JPH09222899A - Word voice recognizing method and device for implementing this method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To reduce the erroneous recognition caused by the addition of noises detected in the voice zone. SOLUTION: A partial standard pattern expressing the voice feature for each standard pattern is extracted in advance, the partial input pattern having the same timewise positional relation as the partial standard pattern is extracted from the start end while each position at each fixed time interval is assumed as the true start end of voice for the input voice pattern, and the pattern matching process is conducted between both partial patterns. The start end and terminal end positions of the true voice zone in the input voice pattern are determined from the position of the partial input pattern capable of obtaining the minimum value of the distance between both partial patterns, and matching is conducted between the true voice zone of the input voice pattern and the standard pattern.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a word voice recognition method and an apparatus for implementing this method, and more particularly to a word voice recognition method for inputting voice in word units and outputting a recognition result and an apparatus for implementing this method. Regarding

[0002]

2. Description of the Related Art A conventional example will be described with reference to the drawings. BACKGROUND ART Various researches and developments have been conventionally performed on a voice recognition device used in a situation where a command is input by voice instead of operating an electric device or other mechanical device by a human hand.

Ideally, the voice recognition technique is capable of recognizing a voice of arbitrary length produced by a human from an arbitrary place at an arbitrary timing with a probability of 100%. However, in order to achieve perfect recognition performance, the speech recognition technology is required to accurately capture speech uttered at arbitrary timing in a real acoustic environment in which various noises are present. As a result, in the voice input processing, the start and end of the voice are detected many times from the signal section observed including the noise, and noise and other complicated signal sections such as unnecessary signal sections are excluded. The section detection algorithm requires a processing operation that is always executed, and the amount of calculation required for the processing is enormous. For this reason, as a simple speech recognition technique, an isolated word speech recognition method is adopted, which detects the beginning and end of speech only once during a certain period of time.

An isolated word speech recognition apparatus will be described with reference to FIG. In FIG. 4, the voice input unit 1 is a place where a voice such as a microphone is converted into an electric voice waveform and input. The conversion unit 2 is to convert a voice waveform into a digital value as preprocessing of voice analysis. The voice feature extraction unit 3 is to extract a feature required for voice recognition from the voice waveform at a fixed time interval, that is, for each short frame, by performing a short-time spectrum analysis of the voice waveform. The voice section detection unit 5 is to determine only one start point and one end point of the voice based on the voice feature amount obtained from the voice feature extraction unit 3. The start switch unit 4 is to give a trigger for starting the start edge detection from the outside when detecting the voice section. The input pattern storage unit 6 takes in the voice feature amount from the voice start end to the voice end determined by the voice section detection unit 5 from the voice feature extraction unit 3 and uses this as the input voice pattern. The standard pattern storage unit 7 is a storage unit that stores a plurality of word voice patterns stored in a procedure similar to that of the input pattern storage unit 6 and used for voice recognition given a label name. The pattern matching unit 8 performs a matching process between an unknown input voice pattern stored in the input pattern storage unit 6 and each standard pattern stored in the standard pattern storage unit 7, and outputs the input voice pattern and each standard pattern. It is about to output the distance value to the standard pattern. The distance comparison unit 9 accumulates and compares the distance values output by the pattern matching unit 8 for each standard pattern to be matched, and as a result, finds the smallest distance value from the matching results of each standard pattern for one unknown input pattern. I'm getting it. Result output unit 10
Is a place where the label name of the standard pattern having the smallest distance value among the distance values output from the distance comparison unit 9 is output to the upper host that operates the voice recognition device.

