JP4986028B2 - Speech recognition apparatus, utterance determination method thereof, utterance determination program, and storage medium thereof - Google Patents

Description

  The present invention relates to a speech recognition apparatus, a speech determination method thereof, a speech determination program, and a storage medium thereof, and more particularly to a speech recognition apparatus suitable for recognition of syllable emphasized speech, a speech determination method thereof, a speech determination program, and a storage medium thereof. .

  FIG. 9 is a diagram illustrating a configuration of a main part of a conventional speech recognition apparatus, which is created in advance based on an acoustic analysis unit 51 that extracts an acoustic feature amount from input speech and the extracted acoustic feature amount. And a search processing unit 52 that performs a search process according to a statistical acoustic model 53 and a language model 54 and outputs a speech recognition result.

  The acoustic analysis unit 51 cuts out a frame having a length T from the input speech and extracts an n-dimensional acoustic feature amount representing the feature. As shown in FIG. 10, this process proceeds while shifting the frame position by ΔT and is executed until the end of the voice.

  The search processing unit 52 searches for the most probable input speech among the translatable word strings defined by the language model. As the language model, either a fixed grammar model in which a word transition pattern is defined in advance or a probabilistic grammar model in which a next transitionable word is stochastically determined according to a word string determined by a certain time is used.

  For example, in the fixed grammar model shown as an example in FIG. 11, there are four words “Ito”, “Itoi”, “Imai”, and “Doi” that can be transitioned from the initial silent state “sil”. A word string is defined so as to finally transition to the silent state “sil” again via the word “is” that can only be shifted to. That is, the most likely word string is searched for “[sil] {Ito / Itoi / Imai / Doi} [sil]”.

  Regardless of whether the fixed grammar model or the probabilistic grammar model is used, the search process using the acoustic feature amount for each frame is advanced in units of phonemes obtained by further subdividing words. Each word is represented by concatenating HMM state sequences for each phoneme. FIG. 12 shows an HMM state sequence of the word “Imai”.

  The phoneme representation of “Imai” is “i / m / a / i”, but in general, an HMM state sequence depending on the preceding and following phonemes as shown in FIG. 12 is used to improve search processing performance. Here, “sil-i + m” represents an HMM state sequence when the preceding phoneme of the phoneme “i” is “sil” and the subsequent phoneme is “m”. Each HMM state is allowed to transition to itself (self-transition) and to the right-hand HMM state (LR transition), and the self-transition probability and LR transition probability are described in the acoustic model. The acoustic model describes a probability distribution for calculating the likelihood (acoustic likelihood) of each acoustic feature quantity obtained for each frame with respect to each HMM state.

  The search process calculates the sum of the transition probability and acoustic likelihood (cumulative likelihood) for each of the self-transition and LR transition for every HMM state that should be considered in that frame for each frame. This is equivalent to repeatedly selecting an apparent transition (highest cumulative likelihood) and finally determining the HMM state sequence having the highest cumulative likelihood. The algorithm for searching for the maximum likelihood path in this way is called the Viterbi algorithm.

  If the user's re-utterance is required because the recognition result is incorrect, even if the initial utterance by the user is a normal utterance that is uttered against a human being in their lives, There is often a phenomenon in which a recurrent voice becomes a syllable emphasized utterance in which each syllable is emphasized with the same intention as utterance that is easy to hear.

  FIGS. 13 and 14 respectively show the waveform of the normal utterance “Kanagawa” of the same utterance content “Kanagawa” by the same utterer and the waveform of the syllable-emphasized utterance “Kana-ga-wa”. In the syllable emphasis utterance, it can be seen that there is a silent section between syllables in the middle of utterance that cannot be seen in normal utterance, and the waveform is as if the syllable was uttered individually.

  In the syllable emphasis utterance, the utterance sections are not continuous like the normal utterance, and there are silent sections between the syllables as shown in FIG. However, in an ordinary speech recognition apparatus, as shown in FIG. 12, the HMM state sequence of each word described in the language model does not allow a transition to “sil” between syllables. The search process must proceed as if some phoneme existed in the silent section between syllables. As a result, the cumulative likelihood of the HMM state sequence corresponding to the utterance content decreases due to a decrease in the acoustic likelihood of a silent section between syllables, which may cause erroneous recognition.

  In order to deal with such technical problems, a description that allows a transition to “sil” after each syllable of the HMM state sequence of a word for a silent section between syllables of syllable-emphasized utterances is conventionally used. It is supported by adding to the model.

