JP3615088B2 - Speech recognition method and apparatus - Google Patents

Description

[0001]
BACKGROUND OF THE INVENTION
The present invention relates to a speech recognition method and apparatus suitable for accurately recognizing spoken speech.
[0002]
[Prior art]
In recent years, speech recognition technology has played an important role in realizing an excellent man-machine interface. Recently, research and development for natural utterance recognition, which does not require restrictions on the utterance method of the speaker, such as word spotting using HMM and continuous speech recognition, have been actively conducted. Conventionally, in these speech recognition methods, an utterance content is recognized by cutting out a section in which it is determined that a speaker is speaking from an input signal and matching that portion with a standard pattern.
[0003]
However, in actual natural utterances, silent sections with low signal power, such as urging sounds, friction sounds, and voiced voices, may occur even in portions that are determined to be speech sections. In the section where the signal power is low, the influence of the background noise becomes relatively large, so that the spectrum of the signal is not stable, and as a result, it is matched with an incorrect pattern, and erroneous recognition often occurs.
[0004]
Furthermore, since the silent section with low power generated in such natural speech is difficult to anticipate in advance, it cannot be registered as a standard pattern.
[0005]
[Problems to be solved by the invention]
As described above, conventionally, when a silent section with low power is present in the section detected as the utterance section, there is a problem that the background noise spectrum becomes dominant in the section and erroneous pattern matching occurs. Further, in the utterance section, it is difficult to predict a section where the power is low in advance, and therefore, there is a problem that those patterns cannot be registered as a standard pattern.
[0006]
The present invention has been made in consideration of the above circumstances, and even if there is a silent section with low power that occurs irregularly in the utterance section, it is possible to perform highly accurate recognition without being affected by it. An object of the present invention is to provide a speech recognition method and apparatus.
[0007]
[Means for Solving the Problems]
The present invention acoustically analyzes an input signal to detect a section where speech is uttered, extracts a feature vector series from the detected speech signal of the utterance section, and extracts the extracted feature vector series and a predetermined recognition candidate A collation score representing the similarity or distance between the two is calculated by collating with a standard pattern of an audio signal prepared in advance for each of the first patterns, and based on the collation score for each recognition candidate In the speech recognition method for determining the recognition result, a silent section of the speech signal is detected from the short-time power of the detected speech signal of the speech section, and the feature vector sequence of the silent section is excluded from pattern matching, The collation score is obtained by collating the feature vector sequence corresponding to the time changing from the section to the voiced section using the second collation method considering the influence of the silent section. Characterized in that the calculation be. Here, an HMM (Hidden Markov Model) matching method may be applied to the first matching method, and an HMM matching method allowing null transition may be applied to the second matching method.
[0008]
According to the present invention, even if there is a silent section of unexpected low power in the utterance section, the silent section is detected, and when the matching with the standard pattern is performed, the silent section is excluded and the matching is performed. As a result, erroneous pattern matching in a silent section can be avoided, and highly accurate recognition is possible. In addition, in the present invention, since the second collation method considering the influence of the silent section is applied to the feature vector sequence corresponding to the time changing from the silent section to the voiced section, for example, the HMM collation scheme that allows the null transition, There is no inconsistency in the state transition due to the fact that the silent section (feature vector) is not used for collation.
[0009]
Here, when the null transition is allowed at the time when the silent section is switched to the voiced section, the immediately preceding time (frame) is set as the HMM state that causes the null transition to the HMM state (first state i) at that time. By selecting a state (second state j) in which the matching score of the optimum route is the maximum among states before state i in FIG. 8, a null transition from state j to state i occurs, and the time immediately before state i It is preferable to replace the collation score in (1) with the collation score at the same time in state j. You may make it restrict | limit the state in which the null transition to this state i is possible by the duration of a silence area.
[0010]
In addition, the present invention independently detects a silence interval based on the short-time power of the speech signal in the utterance interval, using different threshold values, and for each feature vector sequence extracted from the utterance interval speech signal and a predetermined recognition candidate. By comparing the standard pattern of the audio signal prepared in advance with the hidden Markov model matching method based on the silent section information detected independently based on the different threshold values, a matching score is obtained for each threshold value. Hidden Markov Matching that excludes the feature vector sequence of silence intervals calculated based on the corresponding thresholds at that time, and that allows null transitions only at the time of transition from silence intervals to speech intervals It is also characterized in that a recognition result is determined on the basis of a matching score for each recognition candidate obtained by applying a method to detect each threshold.
