JPH0876789A - System and device for voice recognition unspecified speaker word - Google Patents

Abstract

PURPOSE: To recognize a voice with high precision with a smaller amounts of computations and memory size by using inputted voice pitch and line spectrum pair(LSP) parameters for voice recognition. CONSTITUTION: A power/LSP/pitch detecting section 12 detects the power, the LSP parameters and the pitch of voice signals. A sampling section 14 samples the segment between the start and the end of the uttered voice detected by a voice segment detecting section 13 with a prescribed sampling period. A first dictionary section 16 accumulates the sampled pitch and a second dictionary section 18 accumulates the sampled LSP parameters. A first matching section 15 compares the output of the section 14 and the pitch accumulated in the section 16 and recognizes a word. A second matching section 17 compares the recognition result of the section 15 and the LSP parameters accumulated in the section 18 and recognizes the word.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an unspecified speaker word voice recognition system used for voice dials of mobile phones and the like.

[0002]

2. Description of the Related Art Conventionally, many studies have been made on word speech recognition systems. Generally, a word speech recognition system recognizes speech by comparing an input pattern with a reference pattern stored in a dictionary,
The best match is considered the spoken word. However, even the same word has different speeds and lengths depending on how it is spoken (that is, there are variations), so there are various difficulties in improving recognition accuracy. The systems that have been developed so far include a speaker-specific word speech recognition system and a speaker-unspecified word speech recognition system. An example of a speaker-unspecified word speech recognition system is Satoshi Furui, Tokai University Press. Meeting, "Digital Technology Series Digital Speech Processing", P185-P187. FIG. 3 is a block diagram showing the unspecified speaker word speech recognition system described in this document. The system shown in FIG. 3 is based on a method called a multi-template method, and it is clear that in this method, 12 reference templates are required for each word in order to obtain good recognition accuracy. Therefore, this method has a drawback that the calculation amount and the memory size increase as the number of reference templates for each word increases. If there is no limit on the size or complexity of the device, this method will work, but if you try to adopt this type of word recognition system for something that is small and requires low power consumption, such as a mobile phone, The amount of calculation and the memory size are big problems.

[0003]

As described above, the conventional unspecified speaker word speech recognition system has a large amount of calculation and a large memory size, and is not suitable for a mobile phone or the like which is small and requires low power consumption. . The present invention has been made in consideration of the above problems, and provides a suitable unspecified speaker word speech recognition system such as a mobile phone capable of recognizing speech with a small amount of calculation and memory size while maintaining reasonable recognition accuracy. The purpose is to do.

[0004]

SUMMARY OF THE INVENTION In order to solve the above problems, the present invention is directed to the power of digitalized input voice, LS.
Power / LSP to detect P parameter and pitch
/ A pitch detection section, a voice section detection section for detecting the start and end points of the detected utterance of the input voice, and a voice section detected in the section from the start to the end of the utterance detected by the voice section detection section. A sampling unit that samples the power, the LSP parameter, and the pitch at a predetermined sampling period, a first dictionary unit that stores the pitch sampled by the sampling unit, and the LSP parameter that is sampled by the sampling unit. The second dictionary unit for storing, the first matching unit for recognizing a word by comparing the output from the sampling unit and the pitch stored in the first dictionary unit, and the first matching unit for the first matching unit. Recognition result and L stored in the second dictionary section
A second matching unit that recognizes a word by comparing it with an SP parameter is provided.

The method of the present invention further comprises the steps of detecting the power, LSP parameter, and pitch of the digitized input speech, detecting the start and end points of utterance of the detected input speech, and A step of sampling the power, the LSP parameter, and the pitch detected in a section from the start to the end of the utterance detected by the voice section detection section at a predetermined sampling period; and the sampled pitch in the first dictionary section. And storing the sampled LSP
Accumulating a parameter in the second dictionary unit, recognizing a word by comparing the sampling result with the pitch accumulated in the first dictionary unit, the recognition result and the second dictionary It comprises the step of recognizing a word by comparing it with the LSP parameters stored in the copy.