  Patent Document 1 adds an HMM state sequence that targets multiple syllable-enhanced utterances to an HMM state sequence that targets normal speech, such as adding a transition to silence that can be skipped as a subsequent phoneme environment to the HMM state sequence By doing so (multipathing), there is disclosed a technique for maintaining recognition performance even for syllable emphasized utterance in which silence is inserted or acoustic characteristics between syllables are changed from normal utterance.

In Japanese Patent Application Laid-Open No. 2004-228688, not only a language in which a syllable is necessarily separated after a vowel, but a syllable can be divided after any phoneme including English, as in Japanese which is the subject of the above-mentioned Patent Document 1. A technique for maintaining recognition performance for syllable-emphasized utterances by using a model that allows insertion of silence after each syllable is also disclosed for possible languages.
JP 2002-189494 A JP 2006-243123 A

  FIG. 15 is a diagram in which a transition to “sil” is added after the syllable “i” to the HMM state sequence continuous with “sil-i + m” and “i-m + a” in FIG. is there. At the end of the first half syllable “i” of two consecutive syllables “i ・ ma”, in addition to “sil-i + m”, the “sil” environment-dependent HMM state sequence such as “sil-i + sil” Or, it is allowed to transition to “sil” of one state via an HMM state sequence without post-environment dependency like “sil-i + *”. Accordingly, at the beginning of the second syllable “ma”, in addition to “i-m + a”, a transition to “sil-m + a”, which depends on the previous “sil” environment, is added.

  It is also possible to skip the transition to one state “sil” at the end of the first half syllable. The most different point between syllable-enhanced utterances and normal utterances is the existence of silent intervals between syllables. is there. As shown in FIG. 15, a number of transitions are possible in response to the difference between the syllable-emphasized utterance and the normal utterance.

  However, considering a plurality of transitions as shown in FIG. 15 for each syllable included in all words causes an increase in the amount of processing required for the search process, and there is a possibility that a delay occurs in obtaining the recognition result. . In addition, since the same language model is used even when the input is a normal utterance, the recognition performance is lowered due to the expansion of the search space due to the presence of an unnecessary HMM state sequence for syllable emphasis utterance.

  Thus, when the input to the speech recognition apparatus is syllable-weighted utterance, the possibility of misrecognition increases in search processing for normal utterance. In order to prevent adverse effects due to misrecognition, when the input is syllable-enhanced utterance, measures such as prompting the user again for normal utterance without switching the search process or switching to search process targeting syllable-enhanced utterance However, it is necessary to determine whether or not the input is syllable emphasis utterance before the search process.

  SUMMARY OF THE INVENTION An object of the present invention is to solve the above-described problems of the prior art and enable a speech recognition apparatus, an utterance determination method thereof, and an utterance determination program capable of determining whether or not an input is syllable emphasized utterance before search processing. And providing a storage medium thereof.

In order to achieve the above object, the speech recognition apparatus of the present invention is characterized in that the following means are taken.
(1) Acoustic analysis means for extracting acoustic features of the input speech, a statistical model for performing speech recognition based on the extracted acoustic features, and input based on the periodicity of the extracted acoustic features A syllable-enhanced utterance determination unit that determines whether or not the speech is syllable-enhanced utterance; and a search processing unit that executes the search process by applying the statistical model to the acoustic feature amount, and the input speech is syllable-enhanced utterance A unique voice recognition operation is executed according to the determination result of whether or not there is.
(2) It is characterized by including means for requesting the utterer to speak again when the input speech is determined to be syllable emphasized utterance.
(3) The search processing unit includes a first statistical model corresponding to an acoustic feature amount specific to normal utterance and a second statistical model corresponding to an acoustic feature amount specific to syllable emphasized utterance. The search process is executed using the second statistical model when it is determined to be utterance.

According to the present invention, the following effects are achieved.
(1) Since it is possible to determine whether the user's utterance is a normal utterance or a syllable-emphasized utterance before the search process is started, it is possible to shift to an appropriate process corresponding to the user's utterance in a short time.
(2) Whether the user's utterance is a normal utterance or a syllable-emphasized utterance is determined by paying attention to the periodicity of the acoustic feature amount, so that an accurate determination can be performed with a small processing load.
(3) When the user's utterance is determined to be syllable-weighted utterance, the user is prompted to re-utter with the normal utterance, so that accurate speech recognition based on the normal utterance is possible.
(4) A statistical model for normal utterance and a statistical model for syllable-enhanced utterance are provided, and the statistical model is used properly depending on whether the user's utterance is normal utterance or syllable-enhanced utterance. Good speech recognition is possible regardless of whether the person utters either normal speech or syllable-emphasized speech.