The operation of the voice recognition apparatus shown in FIG. 4 will be described below. It is assumed that the standard pattern analyzed and prepared in the same manner as the input voice pattern is registered in the standard pattern storage unit 7 in advance. The voice is always 10 to 30 m through the voice input unit 1, the conversion unit 2, and the voice feature extraction unit 3.
It is input and analyzed for a fixed time interval of about sec, that is, for each short-time frame, and a part of the information of the analysis result, for example, the logarithmic power of the voice signal is sent to the voice section detection unit 5 and is used as the voice section detection information. It Here, it is assumed that the start switch unit 4 is driven by the operation of the speaker or the host host operating the voice recognition device, and a trigger for starting the voice section detection is generated. As a result, the voice section detection unit 5 is initialized, and the voice start end of the information input from the voice feature extraction unit 3 is detected. As a method of detecting the voice start point, for example, there is a method of setting the rising position of the signal power value as the start point when the signal power value continues from a state without voice to a large value of a certain threshold value or more for a predetermined time. After that, the voice section detection unit 5 detects the attenuation point of the signal power value of the voice and uses this as the end of the voice, and ends the operation. The analysis result of the voice feature extraction unit 3 for the section from the beginning to the end of the voice detected in this way is stored as an input voice pattern in the input pattern storage unit 6. At the time when the storage is completed, the pattern matching unit 8 compares the contents of the input voice pattern stored in the input pattern storage unit 6 and the contents of each standard pattern stored in the standard pattern storage unit 7 with D
Distance matching is performed by matching using P matching or another pattern matching method. The distance calculation result for each standard pattern is sorted in the order of the smallest distance value in the distance comparison unit 9, and the label name of the standard pattern having the smallest distance value is output to the upper host via the result output unit 10.

However, such an isolated word voice recognition apparatus requires a technique for accurately detecting a voice section. For the purpose of voice section detection, there is a method of inputting all the information obtained by the voice feature extraction section to a neural network or other filtering operation section to accurately extract only the voice section. The calculation method used for this extraction Is a large-scale one and requires a calculation amount that is not so different from so-called sentence-speech recognition. Therefore, in general, when trying to detect a voice section with a small amount of calculation, the voice section is detected based on voice power, the number of zero crossings, and other simple information. Further, in order to prevent a part of the voice section from being undetected at the voice section detection stage, a method of collectively detecting short noise sections sandwiched between two voice sections as one voice section at the time of section detection. Can be adopted.

FIG. 5 shows a schematic diagram of a voice cut out by such a voice section detection method. This is an example in which a voice section is cut out by focusing on the voice signal power. In FIG. 5, the section (1) is noise recorded when the lip movement is started, and is generally called lip noise. Section (2)
Indicates a true voice section intended for detection. The section (3) shows the breathing sound received after utterance, and the section (4) shows the ambient noise or the noise on the transmission generated from the voice input section 1 to the conversion section 2. FIG. 5 shows that when the voice input unit 1 is close to the mouth of the speaker such as a handset of a telephone, respiratory sounds or lip noises are attached before and after the true voice section, ambient noise or transmission system. It also indicates that noise may occur when it is erroneously determined as a voice segment. In this way, in the case where an unnecessary signal section other than the true speech section occurs, even if the standard pattern is the same as the true speech section, pattern matching is performed including the unnecessary signal section. From the execution of the process, a discrepancy occurs between the two patterns, resulting in a speech recognition error. Such a state of section detection that causes a voice recognition error is generally called a voice section detection error.

False recognition due to a voice section detection error must be avoided by adjusting the input voice pattern. The reason is that the standard pattern is created by the speaker's careful utterance so that the optimum recognition rate is obtained, or it is automatically generated on the computer. , While the section detection error is excluded, the input voice pattern is for the voice recorded under the actual environment, so the cause and the occurrence state of the section detection error are different for each utterance. Is. Further, the voice recognition device is also required to be an avoidance method that works effectively even in an unknown acoustic environment that cannot be assumed in advance.