[0011]
In this way, using the silent section information obtained for each threshold, for each threshold, the corresponding silent section is excluded from pattern matching, and a matching score is obtained for each recognition candidate, and the recognition result is obtained based on the matching score. By determining, it is possible to reduce the influence of erroneous matching in the silent section.
[0012]
Here, each time a collation score for each recognition candidate is calculated for one threshold value, the recognition candidates are narrowed down based on the collation score for each recognition candidate, and the operation is repeated while gradually switching the threshold value in a certain direction. It is good to do so. The silent section detection using different threshold values may be performed in parallel or sequentially while switching the threshold values. In the former case, it is necessary to store the detection result of the silent section. In the latter case, it is necessary to store at least the voice signal of the utterance section.
[0013]
In this way, it is possible to narrow down the recognition candidates step by step by changing the threshold for silent section detection (power threshold) stepwise in a certain direction and performing pattern matching while pruning the recognition candidates. , Improve the accuracy of recognition and reduce misrecognition.
[0014]
Here, if the threshold value is switched in the direction in which the threshold value becomes smaller, when selecting a recognition candidate, it is possible to apply a weight to a large portion of the power where the spectrum is stable, and the spectrum is unstable and the power is low. The influence of a section can be reduced.
[0015]
Further, the threshold value may be switched in the direction in which the threshold value increases. In this case, although erroneous matching in the silent section is allowed at first and a plurality of recognition candidates are selected, the correct answer candidates are correctly matched in other than the silent section, so the higher candidates are entered, and the threshold is gradually increased to perform matching. By doing so, it is possible to reduce the influence of erroneous matching in the silent section, and finally correct correct candidates can be detected.
[0016]
Also, instead of narrowing down recognition candidates while switching the threshold stepwise in a certain direction, a process of calculating a weighted sum of matching scores obtained for each threshold value for the same recognition candidate is executed for all recognition candidates, and all recognition It is also possible to determine the recognition result based on the weighted sum of the matching scores of each candidate.
In this case, the influence of the silent section can be arbitrarily reflected in the matching score, thereby reducing the influence of erroneous matching in the silent section.
[0017]
Note that the present invention relating to a method is also established as an invention relating to an apparatus.
The present invention also allows a computer to execute a procedure corresponding to the invention (or causes a computer to function as a means corresponding to the invention, or causes a computer to realize a function corresponding to the invention). It can also be realized as a computer-readable recording medium on which a program is recorded.
[0018]
DETAILED DESCRIPTION OF THE INVENTION
Embodiments of the present invention will be described below with reference to the drawings.
[0019]
[First Embodiment]
FIG. 1 schematically shows a speech recognition apparatus according to a first embodiment of the present invention.
The speech recognition apparatus shown in FIG. 1 analyzes an input signal to detect a speech segment, and performs acoustic analysis on the speech signal of the speech segment detected by the speech segment detection unit 101. A feature vector extracting unit 102 for extracting a feature vector, a silent segment detecting unit 106 for detecting a silent segment from the speech signal of the speech segment detected by the speech segment detecting unit 101 using the power of the speech signal, The feature vector extraction unit 102 uses the standard feature pattern storage unit 104 that stores the learned standard feature patterns of each recognized recognition candidate and the silent section information detected by the silent section detection unit 106. A pattern for collating a feature vector series with a standard feature pattern of each recognition candidate stored in the standard feature pattern storage unit 104 by a collation method using an HMM A verification unit 103, based on the comparison result for each recognition candidate obtained by the pattern matching unit 103 determines the recognized utterance contents, and a recognition result determining unit 105.
[0020]
In FIG. 1, a voice input unit including a microphone and an A / D (analog / digital) converter that inputs a voice uttered by a speaker and converts the voice into a digital electric signal (digital voice signal) is omitted. .
[0021]
Next, the processing concept of the speech recognition apparatus configured as shown in FIG. 1 will be described.
The speech signal in the utterance section detected by the utterance section detection unit 101 is frequency-analyzed for each of a plurality of predetermined frequency bands by the feature vector extraction unit 102, and is converted into a feature vector series (feature vector time series) {xt}. Converted. A feature vector (feature parameter) is obtained in units of a fixed time length called a frame. As typical feature vectors used for speech recognition, a power spectrum that can be obtained by a bandpass filter or Fourier transform, a cepstrum coefficient that is obtained by LPC (linear prediction) analysis, and the like are well known. However, in this embodiment, the type of feature vector to be used is not limited.