  DESCRIPTION OF THE PREFERRED EMBODIMENTS Hereinafter, the best embodiment of the present invention will be described in detail with reference to the drawings. FIG. 1 is a block diagram showing the configuration of the main part of a first embodiment of a speech recognition apparatus according to the present invention.

  When the input control unit 11 determines that the input is syllable emphasized utterance in the syllable emphasized utterance determination unit 13 described in detail later, for example, “ A re-generation request unit 14 that outputs “Please speak as in normal conversation once more” is included. The acoustic analysis unit 12 extracts an acoustic feature amount from the input voice.

  The syllable-enhanced utterance determination unit 13 detects the appearance periodicity of each syllable uttered at regular intervals using the acoustic feature amount extracted by the acoustic analysis unit 12, and determines whether or not the input speech is a syllable-enhanced utterance. Determine whether. The search processing unit 15 performs a search process according to the statistical acoustic model 16 and the language model 17 created in advance based on the extracted acoustic feature quantity, and outputs a speech recognition result.

  FIG. 2 is a block diagram showing the configuration of the main part of the second embodiment of the speech recognition apparatus according to the present invention, where the same reference numerals as those described above represent the same or equivalent parts.

  In the first acoustic model 16 and the first language model 17, a statistical model for performing speech recognition based on the acoustic feature amount of the normal utterance is registered. In the second acoustic model 18 and the second language model 19, a statistical model for performing speech recognition based on the acoustic feature amount of the syllable-emphasized normal utterance is registered. The statistical model selection unit 20 selects the first statistical models 16 and 17 if the input speech is a normal utterance, and selects the second statistical models 18 and 19 if the input speech is a syllable emphasized utterance. The search processing unit 15 performs a search process using the selected statistical model.

  FIG. 3 is a diagram schematically showing the configuration of the syllable-emphasized utterance determination unit 13. The input speech power (E) extracted by the acoustic analysis unit 12 and the n-dimensional MFCC (MFCC zero-order term C0) A speech section detection unit 131 that detects a speech section of the input speech based on an acoustic feature amount, and a periodicity determination unit 132 that determines the appearance periodicity of the detected speech section. The determination result as to whether or not the input speech is syllable emphasis is output to the input control unit 11 in the first embodiment of FIG. 1 and to the statistical model selection unit 20 in the second embodiment of FIG. .

  Next, the operation of the syllable emphasized utterance determination unit 13 will be described in detail. One of the acoustic feature quantities often used in speech recognition is a cepstrum domain feature quantity (MFCC: Mel Frequency Cepstrum Coefficient) and power. MFCC is extracted by applying a Mel-scale filter bank to the power spectrum obtained by FFT analysis of audio data for each frame, and performing discrete cosine transform (DCT) on the frequency spectrum converted power spectrum This is a parameter that represents the spectral envelope, and the details are “speech recognition system” (edited by Kiyoshi Hiroshi, Katsunobu Ito, Tatsuya Kawahara, Kazuya Takeda, edited by Mikio Yamamoto, Ohm Publishing Co., Ltd .; Explained.

  In speech recognition, 12-dimensional MFCC (MFCC1, MFCC2) obtained by performing discrete cosine transform on spectral features of input speech and executing three processes in the cepstrum domain (DC component removal, liftering processing, and cepstrum average removal) ,... MFCC12) and its first time derivative (ΔMFCC1, ΔMFCC2,..., ΔMFCC12), and a first time derivative (ΔE) of power E, and a 25-dimensional acoustic feature, and further, a second time derivative ( A 38-dimensional MFCC including ΔΔMFCC1, ΔΔMFCC2,... ΔΔMFCC12, ΔΔE) is often used as an acoustic feature.

  ΔE represents a temporal change in power, and ΔE has a large positive peak because the power greatly increases at the boundary where the silent section switches to the utterance section. On the other hand, ΔE has a negative peak with a large absolute value at the boundary from the utterance section to the silent section. Therefore, it is possible to distinguish the utterance section and the silent section from the positive and negative peaks (or the maximum amplitude) of ΔE.