A so-called word spotting method is used as a method for avoiding erroneous recognition caused by an error in detecting a section of an input voice. That is, for the input voice, a section that seems to contain the voice is roughly detected in advance, and each position in the section at constant time intervals is considered as the start of the true input voice, and the input pattern is separated from the standard pattern. The terminal-free pattern matching is repeated with, and the minimum distance value obtained as a result is used as the matching result between the two patterns. However, this method repeats matching corresponding to the length of the speech segment roughly cut out, and the amount of calculation becomes enormous.

[0010]

As described above, the word voice recognition device intended to be executed with a small amount of calculation scale simplifies the voice section detection process, and therefore, it is not necessary to use a section other than the true voice section. As a result of section detection in which the signal section is attached to the front and rear, a correct recognition result cannot be obtained for such a voice. In order to solve this problem, an algorithm that requires a huge amount of calculation must be used.

Here, the effect of the word spotting technique, which is an effective method for avoiding erroneous recognition due to section detection error, will be verified as follows. FIG. 6 is a diagram illustrating an example of word spotting. This method performs end-free matching of an input voice pattern in the horizontal axis direction detected as a long section and a short standard pattern in the vertical axis direction with each position at a constant time interval in the input voice pattern as a starting point. , The distance value is calculated. However, it can be seen that the range in which the pattern matching between the two patterns is performed extends over the entire section where the two patterns intersect, as shown by the shaded area in FIG. 6, and a huge amount of calculation is required. The distance value calculated as the matching result changes so as to take the minimum value in the partial section having the highest similarity to the standard pattern, as in the example of the graph in the upper part of FIG. The effect of word spotting is that pattern matching itself also serves as voice segment detection, and a sufficient condition is used that the segment where the distance value is the minimum value is the true voice segment.

However, in pattern matching, as typified by the DP matching method, if the voice section can be roughly estimated, fluctuations in the voice start and end portions,
It is possible to absorb fluctuations in time expansion and contraction between patterns. Using this characteristic, the speech segment is roughly estimated by the same procedure as word spotting using the segment where the characteristic of speech in the standard pattern appears, and then the input estimated as the entire segment of the standard pattern is estimated. If pattern matching is executed between the sub-sections of the voice pattern, the voice recognition performance comparable to that of the conventional word spotting can be obtained with a smaller calculation amount as compared with the word spotting.

According to the present invention, even when a section detection result accompanied by an unnecessary signal section before and after the input speech occurs, a simple word spotting algorithm using a section of a standard pattern in which the characteristic of the speech appears. The present invention provides a voice recognition device that can accurately recognize a voice without enlarging the calculation amount by estimating a voice section and performing matching.

[0014]

A word for matching is derived by deriving a sub-segment having a high similarity to a sub-segment of each standard pattern to be compared from an input voice pattern, and using this as a clue to estimate a true voice segment. A voice recognition method was constructed.
Then, in the above word speech recognition method, for each standard pattern to be compared, a partial standard pattern that is a partial section in which the characteristics of the speech are present is extracted in advance,
With respect to the input speech pattern to be recognized, assuming that each position at constant time intervals is the starting end of the true speech, a partial input pattern that is a partial section having the same temporal positional relationship as the partial standard pattern is extracted from the starting end. , The start and end positions of the true voice section where the minimum value of the distance between the two partial patterns is obtained by performing the pattern spotting process between the two partial patterns, and between the true voice section of the input voice pattern and the standard pattern. We constructed a word speech recognition method that matches in.