The feature vector time series extracted by the feature vector extraction unit 102 is sent to the pattern matching unit 103.
[0022]
On the other hand, the speech signal in the utterance interval is also sent to the silence interval detection unit 106, and the silence interval is detected in synchronization with the frame of the feature vector series from the short-time power of the audio signal. FIG. 2 conceptually shows the state of a signal in which a silent section is detected by this part of processing. In FIG. 2, the horizontal axis represents time, the vertical axis represents the short-time power of the signal, and TH represents a preset power threshold.
[0023]
The silent section detecting unit 106 compares the short-time power value Pt at each time t with the power threshold TH every time, and determines a section where Pt <TH is a silent section. Information (silent section information) indicating the silent section obtained in this way is sent to the pattern matching unit 103. Here, the time t indicates the t-th frame in the utterance section.
[0024]
The pattern matching unit 103 performs pattern matching using the input feature vector series, silent section information, and a standard feature pattern (standard pattern) learned in advance. The standard feature pattern is stored in advance in the standard feature pattern storage unit 104 as an HMM for each predetermined recognition candidate (recognition unit). At the time of recognition, this HMM is used as it is or in combination.
[0025]
FIG. 3 shows the structure of the HMM used for collation. Here, of the state transitions, the transition with the symbol c is a null transition, and the transitions with the symbols a and b are a normal state transition and a self-loop, respectively. In the HMM of FIG. 3, null transitions are assumed between all states, but it is also possible to limit the states in which null transitions occur by providing constraints here.
[0026]
Next, a pattern matching method using the HMM having the structure of FIG. 3 applied in the pattern matching unit 103 will be described with reference to the flowchart of FIG.
In step S <b> 101, whether or not the input signal at time t, i.e., the signal of the t-th frame is an utterance interval, is determined based on the detection result of the utterance interval detection unit 101. If the input signal at time t is a signal in the utterance section, the process proceeds to step S102, and if not, the process proceeds to step S106.
[0027]
In step S <b> 102, it is determined whether the input signal at time t is a silent section signal based on the detection result of the silent section detection unit 106. If it is determined that the signal is a silent section, the process proceeds to step S107. If it is determined that the signal is a sound section, the process proceeds to step S103.
[0028]
In step S103, the value of the flag (FLAG) is evaluated. The flag takes a value of 0 or 1, and whether the signal at time t-1 (that is, the signal one frame before) belongs to a silent section (when FLAG = 0) or belongs to a voiced section (FLAG = 1). In the case of When the value of the flag is 0, it is determined that the time t is the time when the silent period (until time t-1) is switched (changed) from the silent period to the last step S108. It is determined that the silent section continues, and the process proceeds to step S104.
[0029]
In step S104, in the HMM shown in FIG. 3, all state transition probabilities except for the null transition and output probabilities of all distributions for the signal at time t are calculated, and the optimum transition is determined. After the determination, the process proceeds to step S105.
In step S105, time t is set to the next time t + 1, and the process returns to step S101.
[0030]
In step S106, for each recognition candidate, in the HMM shown in FIG. 3, a state in which the matching score is maximized at the utterance section end time t is selected, and the matching score for each recognition candidate is input to the recognition result determination unit 105. Sent to finish the process. Here, as is well known, the matching score is an evaluation value representing the similarity or distance between the feature vector series of the input speech signal and the standard feature pattern.
In step S107, when it is determined in step S102 that the signal at time t is a silent section signal, the flag value described above is set to 0, and the process proceeds to step S105.Here, pattern matching is not performed, and therefore the feature vector at time t is not subject to pattern matching, and the matching score of each state is not updated.
[0031]
In step S108, when it is determined in step S103 that the signal at time t is the time when the silent section is switched to the voiced section, a null transition is first performed in the HMM shown in FIG. The collation score at time t-1 is updated. After the matching score is updated, all state transition probabilities except for the null transition and output probabilities of all distributions are calculated, and the optimum transition is determined. After the determination, the process proceeds to step S109. Details of the processing of this part will be described later.
[0032]
In step S109, when it is determined in step S102 that the signal at time t is a signal in a sound section, the value of the flag described above is set to 1, and the process proceeds to step S105.
[0033]
The above is the outline and flow of the pattern matching method directly related to the present invention.