  In addition, by multiplying ΔE by the product (ΔMFCC_1) of absolute values of ΔMFCC (| ΔMFCC1 |, | ΔMFCC2 |, ... | On the other hand, the appearance of peaks due to power changes at other locations can be suppressed.

  FIGS. 4 and 5 show changes in ΔE · | ΔMFCC_1 | of the normal utterance and syllable emphasized utterance of the same utterance content “Kanagawa” by the same utterer whose waveforms are shown in FIGS. It can be seen that for both utterances, a positive peak appears at the start of the utterance interval and a negative peak appears at the end.

  Comparing FIGS. 4 and 5, in the normal utterance in FIG. 4, the peaks corresponding to the syllables are continuous, whereas in the syllable emphasized utterance in FIG. 5, peaks appear at almost constant intervals. . When the peak appearance is completely periodic, that is, when the peak interval is completely constant, the autocorrelation of ΔE · | ΔMFCC_1 | A peak appears when it matches.

  However, in practice, there is a fluctuation in the peak interval, and the peak of ΔE · | ΔMFCC_1 | is very sharp, so there is a high possibility that the peak does not appear clearly in the autocorrelation. In such a case, as shown in FIG. 6, for every certain frame, ΔE · | ΔMFCC_1 | is peak picked (rectangular wave) by representing the section with the maximum amplitude within the certain section. A peak can be made to appear in the autocorrelation result by using a technique for absorbing fluctuations in the equal peak interval. As in the frame processing for speech recognition described with reference to FIG. 10, a part of the certain section may be overlapped or may be continued without overlapping.

  FIG. 7 is a diagram showing the autocorrelation result of ΔE · | ΔMFCC — 1 | when the fluctuation is absorbed, and it can be seen that the delay widths τp1, τp2, and τp3 in which large peaks appear are substantially constant intervals. The closer the peak interval of ΔE · | ΔMFCC_1 | is to a certain value, the greater the value of each peak in the autocorrelation. Therefore, for example, it is possible to provide a criterion for determining that the input speech is syllable emphasized utterance when the first peak of the delay width τp1, that is, the primary autocorrelation exceeds a predetermined threshold.

  Therefore, in the present embodiment, the utterance interval detection unit 131 detects an utterance interval based on the time series information described with reference to FIG. 6, and the periodicity determination unit 132 determines the primary autocorrelation of the time series information. When a predetermined threshold value is exceeded, it is determined that the input voice is a syllable emphasized utterance.

  Further, FIG. 8 shows the case where autocorrelation of ΔC0 · | ΔMFCC_1 | of the same utterance is obtained using MFCC of zero-order term instead of the power E, that is, C0 corresponding to the DC component of the spectrum in each frame. Thus, although the scale is different, the same change as the autocorrelation of ΔE · | ΔMFCC_1 | of FIG. 7 is shown. Therefore, ΔC0 may be used instead of ΔE, and it may be similarly determined whether or not the input is a syllable-emphasized utterance.

  Furthermore, in the above-described embodiment, the time change rate (ΔE) of the power E of the input sound or the time change rate (ΔC0) of the 0th-order term (C0) of the MFCC of the input sound is equal to each n dimension of the MFCC of the input sound. The product of the absolute values of the time change rate (ΔMFCC_1) is multiplied, and this is explained as detecting the utterance interval based on time series information obtained by representing the maximum amplitude of the interval for each predetermined interval. The present invention is not limited to this, and the following modifications are possible.

  As a first modification, the utterance interval may be detected based only on the time change rate (ΔE) of the power (E) of the input speech.

  As a second modification, the utterance interval may be detected based only on the time change rate (ΔC0) of the 0th-order term (C0) of the MFCC of the input speech.

  As a third modification, the utterance period is set to the time change rate (ΔE) of the power E of the input speech or the time change rate (ΔC0) of the 0th-order term (C0) of the MFCC of the input speech. Alternatively, detection may be performed based on time-series information obtained by multiplying the products (ΔMFCC_1) of absolute values of the time change rates.

  As a fourth modification, the utterance period is set to the time change rate (ΔE) of the power E of the input speech or the time change rate (ΔC0) of the 0th-order term (C0) of the MFCC of the input speech to the n-dimensional component of the MFCC of the input speech. The product may be detected based on time-series information obtained by multiplying the product (ΔMFCC_1) of absolute values of the time change rate and smoothing the product.