A voice input unit 1 for inputting a voice signal is provided, a voice feature extracting unit 3 for extracting a voice feature pattern from the input voice signal is provided, and voice feature pattern information output by the voice feature extracting unit 3 is used. An input voice pattern storage unit that includes a voice section detection unit 5 that detects a voice section based on the voice section detection result, determines the start end and the end of the voice section based on the voice section detection result, and stores the voice feature patterns of the sections indicated by these ends. 6, a standard pattern storage unit 7 storing standard patterns used for speech recognition, and a partial standard pattern extraction unit 11 for extracting a partial section pattern in which a characteristic of a voice appears from each stored standard pattern. And a portion having the same time relationship as the partial interval pattern of the standard pattern, starting from each position of the input speech pattern at constant time intervals Comprising a partial input pattern extraction unit 12 for extracting between patterns, comprising the partial pattern spotting unit 13 performing pattern matching between the partial period of the reference pattern and the input speech pattern subintervals,
The partial pattern spotting unit 13 is provided with a section position determination unit 14 that determines the position where the distance value between the partial section pattern and the input voice pattern has a minimum value as the start and end positions of the true voice section. A pattern matching unit 8 for performing pattern matching between the standard pattern and the input voice pattern based on the positional relationship information obtained from the position determination unit 14 and outputting a distance value is provided, and between the standard pattern and the input voice pattern. A distance comparing unit 9 for accumulating the distance values output as the matching result of the standard pattern and specifying the standard pattern of the minimum distance value, and a result output unit 10 for outputting the label name of the standard pattern determined to be the minimum distance value. We constructed a word speech recognition system.

[0016]

BEST MODE FOR CARRYING OUT THE INVENTION According to the present invention, first, each standard pattern used for voice recognition is registered, and the partial sections in which the characteristics of the voice appear are extracted from each standard pattern. Next, input the voice to be recognized, focus on simple information such as signal power, detect the voice section,
Use the input voice pattern. Here, a partial section having the same temporal positional relationship as the partial standard pattern is extracted for each frame with a constant interval from the start end to the end of the detected input voice pattern, and the part of each standard pattern previously extracted is extracted. A simple pattern matching is performed with the section to obtain the cumulative distance value between the two partial section patterns. As a result of this pattern matching, it is possible to estimate the position of the partial section in which the cumulative distance value has the minimum value for the input voice pattern. Based on this start position information, the section length is assumed to be the same as the standard pattern to be compared, the section position of the input voice pattern to be matched is specified, and matching between the entire standard pattern section and the specified input pattern portion is performed. To obtain the distance value. As a result of repeating these processes for each standard pattern, the standard pattern having the smallest cumulative distance value can be obtained as the recognition result, and the recognition error due to the voice section detection error can be avoided.

[0017]

Embodiment An embodiment of the present invention will be described with reference to FIG. In FIG. 1, a voice input unit 1 is where a voice is input and uses an audio microphone and an audio input terminal. The conversion unit 2 is to convert a voice waveform into a digital value as preprocessing of voice analysis. The voice feature extraction unit 3 analyzes the voice waveform obtained by the conversion unit 2 for a short time to obtain a feature amount required for voice recognition from the voice waveform at a fixed time interval of about 10 to 30 msec, that is, for each short time frame. A method such as short-time logarithmic power analysis and cepstrum analysis is adopted as the analysis method for extraction. The activation switch unit 4 is to give a trigger for starting the start edge detection at the time of detecting the voice section which is required to realize the isolated word voice recognition. The voice section detection unit 5 is to determine only one start point and one end point of the voice based on the voice feature amount obtained from the voice feature extraction unit 3, and the detection method is as follows.
It is possible to adopt a method in which the noise level before the voice is generated is measured in advance, and a section in which a signal component having a logarithmic power value equal to or higher than a certain threshold derived from the noise level transits within a certain period is set as the voice section . Further, when there are two sections that exceed a certain threshold value across a short section that is less than the threshold value, it is possible to adopt a method of considering these three sections as one partial section. The input pattern storage unit 6 takes in the voice feature amount from the voice start end to the voice end determined by the voice section detection unit 5 from the voice feature extraction unit 3 and uses this as the input voice pattern. The standard pattern storage unit 7 is a storage unit that stores a plurality of word voice patterns stored in a procedure similar to that of the input pattern storage unit 6 and used for voice recognition given a label name.