By the processing in the pattern matching unit 103 to which the pattern matching method is applied, the matching scores of all the recognition candidates are calculated, and the recognition candidate having the maximum score is selected as the recognition result in the recognition result determination unit 105.
[0034]
Here, the details of the processing in step S108 at time t when the silent section is switched to the voice section will be described with reference to the flowchart of FIG.
At time t, first, in step S401, the state number i is set to the final state.
[0035]
In step S402, for state i, state j is selected from state 0 to state i that maximizes the matching score of the optimum route at time t-1 (one frame before).
[0036]
In step S403, a null transition from state j to state i occurs, and the matching score at time t-1 (one frame before) of state i is replaced with the matching score at time t-1 of state j.
[0037]
In step S404, it is determined whether or not the state i is the leading state 0. If the state is 0, the process proceeds to the final step S406, and if not, the process proceeds to step S405.
[0038]
In step S405, i is counted down by 1, and the process returns to step S402.
In step S406, the optimum route excluding the null transition and its matching score at time t are obtained for all states.
[0039]
Thus, by considering the null transition at the time when the silent section is switched to the voiced section, it is possible to remove the influence that the feature vector of the silent section is not used for matching. Here, it is assumed that the null transition to state i can occur from all states from state 0 to state i. However, there is a restriction here, for example, the null to state i depends on the duration of a silent interval, etc. It is also possible to limit the states in which transition is possible (reducing the number of states as the duration is shorter). Further, when the duration of the silent section is equal to or less than a predetermined threshold, it is possible to prevent a null transition from occurring. Further, here, a state in which a null transition is possible is searched for all the states from the final state, but this need not necessarily be performed for all states, and a silent section is likely to occur based on prior information in advance. It is also possible to make a null transition only for the state.
[0040]
Next, the effect of this embodiment will be described with reference to FIGS.
FIG. 6 is an image diagram of signal power when “TOSAKA” is spoken. Here, times T0 and T7 indicate the start time and end time of the utterance section, respectively. In addition, each section of time T0-T1, T2-T3, T4-T5, T6-T7 is a section determined as a silent section by the power threshold TH.
[0041]
In general, the silent section in the utterance section is generated by devoicing of the urging sound, the friction sound, and the voiced sound. In this section, the influence of the background noise becomes relatively large, so that matching with an erroneous pattern is likely to occur. As a result, erroneous recognition may occur. According to FIG. 6, there is a possibility that erroneous pattern matching may occur in the sections from T0 to T1, T2 to T3, T4 to T5, and T6 to T7.
[0042]
FIG. 7 shows in more detail the state of the short-time power of the audio signal and the generated content (expressed here as a phoneme string) in the section from T2 to T3. In this example, the section corresponding to the frictional sound / S / is completely below the threshold value TH. As described above, in this case, erroneous matching is likely to occur in the section from T2 to T3 that is equal to or less than the power threshold TH.
[0043]
FIG. 8 is an HMM representing “TOSAKA” in which one phoneme is represented in one state for the sake of simplicity. Here, for the sake of simplicity, the null transition from the state / O // S // A // K // A / is omitted.
[0044]
When the pattern matching method described above is applied to the HMM as shown in FIG. 8, control is performed so that the feature vector series is not used for matching in the section from T2 to T3 (silent section of the audio signal). For this reason, unlike the prior art in which the feature vector series is used for matching regardless of the voiced and silent sections of the speech signal, there is no false matching in the T2 to T3 section (silent section). Does not adversely affect the matching score. Moreover, in the pattern matching method applied in the present embodiment, null transition is allowed at the time when the silent section changes to the voiced section, so that there is no state transition inconsistency caused by not using the silent section for matching. .
[0045]
As a result, in the present embodiment, the transition as shown in FIG. 9 is possible without adversely affecting the matching score. In this example, since the power of the feature vector corresponding to the phoneme / S / is equal to or less than the power threshold TH, the feature vector of this portion is not used for matching, and the phoneme / O / Is allowed to pass from the phoneme / S / state to the phoneme / A /.
This can be considered in the same manner for silent sections other than T2-T3 (T0-T1, T4-T5, T6-T7).
[0046]
When the utterance period ends, the optimal state transition path at time T7 and the matching score at that time are obtained for all states, and the maximum score may be used for determination of the recognition result.
[0047]
If this method is used, it is possible to avoid an increase in the verification score of an incorrect recognition candidate due to an incorrect matching of a silent section in the verification of a recognition candidate for utterance. As a result, the accuracy of the matching score is improved, leading to an improvement in recognition rate. The above is the detailed description of the configuration, operation, and effect of the speech recognition apparatus according to the first embodiment of the present invention.