1 is a block diagram of a first embodiment of a speech recognition apparatus according to the present invention. It is a block diagram of 2nd Embodiment of the speech recognition apparatus which concerns on this invention. It is the figure which expressed typically the structure of the syllable emphasis utterance determination part. FIG. 6 is a waveform diagram showing a change in ΔE · | ΔMFCC_1 | of a normal utterance “Kanagawa”. FIG. 6 is a waveform diagram showing changes in ΔE · | ΔMFCC_1 | of a syllable-emphasized utterance “Kananagawa”. FIG. 6 is a waveform diagram obtained by applying a technique for absorbing fluctuations to the waveform of FIG. 5. FIG. 6 is a diagram showing an autocorrelation result of ΔE · | ΔMFCC_1 | in which fluctuation is absorbed. FIG. 6 is a diagram illustrating an autocorrelation result of ΔC0 · | ΔMFCC_1 | in which fluctuation is absorbed. It is the figure which showed the structure of the principal part of the conventional speech recognition apparatus. It is a figure for demonstrating the extraction method of the acoustic feature-value in an acoustic analysis part. It is the figure which expressed typically the structure of the fixed grammar model. It is the figure which showed the HMM state series of the word "Imai". It is a waveform diagram of normal utterance "Kanagawa". It is a wave form diagram of syllable emphasis utterance "Kana naga wawa". It is the figure which showed the HMM state series which added the transition to "sil".

Explanation of symbols

DESCRIPTION OF SYMBOLS 11 ... Input control part, 12 ... Acoustic analysis part, 13 ... Syllable emphasis utterance determination part, 14 ... Regeneration request part, 15 ... Search processing part, 16 ... Acoustic model, 17 ... Language model, 18 ... 2nd acoustic model, 19 ... Second language model, 20 ... Statistical model selection unit

Claims (22)