The partial standard pattern extraction unit 11 added according to the present invention extracts a partial standard pattern, which is a partial section in which a voice feature appears, from the standard pattern storage unit 7, and a partial pattern spotting unit described later. Thirteen
It is used for matching. Similarly, the partial input pattern extraction unit 1 added according to the present invention
2 is for extracting a partial input pattern which is the same partial section as that extracted by the partial standard pattern extraction unit 11 while shifting the position of the voice pattern input from the input pattern storage unit 6 from the start end to the end at regular intervals. I am about to do it. The partial pattern spotting unit 13 added by the present invention is a partial standard pattern extraction unit 11
A simple pattern matching is performed between the partial section pattern output by the above and the partial section pattern output by the partial input pattern extraction unit 12, and the distance value between both partial section patterns is output. Section position determination unit 14
This is also a configuration added by the present invention. The distance value output from the partial pattern spotting unit 13 is accumulated and compared for each extraction position of the partial input pattern, and the position of the input pattern having the minimum distance value is specified. Further, the position of the input voice pattern section to be collated in the pattern matching section 8 is specified by further referring to the standard partial pattern to be compared. Pattern matching unit 8
Between the unknown input voice pattern stored in the input pattern storage unit 6 and each standard pattern stored in the standard pattern storage unit 7 based on the information of the section position determination unit 14 Matching is performed and a distance value to the input voice pattern is output. The pattern matching method is well known as a voice recognition pattern matching method.
The P matching method can be adopted. Distance comparison unit 9
Is to accumulate and compare the distance values output by the pattern matching unit 8 for each standard pattern to be matched, and as a result, obtain the minimum distance value from the matching results of each standard pattern with respect to one unknown input pattern. The result output unit 10 derives the standard pattern having the smallest distance value out of the distance values output from the distance comparison unit 9, and outputs the label name of the standard pattern to the host host operating the speech recognition apparatus. .

The operation of the voice recognition apparatus shown in FIG. 1 will be described below. It is assumed that the standard pattern analyzed and prepared in the same manner as the input voice pattern is registered in the standard pattern storage unit 7 in advance. The voice is always 10 to 30 m through the voice input unit 1, the conversion unit 2, and the voice feature extraction unit 3.
It is input and analyzed for a fixed time interval of about sec, that is, for each short-time frame, and a part of the information of the analysis result, for example, the logarithmic power of the voice signal is sent to the voice section detection unit 5 and is used as the voice section detection information. It Here, it is assumed that the start switch unit 4 is driven by the operation of the speaker or the host host operating the voice recognition device, and a trigger for starting the voice section detection is generated. As a result, the voice section detection unit 5 is initialized, and the voice start end of the information input from the voice feature extraction unit 3 is detected. As a method of detecting the voice start point, for example, there is a method of setting the rising position of the signal power value as the start point when the signal power value continues from a state without voice to a large value of a certain threshold value or more for a predetermined time. After that, the voice section detection unit 5 detects the attenuation point of the signal power value of the voice and uses this as the end of the voice, and ends the operation. The analysis result of the voice feature extraction unit 3 for the section from the beginning to the end of the voice detected in this way is stored as an input voice pattern in the input pattern storage unit 6.

When the storage of the input voice pattern in the input pattern storage unit 6 is completed, matching between this input voice pattern and each registered standard pattern is started. First, in the partial standard pattern extraction unit 11,
A partial standard pattern, which is a partial section in which the characteristic of voice appears, is extracted from the standard pattern and supplied to the partial pattern spotting unit 13. A method of extracting the partial standard pattern will be described with reference to FIG. Figure 2 shows the entire standard pattern
As in (a). In order to simplify the explanation, the voice waveform is expressed only by the logarithmic voice power. As an example of extraction of the sub-sections, one or more sub-sections in which the characteristics of the voice as shown by the slanted line in (b) appear, short start and end sub-sections shown by the slanted line in (c),
It is preferable to adopt a partial section at a position which is graphically easy to calculate, such as a position corresponding to 1/4 and 3/4 of the total length of the standard pattern, which is indicated by the diagonal line in (d).
In this case, the mutual positional relationship of the partial standard patterns is maintained while maintaining the positional relationship on the original standard pattern section, that is, time expansion / contraction is not considered.