[0048]
[Second Embodiment]
FIG. 10 schematically shows a speech recognition apparatus according to the second embodiment of the present invention.
[0049]
The speech recognition apparatus shown in FIG. 10 includes an utterance section detection unit 201, a feature vector extraction unit 202, a pattern matching unit 203, a standard feature pattern storage unit 204, a recognition result determination unit 205, and N silent section detection units (# 1). ) 206-1 to (#N) 206-N.
[0050]
The characteristics of the configuration of FIG. 10 are prepared in advance by N silent section detectors # 1 (206-1) to #N (206-N) (corresponding to the silent section detector 106 in FIG. 1). The silent section of the voice signal (of the utterance section) is detected based on the threshold values TH1 to THN of the powers of the different signals. For this reason, the functions of the pattern matching unit 203 (corresponding to the pattern matching unit 103 and the recognition result determination unit 105 in FIG. 1) and the recognition result determination unit 205 are also partially different as will be described later. The other components, that is, the utterance section detection unit 201, the feature vector extraction unit 202, and the standard feature pattern storage unit 204 are the utterance section detection unit 101, the feature vector extraction unit 102, and the standard feature pattern storage unit in FIG. Similar to 104.
[0051]
Therefore, the operation of the speech recognition apparatus having the configuration shown in FIG. 2 will be described focusing on the differences from the speech recognition apparatus shown in FIG.
Audio signals detected by the utterance section detection unit 201 are input in parallel to the silent section detection units # 1 (206-1) to #N (206-N). Each silent section detector #i (i = 1 to N) has thresholds THi of different powers, and silent sections of the audio signal are detected independently using these threshold values.
[0052]
FIG. 11 shows a silent threshold detection unit # 1 (206-1), a silent segment detection unit # 2 (206-2),... A silent segment detection unit #N (206-N), and a preset power threshold TH1. , TH2,... THN represents a state in which a silent section of the voice signal in the utterance section is detected. Here, it is assumed that THi> THi + 1 is set.
[0053]
Information indicating silence sections (silence section information) independently detected by the silence section detection units # 1 (206-1) to #N (206-N) is sent to the pattern matching unit 203. A time series (feature vector series) of feature vectors extracted by the feature vector extraction unit 202 is also sent to the pattern matching unit 203. The pattern matching unit 203 uses the feature vector series input from the feature vector extraction unit 202 and the silent section information input from the silent section detection units # 1 (206-1) to #N (206-N). A collation score for each recognition candidate is calculated.
[0054]
Here, processing in the pattern matching unit 203 and the recognition result determination unit 205 will be described with reference to the flowchart of FIG.
In step S201, an initial setting process is performed, and i = 1 is set as a parameter (silent section detector number) indicating the silent section detector #i.
[0055]
In step S202, for all recognition candidates, the matching score is calculated by the pattern matching unit 203 using the silent interval information from the silent interval detecting unit #i. For the matching score calculation in the pattern matching unit 203, the matching method (in the pattern matching unit 103) described in the first embodiment is used.
[0056]
In step S203, recognition candidates up to the top Mi are selected from the collation score for each recognition candidate calculated in step S202 according to the number of recognition candidates Mi for pruning prepared in advance, and recognition in the next step is performed. Left as a candidate. Here, it is assumed that Mi> Mi + 1 is set.
[0057]
In step S204, it is determined whether i has reached a parameter value (silent section detector number) N representing the silent section detector #N. If i = N, the process proceeds to the final step S206; otherwise, the process proceeds to step S205.
In step S205, i is incremented by 1, and the process returns to step S202.
[0058]
In step S206, the recognition result determination unit 205 selects the one having the largest matching score from the recognition candidates remaining up to that point (up to the higher MN rank), and outputs it as a recognition result.
The processing in the pattern matching unit 203 and the recognition result determination unit 205 in the second embodiment has been described above.
[0059]
When the above method is used, when a recognition candidate is selected, it is possible to apply a weight to a large power portion where the spectrum is stable, and it is possible to reduce the influence of a low power section where the spectrum is unstable. In addition, by changing the power threshold in stages and performing pattern matching while pruning recognition candidates, the recognition candidates can be narrowed down in stages, improving recognition accuracy and reducing misrecognition. Can do.
