sent to the voiced pause correction unit 104, and the voiced pause continuation length (hereinafter, also referred to as average section length) L (m, n) is passed to the distribution management unit 106.

<Voice pause correction unit 104>
The voiced pause correction unit 104 receives the voiced pause section information B (m, n) as input and performs processing on the detected voiced pause section based on the voiced pause section information B (m, n) (S104). Output the processing result. The contents of the processing differ depending on the purpose.

<Speech speed calculation unit 105>
The voice speed calculation unit 105 receives the digital voice data p (τ) as input, calculates the voice speed v (m) from the digital voice data p (τ) (S105), and outputs it. The preceding example is used for the method of calculating the speech speed. For example, the speech speed can be obtained from the recognition result in the voice recognition and the length (file length) of the digital voice data p (τ).

<Distribution management unit 106>
The distribution management unit 106 manages the probability distribution _{fv, L} using the speech speed v and the voiced pause duration L as variables. The distribution management unit 106 receives the speech speed v (m) as an input, obtains a probability distribution f _{v (m), L} corresponding to the speech speed v (m) (S106), and outputs it.

The distribution management unit 106 has two major roles. One is to the voiced pause detection unit 103, from the probability distribution f _{v, L} currently held, the probability distribution f (m) corresponding to the speech speed v (m). _{v (m), L.} That is, when the speech speed v (m) is given as an argument from the speech speed calculation unit 105, the distribution management unit 106 determines the probability distribution f _{v (m), L} at the speech speed v (m) as a voiced pause detection unit 103 (See FIG. 4). The voiced pause detection unit 103 performs a voiced pause detection of the next utterance m + 1 using the given probability distribution _{fv (m), L.}

The second role is to update statistical information (probability distribution f _{v, L} ). At the start of processing, the distribution management unit 106 holds, as initial values _, statistical information (probability distribution f _{v, L} ) obtained from data prepared in advance. Here, the voiced pause detection unit 103 receives the voiced pause duration L (m, n), and the voice speed calculation unit 105 receives the voice speed v (m) _, thereby sequentially updating the probability distribution f _{v, L.} Go.

<Modification>
The voiced pause detection device 100 may not include the voiced pause correction unit 104. In that case, the voiced pause detection device 100 outputs the voiced pause section information B (m, n). For example, some processing is performed using the voiced pause section information B (m, n) by another device provided at the subsequent stage.

In the present embodiment, it is assumed that the speakers of the utterances included in the digital voice data p (τ) are the same, and that the m-th utterance and the (m + 1) -th utterance do not suddenly change in speech speed. The probability distribution f _{v (m), L} corresponding to the speech speed v (m) of the m-th utterance is obtained, and this probability distribution f _{v (m), Use L.} However, there are a plurality of speakers of the utterance included in the digital voice data p (τ), and there is a possibility that the speech speed differs for each speaker. In such a case, the following configuration is effective. A speaker identification unit is provided before the speech speed calculation unit 105 to identify the m-th speaker of the digital voice data p (τ) and output speaker information w (m). Note that w (m) is information indicating the speaker of the m-th utterance. Various existing technologies can be used as the speaker identification technology, and may be appropriately selected. For example, speaker identification can be performed online using the technique of Reference 1.
(Reference 1) Weizhong Zhu and Jason Pelecanos, "ONLINE SPEAKER DIARIZATION USING ADAPTED I-VECTOR TRANSFORMS", ICASSP 2016, pp. 5045-5049.
The voice speed calculation unit 105 receives the digital voice data p (τ) and the speaker information w (m) as input, calculates the voice speed v (m) from the digital voice data p (τ) (S105), and A combination (v (m), w (m)) of the speed v (m) and the speaker information w (m) is output. The distribution management unit 106 receives the combination (v (m), w (m)) as input, obtains a probability distribution f _{v (m), L} corresponding to the speech speed v (m), and obtains speaker information w (m). The combination (w (m), _{fv (m), L} ) of the probability distribution _{fv (m), L} is stored.

The voiced pause detection unit 103 receives a combination of the voiced pause likelihood P _f (t) and the speaker information w (m) (output value of a speaker identification unit (not shown)) as input, and outputs the speaker information w (m). To the distribution management unit 106 based on

  The distribution management unit 106 provides the probability distribution corresponding to the speaker information w (m) to the voiced pause detection unit 103. When there is a plurality of probability distributions corresponding to the speaker information w (m), the latest probability distribution is provided. With such a configuration, by managing the speech speed for each speaker, even if the speech speed differs for each speaker, erroneous detection can be suppressed and the detection accuracy can be improved as compared with the related art.