Acoustic analysis means for extracting acoustic features of the input speech;
A statistical model for performing speech recognition based on the extracted acoustic features;
Syllable emphasized utterance determination means for determining whether or not the input speech is syllable emphasized utterance based on the periodicity of the extracted acoustic feature amount;
A search processing unit that executes a search process by applying the statistical model to the acoustic feature amount,
Depending on the determination result of whether or not the input speech is syllable emphasis utterance, perform a unique speech recognition operation,
The syllable-emphasized utterance determination means includes
Utterance section detection means for detecting the utterance section of each syllable based on a boundary that switches from the silent section to the utterance section and a boundary that switches from the utterance section to the silent section;
Periodicity determining means for determining the periodicity in which the utterance section of each syllable appears,
A speech recognition apparatus, characterized in that the input speech having periodicity is determined as syllable emphasized utterance . The syllable-emphasized utterance determination means is
The speech recognition apparatus according to claim 1, wherein when the appearance periodicity is higher than a predetermined reference value, the input speech is determined as a syllable emphasized utterance. The speech recognition apparatus according to claim 1, wherein the periodicity determination unit determines appearance periodicity based on an autocorrelation of an appearance period of the utterance section of each syllable . The speech recognition apparatus according to claim 1, wherein the periodicity determination unit determines appearance periodicity based on a first-order autocorrelation of an appearance period of the utterance section of each syllable .   5. The speech recognition apparatus according to claim 1, further comprising means for requesting the utterer to utter again when the input speech is determined to be syllable emphasized utterance. A first statistical model corresponding to acoustic features specific to normal speech;
A second statistical model corresponding to acoustic features specific to syllable-emphasized utterances,
5. The speech recognition according to claim 1, wherein the search processing unit performs a search process using the second statistical model when it is determined that the input speech is a syllable emphasized utterance. 6. apparatus. The acoustic feature amount includes power (E) of input voice,
7. The speech recognition apparatus according to claim 2, wherein the utterance section detecting unit detects a utterance section of each syllable based on a time change rate (ΔE) of power of input speech. The acoustic feature amount includes MFCC of input speech,
The speech recognition according to any one of claims 2 to 6, wherein the speech section detecting means detects a speech section of each syllable based on a time change rate (ΔC0) of a zero-order term of MFCC of input speech. apparatus. The acoustic feature amount includes power (E) of input speech and n-dimensional MFCC,
The utterance section detecting means is configured to take either the time change rate (ΔE) of the power of the input voice or the time change rate (ΔC0) of the MFCC of the input voice for the nth dimension of the MFCC of the input voice. The speech recognition apparatus according to any one of claims 2 to 6, wherein an utterance section of each syllable is detected based on time-series information obtained by multiplying products of absolute values of ΔMFCC. The acoustic feature amount includes power (E) of input speech and n-dimensional MFCC,
The utterance section detecting means is configured to take either the time change rate (ΔE) of the power of the input voice or the time change rate (ΔC0) of the MFCC of the input voice for the nth dimension of the MFCC of the input voice. 7. The speech recognition apparatus according to claim 2, wherein a speech section of each syllable is detected based on time series information obtained by multiplying the products of absolute values of ΔMFCC and smoothing the product. . The acoustic feature amount includes power (E) of input speech and n-dimensional MFCC,
The utterance section detecting means is configured to take either the time change rate (ΔE) of the power of the input voice or the time change rate (ΔC0) of the MFCC of the input voice for the nth dimension of the MFCC of the input voice. A product of absolute values of ΔMFCC is multiplied, and the utterance section of each syllable is detected based on time-series information obtained by representing the maximum amplitude of the section for each predetermined constant section. The speech recognition apparatus according to any one of 2 to 6. In the utterance determination method for determining whether or not the input speech is syllable emphasis utterance,
A procedure for extracting acoustic features of the input speech;
On the basis of the extracted acoustic features with a vocal section of each syllable of the input speech, and procedures for detecting the boundary and vocal section switched to speech section from a silent section based on the boundaries switched to silent section,
A procedure for determining the appearance periodicity of the utterance section of each syllable ;
And a procedure for determining whether or not the input speech is a syllable emphasized utterance based on the appearance periodicity of the utterance section of each syllable . A procedure for determining whether or not the input speech is a syllable emphasis utterance,
Wherein when occurrence periodicity is higher than the predetermined reference value, the utterance determination method according to claim 12, wherein the input speech is a procedure for determining the syllable emphasis utterance. The utterance determination method according to claim 12 or 13, wherein the procedure for determining the appearance periodicity determines the appearance periodicity based on an autocorrelation of an appearance period of the utterance section of each syllable . The utterance determination method according to claim 12 or 13, wherein the step of determining the appearance periodicity determines the appearance periodicity based on a primary autocorrelation of the appearance period of the utterance section of each syllable. . The acoustic feature amount includes power (E) of input voice,
The utterance determination method according to any one of claims 12 to 15, wherein the procedure of detecting the utterance section of each syllable detects the utterance section based on a time change rate (ΔE) of power of the input speech. . The acoustic feature amount includes MFCC of input speech,
16. The procedure for detecting the utterance interval of each syllable detects the utterance interval based on a time change rate (ΔC0) of the 0th-order term of the MFCC of the input speech. Utterance determination method. The acoustic feature amount includes power (E) of input speech and n-dimensional MFCC,
The procedure for detecting the utterance interval of each syllable is determined by either the time change rate (ΔE) of the power of the input speech or the time change rate (ΔC0) of the 0th-order term of the MFCC of the input speech. The utterance determination method according to any one of claims 12 to 15, wherein the utterance section is detected based on time-series information obtained by multiplying products of absolute values of minute time change rates ΔMFCC. The acoustic feature amount includes power (E) of input speech and n-dimensional MFCC,
The procedure for detecting the utterance interval of each syllable is determined by either the time change rate (ΔE) of the power of the input speech or the time change rate (ΔC0) of the 0th-order term of the MFCC of the input speech. The utterance section according to claim 12, wherein the utterance section is detected based on time-series information obtained by multiplying products of absolute values of the time change rate ΔMFCC of minutes and smoothing the product. Judgment method. The acoustic feature amount includes power (E) of input speech and n-dimensional MFCC,
The procedure for detecting the utterance interval of each syllable is determined by either the time change rate (ΔE) of the power of the input speech or the time change rate (ΔC0) of the 0th-order term of the MFCC of the input speech. Multiplying the products of the absolute values of the time change rate ΔMFCC of minutes and detecting the utterance interval based on time series information obtained by representing the maximum amplitude of the interval for each predetermined interval The utterance determination method according to claim 12.   An utterance determination program for causing a computer to execute the utterance determination method according to any one of claims 12 to 20.   A storage medium storing the utterance determination program according to claim 21 so as to be readable by a computer.