JPH0679232B2 - Voice recognizer - Google Patents

Description

Description: TECHNICAL FIELD The present invention relates to a voice recognition device that performs pattern matching in voice recognition.

2. Description of the Related Art Voice recognition is required to operate a device by voice, and a voice recognition device is being actively developed. Thus, word speech recognition is widely used at present,
The feature pattern of the word voice registered in advance and the feature pattern of the input voice are compared to select a similar pattern, which is used as the recognition result. When comparing these patterns, methods have been proposed for removing the influence of pattern variations, and they are roughly classified into those that use nonlinear stretching and those that use linear stretching. The former can remove the influence of the pattern variation with high accuracy, but has a drawback that the calculation amount is large, while the latter has a disadvantage that the calculation amount is small but not as accurate as the former.

Purpose The present invention has been made in view of the above-mentioned accuracy,
In particular, the present invention has been made for the purpose of providing a voice recognition device capable of performing highly accurate pattern matching with a small amount of calculation.

Configuration In order to achieve the above object, the present invention provides a conversion unit for converting an input voice into a characteristic parameter, a characteristic parameter converted by the conversion unit, and a portion in which the value of the characteristic parameter is smaller than a predetermined value. A dictionary storage unit that stores the existence position, a silent parameter determination unit that finds the existence position of a portion where the value of the feature parameter obtained at the same time when an unknown voice is input is less than a predetermined value, and the input voice When there is a portion smaller than the value specified for the characteristic parameter,
The length of the comparison data is made uniform by linearly expanding and contracting the front and rear of the portion in association with the portion and the existing position of the portion in which the value of the characteristic parameter stored in the dictionary storage unit is smaller than a predetermined value. It is characterized in that it comprises a collating unit for later comparing and collating.

FIG. 1 is a configuration diagram for explaining an embodiment of the present invention, in which 1 is a microphone, 2 is a bandpass filter group,
3 is a silent section determination unit, 4 is a matching unit, 5 is a dictionary storage unit,
In the present invention, in order to achieve the above-mentioned object, a portion where each parameter is equal to or less than a certain value at the same time in a voice pattern converted into a characteristic parameter, that is, a silent section is found and recorded together with a registered pattern, If a similar part exists in the input pattern of, that is, if there is a silent section and it exists at the same position on the characteristic pattern, the registered pattern and this part of the unknown input pattern are associated with each other. I try to compare and collate.

In the embodiment shown in FIG. 1, a bandpass filter group is used to extract the characteristic parameter, and the voice input from the microphone 1 is converted into a voice pattern by the bandpass filter group 2. In the silent section discrimination unit 3,
The part where the output of each filter is below a certain value is found and the position is stored. As an example, the threshold value may be determined as an output value of each filter of noise when there is no input voice, and when there is a portion equal to or less than the threshold value in the middle of the input voice, that portion is regarded as a silent section. The matching unit 4 associates the input and the silent sections of the dictionary 5 with each other, and linearly expands and contracts before and after the matching.

2 and 3 show the conventional linear expansion / contraction method (second
Fig.) And the matching method according to the present invention (Fig. 3) for explaining the difference, both of which use lower (/ sita /) as a word example. Thus, if the whole word is linearly expanded / contracted as shown in FIG. 2 for matching, an appropriate correspondence such as the / t / part of the input corresponding to the silent section of the dictionary cannot be made. As shown in Fig. 3, the silent sections are associated with each other,
By performing linear expansion and contraction before and after that, it is possible to perform linear expansion and contraction between the input and / si / between dictionary and / ta /.

Further, as shown in FIGS. 2 and 3, when the interval between the beginning and the end and the silence is considered as one interval of the voice, as long as the same person utters, the length of time is about ± 30%. It is known to stretch. Using this fact, the positions from the beginning of the two patterns of speech to be compared to the silent section are within ± 30% of each, which means that there is a silent section at the same place. It's easy to think of. Therefore, a silent section at the same position means a silent section located at a position within ± 30% of the length from the beginning or end of the pattern to that point. Therefore, in the determination of the silent section in the invention of the present application, only those in these ranges are matched with the silent section, and other matching is not performed.

Effects As is clear from the above description, according to the present invention, in the pattern of the input voice converted into the characteristic parameter, a silent section in which each parameter is equal to or less than a certain value is found at the same time and recorded together with the registered pattern. If there is a silent section at the same position above in the input pattern of, the registered pattern and the silent section of the unknown input pattern are matched and compared, so high accuracy is achieved with a small amount of calculation. Pattern matching is possible.

[Brief description of drawings]

FIG. 1 is a block diagram for explaining an embodiment of a voice recognition system according to the present invention, FIG. 2 is a diagram showing an example of conventional word recognition, and FIG. 3 is an example of word recognition according to the present invention. FIG. 1 ... Mike, 2 ... Bandpass filter group, 3 ...
Silent section determination unit, 4 ... collation unit, 5 ... dictionary storage unit.

Claims (1)

[Claims] 1. A dictionary storage for storing a conversion unit for converting an input voice into a characteristic parameter, a characteristic parameter converted by the conversion unit, and an existing position of a portion in which the value of the characteristic parameter is smaller than a predetermined value. Section, when the unknown voice is input, the value of the characteristic parameter obtained at the same time is a silent section determination section that finds the existing position of the portion that is smaller than the specified value, and the value determined by the characteristic parameter of the input voice. When a small part exists, the part and the front and back of the part are linearly expanded / contracted by associating the part and the existence position of the part in which the value of the characteristic parameter stored in the dictionary storage is smaller than a predetermined value. A speech recognition device comprising: a collating unit for collating and collating after making the lengths of comparison data uniform.