Next, in the partial input pattern extraction unit 12, the section from the start end to the end of the input voice pattern stored in the input pattern storage unit 6 is subdivided into fixed time intervals in frame units. Each frame at this time is assumed to be the start end of the true input voice pattern, and a partial input pattern that is a partial pattern at the same position as the partial standard pattern is extracted from the start end. In this way, the partial pattern spotting unit 13 performs matching between the partial standard pattern extracted by the partial standard pattern extracting unit 11 and the partial input pattern extracted by the partial input pattern extracting unit 12. As a matching method for each partial section, for example, the cumulative distance value of the Euclidean distance of the cepstrum is used. The distance calculation result is sent to the section position determining unit 14, and the section position determining unit 14 looks at the transition of the cumulative distance value for each hypothetical start end of the input voice pattern scanned by the partial input pattern extraction unit 12. An example of this transition is shown in FIG. In FIG. 3, the region requiring the spectral distance calculation for the partial pattern spotting between the partial standard pattern and the input voice pattern is only the shaded portion in the central frame of FIG. By comparison, it can be seen that the calculation area is obviously smaller. As a result of the partial pattern spotting, the distance values between the partial sections change as shown in the upper part of FIG. 3, but the positional relationship in which the total length of the standard pattern and the true input voice part of the input voice pattern almost match is shown. When taken, the distance value becomes a minimum value. The section position determining unit 14 determines the assumed start end of the input voice pattern when the minimum value is taken as the start end for the true voice section, and sends the information to the pattern matching unit 8. The pattern matching unit 8 inputs the partial section position information sent from the section determination unit 14, and regarding the total length of the standard pattern supplied from the standard pattern storage unit 7 and the input voice pattern supplied from the input pattern storage unit 6. The matching is performed based on the partial section position information determined by the section determining unit 14. At this time, the section length of the input voice is the same as the standard pattern section length. The matching results are accumulated for each standard pattern in the distance comparison unit 9, arranged in the order of smaller distance values, and sent to the result output unit 10. The label name of the standard pattern having the smallest distance value is output to the upper host via the result output unit 10. In comparison between each standard pattern and the input voice pattern, the input voice pattern length may be shorter than the standard pattern length. In this case, the section position determination unit 14 instructs the pattern matching unit 8 to perform pattern matching between the total length of the input voice pattern and the standard pattern.

With respect to the above algorithm, the results of experiments conducted on actual speech will be described. The recognition target is one of the top 20 words in the 100 words of the city of Japan described in the document "Common Speech Data for Speech Recognition" (author Itabashi, Proceedings of the Acoustical Society of Japan, published in 1985). . Voice is A / D converted at 8kHz through a telephone band (300Hz to 3.4kHz) filter, and 16mse.
LPC for a short frame of 32 msec length for each c
Cepstrum analysis is performed. The speech section was detected by focusing on the short-time logarithmic power. In utterance, in order to clarify the effectiveness of the speech recognition method of the present invention, (1) utterance is made by attaching "Eh" before the recognition vocabulary, and (2) Attaching "is" after the recognition vocabulary. They were uttered in the style of (3) uttering with "er" and "da" before and after the recognition vocabulary. As shown in FIG. 2D, the method of extracting the partial standard pattern is 1/4 from the starting end and 1/4 from the ending end of the standard pattern, that is, 3 from the starting end.
The method of extracting only in two sections of / 4 was adopted. As a result, in the conventional recognition method of fixing one voice section, the recognition rates by the utterance styles of (1), (2), and (3) were 10%, 40%, and 5%, respectively. On the other hand, in the method of the present invention in which the true speech segment is estimated from the input pattern and the matching is performed, the recognition rates are 100%, 100% and 85%, respectively, which confirms that the method of the present invention is effective. .