Further, in the first embodiment and the above-described modified example, the speech speed calculation unit 105 calculates the speech speed v (m) using the speech recognition result. Alternatively, the speech speed v (m) may be calculated only from the digital audio data p (τ). For example, by using the technique of Reference 2, it is possible to calculate the m-th utterance and the speech speed up to immediately before the voiced pause section candidate in real time using the RNN without performing the voice recognition processing.
(Reference 2) Yishan Jiao, Ming Tu, Visar Berisha and Julie Liss, "ONLINE SPEAKING RATE ESTIMATION USING RECURRENT NEURAL NETWORKS", ICASSP 2016.
In this case, first, the speech speed calculation unit 105 receives the digital voice data p (τ) as input, calculates the voice speed v (m) from the digital voice data p (τ) (S105), and outputs it. Next, the distribution management unit 106 obtains a probability distribution _{fv (m), L} corresponding to the speech speed v (m) (S106) and outputs it. The voiced pause detection unit 103 receives the voiced pause likelihood P _f (t) and the probability distribution f _{v (m), L} and uses these values to calculate the voiced pause included in the digital voice data p (τ). A pause section is detected (S103). That is, in the first embodiment, processing is performed in the order of voiced pause detection (S104), voiced pause correction (S104), voice recognition (S107), speech speed calculation (S105), and distribution management (S106). When the speech speed v (m) is calculated in real time using the RNN, speech speed calculation (S105), distribution management (S106), voiced pause detection (S104), voiced pause correction (S104), and speech recognition (S107) ) Can be performed in this order. With such a configuration, in the m-th utterance and the (m + 1) -th utterance, the probability distribution f _{v (m)} is appropriate even when the speech speed changes suddenly or when there are multiple speakers. _{, L} can be used to detect voiced pauses.

<Example 1>
In the present embodiment, an example will be described in which speech recognition is improved using the detection result of a voiced pause section. The voice recognition device includes a voice recognition unit 107 in addition to the configuration of the voiced pause detection device 100 (indicated by a broken line in FIG. 2).

<Feature extraction unit 101>
FIG. 5 shows an example of a functional block diagram of the feature amount extraction unit 101. The feature amount extraction unit 101 can use, for example, Non-Patent Documents 1 and 2 for calculating the change amounts of F0 and the spectrum envelope. For example, the sampling frequency of digital audio data p (τ) is 16 [kHz], the frame width is 64 [ms], and the frame shift is 10 [ms].

First, the frequency domain transforming unit 101-1 divides digital audio data p (τ) into short units called frames, and performs short-time Fourier transform (STFT). From this result, F0 is obtained by calculating the instantaneous frequency in the F0 extraction unit 101-2. The F0 change amount calculation unit 101-3 calculates the change amount of each frame time (for example, the difference between F0 of the current frame time and F0 of the immediately preceding frame time) A _f (t) from F0 in the time series.

In addition, a harmonic filter (comb filter) 101-4 that passes only the harmonic component of F0 estimated previously to the L-order component is applied to the output of the frequency domain transforming unit 101-1. Get the quantity. For these points, the spectrum envelope generation unit 101-5 performs linear interpolation, and then performs resampling at regular frequency intervals to generate a spectral envelope. Further, the spectrum envelope generation unit 101-5 shortens the calculation time by setting the upper limit of the frequency at which resampling is performed, because the high-frequency component does not have much information important for the determination of the voiced pause. For example, when the frequency upper limit is determined so that only two local peaks (formants) in human speech can be taken from the lower order, the upper limit frequency is 3,200 [Hz]. Thereafter, the spectrum envelope change amount calculation unit 101-6 obtains a change amount between frames of each dimension (for example, a difference between a value of each dimension of the current frame time and a value of each dimension of the immediately preceding frame time). Then, by adding them over all dimensions, the variation A _s (t) of the spectral envelope of each frame is calculated.