The above is the detailed description of the configuration, operation, and effect of the speech recognition apparatus according to the second embodiment of the present invention.
[0060]
(First Modification of Second Embodiment)
In the second embodiment described above, the pruning of recognition candidates in the pattern matching unit 203 has been described as being performed using the power threshold in descending order, but conversely, the power threshold is decreasing in order. It is also possible to do this.
[0061]
Therefore, a first modification of the second embodiment to which the method of pruning recognition candidates in order from the power threshold in the configuration of FIG. 10 is applied will be described with reference to the flowchart of FIG.
[0062]
In step S301, i = N is initialized.
In step S302, for all recognition candidates, the matching score is calculated by the pattern matching unit 203 using the silent section information from the silent section detecting unit #i. For the matching score calculation in the pattern matching unit 203, the matching method (in the pattern matching unit 103) described in the first embodiment is used.
[0063]
In step S303, recognition candidates up to the upper Mi position are selected from the matching score calculated in step S302 according to the number of recognition candidates Mi for pruning prepared in advance, and are left as recognition candidates in the next step. Here, it is assumed that Mi <Mi + 1 is set, unlike the example of the previous collation method.
[0064]
In step S304, it is determined whether i has reached a parameter value (silent section detecting unit number) 1 representing the silent section detecting unit # 1. If i = 1, the process proceeds to the final step S306; otherwise, the process proceeds to step S305.
In step S305, i is counted down by 1, and the process returns to step S302.
[0065]
In step S306, the recognition result determination unit 205 selects the one having the maximum collation score from the recognition candidates remaining up to that point (up to the top M1) and outputs the recognition result.
The processing in the pattern matching unit 203 and the recognition result determination unit 205 according to the first modification of the second embodiment has been described above.
[0066]
In the above method, first, a plurality of recognition candidates are selected with a small threshold of power, allowing erroneous matching in a silent section. Since the correct answer matches correctly except in the silent section, it is included in the top candidate. Then, by gradually increasing the power threshold and performing matching, it is possible to reduce the influence of erroneous matching in the silent section, and finally detect a correct candidate.
[0067]
If such a method is used, when selecting a recognition candidate, first, among the recognition candidates including erroneous matching in the silent section, it is possible to reduce the influence of erroneous matching in the silent section step by step. Recognition accuracy can be improved and misrecognition can be reduced.
The above is the detailed description of the operation and effect of the speech recognition apparatus in the first modification of the second embodiment of the present invention.
[0068]
(Second modification of the second embodiment)
In the second embodiment described above and the first modification of the embodiment, the pattern matching unit 203 obtains the silent section information detected for each threshold THi using different power thresholds THi. Although it has been described that the recognition candidates are pruned using a plurality of matching scores in order and the recognition result is obtained, the present invention is not limited to this. For example, it is possible to determine the recognition result by taking a weighted sum of the matching scores obtained for each threshold THi.
[0069]
Therefore, a second modification of the second embodiment using this method will be described with reference to the flowchart of FIG.
In step S501, the recognition candidate number i is initially set to 1.
[0070]
In step S502, a parameter (silent section detector number) j indicating the silent section detector #j is initially set to 1.
In step S503, the pattern matching unit 203 calculates the matching score sij of the recognition candidate i (recognition candidate whose recognition candidate number is i) using the silent section information from the silent section detection unit #j.
[0071]
In step S504, it is determined whether j has reached a parameter value (silent section detector number) N representing the silent section detector #N. If j = N, the process proceeds to step S506; otherwise, the process proceeds to step S505.
In step S505, j is incremented by 1, and the process returns to step S503.
[0072]
In step S506, each silent section detector #j (j = 1 to N), that is, the weighted sum of the matching scores sij calculated using the silent section information from the silent section detectors # 1 to #N by the weight wj, In other words, the weighted sum of the si1 to siN weights w1 to wN is calculated, and the verification score Si of the recognition candidate i used for recognition result determination is calculated. Here, wj is a predetermined weight (0 ≦ wj ≦ 1), and is a weight for the collation score sij calculated using the silent section information from the silent section detecting unit #j.
[0073]
In step S507, whether or not the matching score Si has been calculated for all recognition candidates is determined based on the value of i. If i has reached the number of recognition candidates, the process proceeds to the final step S509, and if not, the process proceeds to step S508.
In step S508, the recognition candidate number i is incremented by 1, and the process returns to step S502.