The present invention not only intentionally attaches unnecessary voices as in this experiment, but also voices lip noises, breath sounds, background noises, and other noises generated regardless of the intention of the speaker. Speech recognition can be performed by excluding it from the section.

[0024]

As described above, the present invention relates to a word voice recognition device for performing pattern matching after detecting a voice section, unnecessary voices, lip noise, and A recognition error caused by the addition of breath noise, background noise, or other noise is avoided by adding a simple operation to the voice section detection algorithm.

[Brief description of drawings]

FIG. 1 is a block diagram illustrating an embodiment.

FIG. 2 is a diagram illustrating a method of extracting a partial pattern from a standard pattern.

FIG. 3 is a diagram showing a transition of a cumulative distance value with respect to a position of an input voice pattern when pattern spotting is performed.

FIG. 4 is a block diagram illustrating a conventional example.

FIG. 5 is a diagram illustrating a signal phenomenon that occurs when a voice section is detected.

FIG. 6 is a diagram showing a transition of a cumulative distance value with respect to a position of an input voice pattern when performing word spotting.

[Explanation of symbols]

 1 voice input unit 2 conversion unit 3 voice feature extraction unit 4 start switch unit 5 voice section detection unit 6 input pattern storage unit 7 standard pattern storage unit 8 pattern matching unit 9 distance comparison unit 10 result output unit 11 partial standard pattern extraction unit 12 Partial input pattern extraction unit 13 Partial pattern spotting unit 14 Section position determination unit

Claims (3)

[Claims] 1. A word speech recognition method characterized by deriving a partial segment having a high similarity to a partial segment of each standard pattern to be compared from an input speech pattern to estimate a true speech segment and performing matching. . 2. The word voice recognition method according to claim 1, wherein for each standard pattern to be compared, a partial standard pattern that is a partial section in which a characteristic of a voice appears is extracted in advance, and a recognition target is obtained. Assuming that each position at constant time intervals is the start of the true voice, a partial input pattern that is a partial section having the same temporal positional relationship as the partial standard pattern is extracted from the input voice pattern. The start and end positions of the true voice section in the input voice pattern are determined from the position of the partial input pattern where the minimum value of the distance between both partial patterns is obtained by performing the pattern matching process between the partial voice patterns. A word voice recognition method characterized by performing matching between a true voice segment and a standard pattern. 3. A voice input unit for inputting a voice signal, a voice feature extraction unit for extracting a voice feature pattern from the input voice signal, and based on voice feature pattern information output by the voice feature extraction unit. A voice section detection unit for detecting a voice section, and an input voice pattern storage section for determining the start and end of the voice section based on the voice section detection result and storing the voice characteristic pattern of the section indicated by these ends. In addition, a standard pattern storage unit that stores standard patterns used for voice recognition is provided, and a partial standard pattern extraction unit that extracts a partial section pattern in which the characteristic of the voice appears from each stored standard pattern is provided. A partial section pattern having the same time relationship as the partial section pattern of the standard pattern is extracted starting from each position of the voice pattern at a constant time interval. A partial pattern spotting unit for performing pattern matching between a partial section of the standard pattern and a partial section of the input voice pattern, and a partial section pattern based on the matching result of the partial pattern spotting section. And the input voice pattern have a minimum distance value as the start and end positions of the true voice section, and are provided with a section position determination unit, which is a standard based on the positional relationship information obtained from the section position determination unit. The pattern matching unit that performs pattern matching between the pattern and the input voice pattern and outputs the distance value is provided, and the distance value output as the matching result between each standard pattern and the input voice pattern is accumulated to obtain the minimum distance value. Equipped with a distance comparison unit that specifies the standard pattern of Word speech recognition apparatus characterized by comprising a result output unit for outputting the emission of the label name.