<Likelihood calculation unit 102>
FIG. 6 is a diagram for explaining the likelihood calculating unit 102. For example, the likelihood calculation processing using the F0 change amount A _f (t) and the spectrum envelope change amount A _s (t) may be performed according to Non-Patent Documents 1 and 2.

The voiced pause-likeness P _f (t) in a certain frame t is obtained by the following equation using the F0 variation A _f (t) and the spectrum envelope variation A _s (t).
Here, R is a weight, and W is a constant for controlling the magnitude of fluctuation or deformation. When R is large, F0 is emphasized, and when R is small, spectrum envelope is emphasized. When W is large, the amount of change in the feature amount is allowed to be slightly large, and when R is small, the amount of change is sensitive to the amount of change.

<Voice pause detector 103>
FIG. 7 shows an example of a functional block diagram of the voiced pause detection unit 103. The voiced pause detection unit 103 receives the probability distribution f _{v (m), L} passed from the distribution management unit 106 and the voiced pause likelihood P _f (t) passed from the likelihood calculation unit 102 as inputs. Is used to detect a voiced pause section, and a voiced pause section B (m, n) and a voiced pause duration L (m, n) are output.

First, the moving average calculation unit 103-1 calculates a moving average of m frames before and after the voiced pause likelihood P _f (t), and performs smoothing processing. As a result, it is possible to alleviate a phenomenon in which the voice pause is suddenly increased or decreased. The smoothed likelihood obtained in this manner is _defined as P _{f_avg} (t). If the window width cannot be sufficiently secured at the start and end of the audio data, the shortest end element may be copied and supplemented by the shortage.

Thereafter, the first determination unit 103-2, provided certain threshold thr _f, for each frame, there is a possibility of _P f_avg (t)> thr _f if filled pause, if _P f_avg (t) ≦ thr _f It is determined that it is not a voiced pause. For example, if P _{f_avg} (t)> thr _f , A (t) ← 1; if P _{f_avg} (t) ≦ thr _f , A (t) ← 0, and in this determination result, A (t) = 1 It is determined that one or more consecutive appearing sections may be "voiced pause sections". From each detection section, a voiced pause continuation length L (m, n) and average likelihood _{Pt, t + N-1} are calculated from the number of frames belonging to the section and the smoothed likelihood. For example, if there is a section that may be a voiced pause section consisting of N frames at frame numbers t to t + N−1, a voiced pause duration L (m, n) and average likelihood P _{t, t + N-1} is
Becomes Note that 10 [ms] is the above-described frame shift.

  A determination based on the speech speed is performed for each of the sections that may be detected as voiced pause sections in this manner.

In the section likelihood calculation section 103-3, using the probability distribution f _{v (m-1), L} passed from the distribution management section 106, the voiced pause continuation length L ( m, n) probability of density _{f v (m), L (} m, n) and the average likelihood P _t, by obtaining the product of the _{t + n-1,} the final segment likelihoods P _t (= f _{v (m), L (m, n)} × P _{t, t + N−1} ). For example, in a section [t, t + N-1] where there is a possibility of a voiced pause section, L (m, n) = 650 [ms] and _{Pt, t + N-1} = 0.80, and FIG. Given a distribution as shown in the above, the probability density f _{v (m), L (m, n)} at L = 650 [ms] in this distribution is 0.87, so the final interval likelihood Is P _t = 0.80 × 0.87 = 0.696.

The in second determination unit 103-4, thus determined interval likelihood P _t is equal to or smaller than the threshold value thr _s set, (and determined not to be filled pause period) excluding the period, interval likelihood If P _{where (t)} exceeds the threshold value thr _s, determines the section between voiced pause interval. By considering the probability density _{fv (m), L (m, n)} , the absolute value itself of the average likelihood _{Pt, t + N-1} increases or decreases. For thr _s , a variable different from the threshold thr _f used for the voiced pause likelihood of each frame should be set based on the probability density of the distribution.

  The second determination unit 103-4 outputs the voiced pause duration L (m, n) and the voiced pause section information B (m, n) when it is determined that the section is a voiced pause section.

<Voice pause correction unit 104>
In this embodiment, a correction is made to improve the recognition rate.
The simplest correction method for improving the recognition rate is "deletion" of the detected section. FIG. 8 is a diagram for explaining the voiced pause correction unit. That is, the voice (feature amount) of the frame corresponding to the detected section is cut out, and the remaining portions are connected to generate pseudo voice including no voiced pause. However, if simply cut and joined, a discontinuous point occurs at the cut part, which may adversely affect recognition.Therefore, by applying a window function, the cuts of the sections are smoothly joined, Application of an interpolation method such as insertion of a compensation frame can be considered.