[0074]
In step S509, the recognition result determination unit 205 compares the collation scores Si of all recognition candidates, and the recognition candidate having the maximum Si is determined and output as a recognition result.
The processing in the pattern matching unit 203 and the recognition result determination unit 205 according to the second modification example of the second embodiment has been described above.
[0075]
In the above method, appropriate weights w1 to w1 for the collation scores si1 to siN for the same recognition candidate i calculated using the silent section information for each threshold obtained based on the thresholds TH1 to THN of different powers. By taking the sum by multiplying wN, the influence of the silent section can be arbitrarily reflected in the collation score. For this reason, the influence of the incorrect matching in a silence area can be reduced.
The above is the detailed description of the operation and effect of the speech recognition apparatus in the second modification of the second embodiment of the present invention.
[0076]
In the second embodiment, the silent section detectors # 1 (206-1) to #N (206-N) have been described as operating in parallel, but are detected by the utterance section detector 201. The voice signal of the utterance section is stored in the storage means such as a memory, and in this state, the silent section detection units # 1 (206-1) to #N (206-N) are sequentially activated to store the voice signal in the storage means. It is also possible to cause the silent section detection unit to detect a silent section with a specific threshold value and send the detected silent section information to the pattern matching unit 203 each time.
[0077]
In addition, the utterance section detection unit 101 (201), the feature vector extraction unit 102 (202), the pattern matching unit 103 (203), the recognition result determination unit 105 (205), and the silent section detection unit 106 (206−) in the above embodiment. Each function of 1 to 206-N) can also be realized as software.
[0078]
Further, the present embodiment is for causing a computer to execute a predetermined procedure including the pattern matching method applied in the speech recognition apparatus according to the above embodiment (or for causing the computer to function as a predetermined means possessed by the speech recognition apparatus. Or a computer-readable recording medium such as a CD-ROM on which a program (for realizing a predetermined function of the voice recognition apparatus) is recorded. The program may be downloaded via a communication medium.
[0079]
In addition, the embodiment of the present invention can be variously modified with respect to the above-described example, and these are also within the scope of the embodiment of the present invention unless they are contrary to the spirit.
[0080]
【The invention's effect】
As described above, according to the present invention, even if there is a silent section with low unexpected power in the utterance section, by detecting the silent section and using it in matching with the standard feature pattern, An erroneous pattern matching in a silent section can be avoided, and a great practical effect such as enabling highly accurate recognition can be achieved.
[Brief description of the drawings]
FIG. 1 is a block diagram showing a basic configuration of a speech recognition apparatus according to a first embodiment of the present invention.
FIG. 2 is a conceptual diagram showing a silent section in an input voice signal.
FIG. 3 is a diagram showing a configuration of an HMM including a null transition.
FIG. 4 is a diagram showing a flow of a pattern matching method.
FIG. 5 is a diagram showing a flow of processing at a time when a silent section is switched to a voice section.
FIG. 6 is a diagram showing a state of power in an input audio signal.
FIG. 7 is a diagram showing details of the state of power in an input audio signal.
FIG. 8 is a diagram showing a specific example of the configuration of an HMM.
FIG. 9 is a conceptual diagram of an optimum route after pattern matching processing.
FIG. 10 is a block diagram showing a basic configuration of a speech recognition apparatus according to a second embodiment of the present invention.
FIG. 11 is a conceptual diagram showing a silent section of an input signal with a plurality of threshold values.
FIG. 12 is a diagram showing a flow of a pattern matching method using a plurality of threshold values.
FIG. 13 is a diagram showing a first modification of the flow of the pattern matching method using a plurality of threshold values.
FIG. 14 is a diagram showing a second modification of the flow of the pattern matching method using a plurality of threshold values.
[Explanation of symbols]
101, 201 ... utterance section detection unit
102, 202 ... feature vector extraction unit
103, 203 ... pattern matching unit
104, 204 ... Standard feature pattern storage unit
105, 205 ... Recognition result determination unit
106, 206-1 to 206-N: Silent section detector

Claims (6)

An input signal is acoustically analyzed to detect a section in which speech is uttered, a feature vector series is extracted from the detected speech signal in the utterance section, and prepared in advance for each of the extracted feature vector series and predetermined recognition candidates By comparing the standard pattern of the voice signal with a hidden Markov model matching method, a matching score representing the similarity or distance between the two is calculated, and the recognition result is judged based on the matching score for each recognition candidate In the speech recognition method to
Detecting a silent interval of the audio signal from the short-time power of the audio signal of the detected utterance interval,
The feature vector sequence of the silent section is excluded from pattern matching, and a null transition is performed for the feature vector sequence corresponding to the time when the silent section changes to the voiced section, and the matching score one frame before each state is calculated. After updating and updating the matching score, all state transition probabilities except for the null transition and output probability of all distributions are calculated, and the optimal transition is determined, and matching is performed using a hidden Markov model matching method that allows the null transition. A speech recognition method characterized in that a collation score is calculated. While the input signal is acoustically analyzed to detect the section where the voice is uttered, and the feature vector series is extracted from the voice signal of the detected utterance section,
From the short-time power of the audio signal of the detected utterance interval, independently detect the silent interval of the audio signal based on different threshold values,
Hidden Markov model matching method based on information on silent sections that are independently detected based on the different threshold values, the extracted feature vector series and a standard pattern of a speech signal prepared in advance for each predetermined recognition candidate By collating with each other, the collation score representing the similarity or distance between the two is calculated for each threshold, and the feature vector series of the silent section detected based on the corresponding threshold at that time is excluded from the pattern matching target In addition, the null transition is performed only at the time when the silent section changes to the voiced section, the collation score is updated one frame before each state, and after the collation score is updated, all state transition probabilities except for the null transition, and Apply a hidden Markov matching scheme that allows null transitions, where the output probabilities of all distributions are calculated and the optimal transition is determined,
A speech recognition method, wherein a recognition result is determined based on a collation score for each recognition candidate obtained for each threshold. Each time the calculation of the matching score for each recognition candidate is performed for one threshold, the recognition candidates are narrowed down based on the matching score for each recognition candidate, and the operation is repeated while gradually switching the threshold in a certain direction. The speech recognition method according to claim 2. A process of calculating a weighted sum of matching scores obtained for each threshold for the same recognition candidate is performed for all recognition candidates, and a recognition result is determined based on the weighted sum of the matching scores of all the recognition candidates. The speech recognition method according to claim 2. An utterance section detecting means for acoustically analyzing an input signal and detecting a section in which the voice is uttered;
Feature vector extraction means for extracting a feature vector series from the speech signal of the utterance section detected by the utterance section detection means;
A silent section detecting means for detecting a silent section of a voice signal from a short time power of the voice signal of the voice section detected by the voice section detecting means;
A standard pattern storage means for storing a standard pattern of the audio signal of each predetermined recognition candidate;
By comparing the feature vector series extracted by the feature vector extraction unit with the standard pattern of each recognition candidate stored in the standard pattern storage unit using a hidden Markov matching method, the similarity or distance between them is expressed. A pattern matching means for calculating a matching score, wherein the feature vector series of the silent section detected by the silent section detecting means is excluded from pattern matching, and only at the time when the silent section changes to the voiced section. , Perform a null transition, update the matching score one frame before each state, and after updating the matching score, all state transition probabilities except for the null transition and output probabilities of all distributions are calculated to determine the optimal transition Pattern matching means for matching by a hidden Markov matching method that allows null transition,
A speech recognition apparatus comprising: a recognition result determining unit that determines a recognition result based on a matching score for each recognition candidate obtained by the pattern matching unit. An utterance section detecting means for acoustically analyzing an input signal and detecting a section in which the voice is uttered;
Feature vector extraction means for extracting a feature vector series from the speech signal of the utterance section detected by the utterance section detection means;
A plurality of silent section detecting means for detecting a silent section of the voice signal based on different threshold values from the short-time power of the voice signal of the voice section detected by the voice section detecting means;
A standard pattern storage means for storing a standard pattern of the audio signal of each predetermined recognition candidate;
The feature vector series extracted by the feature vector extraction means is stored in the standard pattern storage means for each of the different threshold values, except for the feature vector series of the silence periods detected by the silence period detection means. A pattern matching means for calculating a matching score representing the similarity or distance between the standard patterns of each recognition candidate by the hidden Markov matching method, at a time when the silent section changes to the voiced section. Only the null transition is performed, the collation score is updated one frame before each state, and after the collation score is updated, all the state transition probabilities except for the null transition and the output probabilities of all distributions are calculated. A pattern matching means for matching by a determined Markov matching method that allows a null transition;
A speech recognition apparatus comprising: a recognition result determining unit that determines a recognition result based on a matching score for each recognition candidate obtained by the pattern matching unit for each threshold value.

Images