  Therefore, the voiced pause correction unit 104 receives the digital voice data p (τ) and the voiced pause section information B (m, n) for the digital voice data p (τ), and performs a voice pause from the digital voice data p (τ). The portion corresponding to the section information B (m, n) is deleted, corrected voice data q (τ) is obtained and output.

<Speech speed calculation unit 105>
FIG. 9 is a diagram for explaining the processing content of the speech speed calculation unit 105. The speech speed calculation unit 105 receives the digital speech data p (τ) and the recognition result given from the speech recognition unit 107, which will be described later, and calculates the speech speed v (m) based on the length of the speech file and the number of mora of the recognition result. And output. The mora is a unit of a sound phrase. For example, the word "Kessou" is 4 mora ("Ke | Sho"). By calculating the total number of mora from the recognition result and dividing it by the length of the audio file, it is possible to determine how many mora are speaking per second (unit is [mora / s]). This is used as the speech speed v (m). The calculated speech speed v (m) is passed to the distribution management unit 106, which will be described later, and is used to determine the probability distribution used for updating the distribution and detecting the next voiced pause.

<Distribution management unit 106>
FIG. 10 is a functional block diagram of the distribution management unit 106.
The probability distribution f _{v, L} managed by the distribution management unit 106 is assumed to be a Gaussian distribution with the speech speed and the voiced pause duration as variables (see FIG. 4). Such a probability distribution f _{v, L} is stored in the probability distribution storage unit 106-2. When the voiced pause duration is represented by a variable x and the speech speed is represented by a variable y, the average μ _x , μ _{y, the} variance σ _x ² , σ _y ² , and the covariance σ _xy (= σ _yx ) The distribution can be expressed as:
Here, x ^→ = (x, y) ^t and μ ^→ = (μ _x , μ _y ) ^t (the superscript t represents transpose), and Σ represents a two-dimensional variance-covariance matrix.

Next, updating of the distribution will be described. When the speech speed v (m) and the voiced pause continuation length L (m) at the m-th utterance are given from the voice speed calculation unit 105 and the voiced pause detection unit 103, the distribution update unit 106-1 provides the probability distribution storage unit 106 The probability distribution f _{v, L} at that time is extracted from −2, and the average μ ^→ _m−1 = (μ _{x, m−1} , μ _{y, m−1} ) ^t calculated from the processing up to the m−1 utterance And the variance-covariance matrix
Is used to update the statistical information according to the following equation.
If there are R (R ≧ 2) voiced pauses in the m-th utterance, each of the voiced pause durations L (m, 1), L (m, 2),..., L (m, R) , The probability distribution f _{v, L} may be updated, or the average of R voiced pause durations L (m, 1), L (m, 2), ..., L (m, R) is voiced. The probability distribution _{fv, L} may be updated with the pause duration L (m).

After updating the probability distribution f _{v, L} , the probability distribution calculation unit 106-3 calculates the probability density distribution f _{v (m), L} at the speech speed v (m) using the input speech speed v (m). I do. The calculated probability density distribution f _{v (m), L} is used for calculating the section likelihood in the ₍ m + 1) th input utterance.

  The Gaussian distribution has good properties and is easy to handle statistical information. However, due to the nature of the distribution, there is a probability that the voiced pause duration becomes negative. In order to avoid this, it is conceivable to use a distribution such as a gamma distribution that has a value only in the domain of 0 or more.

<Speech recognition unit 107>
The speech recognition unit 107 receives the modified speech data q (τ) as input, performs speech recognition on the modified speech data q (τ) (S107), and outputs a recognition result as an output value of the speech recognition device. Output to the above-mentioned speech speed calculation unit 105.

  For example, speech recognition by a WFST decoder using an acoustic model and a language model, which is generally performed in current speech recognition, is performed. The voice used for recognition here is not the original voice data p (τ), but the corrected voice data q (τ) after the voiced pause portion has been cut off by the voiced pause correction unit 104. As a result, the influence of voiced pause can be removed, and the recognition rate can be improved. Further, by passing the obtained recognition result to the speech speed calculation unit 105, the speech speed can be calculated from the recognition result. The calculated speech speed is used for voiced pause detection of the next utterance.

<Effect>
With the above configuration, erroneous detection can be suppressed and the detection accuracy can be improved as compared with the related art. Further, the accuracy of voice recognition can be improved by using the detected voiced pause.

  The voiced pause detecting device according to the first embodiment can be applied to various techniques using voiced pauses other than the voice recognition device. For example, it can be applied to an interactive system. In that case, the dialogue system can receive the voiced pause section information and realize a dialogue in which the dialogue system returns to the user. However, in this case, since the speech speed cannot be calculated using the speech recognition result as in the present embodiment, the digital speech data p is used without using the speech recognition result as described in the modification of the first embodiment. It is desirable that the speech speed v (m) be calculated in real time only from (τ).

<Other modifications>
The present invention is not limited to the above embodiments and modifications. For example, the above-described various processes may be executed not only in chronological order as described, but also in parallel or individually according to the processing capability of the device that executes the processes or as necessary. In addition, changes can be made as appropriate without departing from the spirit of the present invention.

<Program and recording medium>
Further, various processing functions in each device described in the above embodiment and the modified examples may be realized by a computer. In this case, the processing content of the function that each device should have is described by a program. By executing this program on a computer, various processing functions of the above-described devices are realized on the computer.

  A program describing this processing content can be recorded on a computer-readable recording medium. The computer-readable recording medium may be any recording medium such as a magnetic recording device, an optical disk, a magneto-optical recording medium, a semiconductor memory, and the like.

  The distribution of the program is performed by, for example, selling, transferring, lending, or the like, a portable recording medium such as a DVD or a CD-ROM on which the program is recorded. Further, the program may be stored in a storage device of a server computer, and the program may be distributed by 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 unit. Then, when executing the processing, the computer reads the program stored in its own storage unit and executes the processing according to the read program. As another embodiment of the program, a computer may directly read the program from a portable recording medium and execute processing according to the program. Further, each time a program is transferred from the server computer to the computer, processing according to the received program may be sequentially performed. A configuration in which the above-described processing is executed by a so-called ASP (Application Service Provider) type service that realizes a processing function only by executing an instruction and acquiring a result without transferring a program from the server computer to the computer. It may be. It should be noted that the program includes information to be used for processing by the computer and which is similar to the program (such as data that is not a direct command to the computer but has properties that define the processing of the computer).

  Further, each device is configured by executing a predetermined program on a computer, but at least a part of the processing contents may be realized by hardware.

Claims (5)

A speech speed calculation unit for calculating a speech speed v (m) of voice data,
The probability distribution _fv, which manages the speech speed v and the voiced pause continuation length L as variables, and a probability distribution fv (m) corresponding to the speech speed v (m), a distribution management unit for determining _L ,
The amount of change in the fundamental frequency of the audio data to be detected, the voiced pause likelihood calculated using the amount of change in the spectral envelope of the audio data to be detected, and the probability distribution _{fv (m), L} or probability Including a voiced pause detection unit that detects a voiced pause section using the distribution _{fv (m-1), L.}
Voiced pause detection device. The voiced pause detection device according to claim 1,
The distribution management unit updates the probability distribution _{fv, L} from the voice speed v (m) and the voiced pause duration L (m) corresponding to the detected voiced pause section,
Voiced pause detection device. A speech speed calculation step of calculating a speech speed v (m) of voice data;
A distribution management step of managing a probability distribution f _{v, L} having a speech speed v and a voiced pause duration L as variables, and obtaining a probability distribution f _{v (m), L} corresponding to the speech speed v (m),
The amount of change in the fundamental frequency of the audio data to be detected, the voiced pause likelihood calculated using the amount of change in the spectral envelope of the audio data to be detected, and the probability distribution _{fv (m), L} or probability Voiced pause detection step of detecting a voiced pause section using the distribution _{fv (m-1), L} and
Voiced pause detection method. 4. The voiced pause detection method according to claim 3, wherein
A probability distribution update step of updating the probability distribution f _{v, L} from the voice speed v (m) and the voiced pause continuation length L (m) corresponding to the detected voiced pause section,
Voiced pause detection method.   A program for causing a computer to function as the voiced pause detection device according to claim 1 or 2.