[0006]

In the above structure, the pitch of the input voice and the LSP parameter are used for voice recognition.
Preliminary selection of words is performed at the first pitch comparison to reduce utterance variations. The next LSP
Since it is possible to reduce the number of times of matching in comparison by parameters, it is possible to recognize voice with higher accuracy and with a small amount of calculation and memory size.

[0007]

An embodiment of the present invention will be described with reference to the drawings.
FIG. 1 is a block diagram showing the configuration of an embodiment of the unspecified speaker word speech recognition system of the present invention. In FIG. 1, 10
Is an analog / digital converter for converting an analog voice into a digital signal, 11 is a pre-processing unit for emphasizing the high frequency region of the input signal, 12 is the power of the voice signal described later, L
Power / LSP / pitch detection unit for detecting SP (Line Spectrum Pair) parameters and pitch, 13 is a voice section detection unit for detecting respective times of start and end of input voice, and 14 is a voice section detection unit 13 The sampling unit that samples the interval from the start to the end of the utterance detected by the sampling unit at a predetermined sampling period, 16 is the first dictionary unit that stores the pitch sampled by the sampling unit 14, and 18 is the resampling unit 14. A second dictionary part that stores the sampled LSP parameters, 1
5 is the output from the sampling unit 14 and the first dictionary unit 16
A first matching unit for recognizing a word by comparing it with the pitch accumulated in the word, and 17 comparing a recognition result of the first matching unit 15 with the LSP parameter accumulated in the second dictionary unit 18 for the word. The second matching unit 19 for recognizing the training process 19 is a training process (indicated by MODE = 1 in FIG. 1) and a recognition process (M in FIG. 1) according to an external input.
A mode switching unit for switching (ODE = 2) is shown.

Next, the operation of each part will be described. Before explaining the operation, the voice input to the system will be described. The voice input to the system is generated by the vibration of the vocal cords. During normal breathing, the vocal cords are wide apart, but when trying to make a voice, the vocal cords approach each other, and the air from the lungs tries to pass through during this, so the interaction between the airflow and the vocal cords The vocal cords open and close in a cyclic manner, resulting in a nearly regular interruption of air. This change in flow rate can be approximated by an asymmetrical triangular wave, which is the source of the audible voice. When the tension in the vocal cords is high and the air pressure from the lungs is high, the opening / closing cycle of the vocal cords, that is, the vibration cycle, becomes shorter, and the pitch of the sound source becomes higher, and vice versa. This corresponds to the pitch. The sound accompanied by vibration of this vocal cord is called a voiced sound, and the sound not accompanied by it is called an unvoiced sound. The voiced sound is a sound having a pitch, but the unvoiced sound has no period and has a pitch value of 0, and is treated as broadband noise by the present system. Therefore, the pitch of any speaker is unvoiced when the value is 0 and voiced when the value is not 0. For example, storms, sneezes, and coughs are treated as unvoiced sounds without pitch. Therefore, when recognizing a voice with reference to the pitch, these unvoiced sounds are excluded, and erroneous recognition is eliminated. The present invention utilizes this characteristic,
This enables more accurate and robust recognition, and extends recognition to untrained speakers while maintaining recognition accuracy. Further, in the present invention, LSP parameters in the frequency domain used are used for recognition. This LSP parameter is used as an approximation of the formant frequency that characterizes the speech. This formant frequency largely depends on the vocal tract that characterizes the utterance of the speaker, and since the candidates are narrowed down by the preliminary selection by the first pitch, the recognition accuracy is improved.

In FIG. 1, the sound generated by the above mechanism is input to the analog / digital converter 10 and converted into a digital signal. The high-frequency region of the digitized voice is emphasized by the preprocessing unit 11. Next, in the power / LSP / pitch detection unit 12, the power of the input voice,
Get LSP parameters and pitch. Here, each value is detected in frame units of a predetermined length.
Pitch and LSP parameters are stored in the buffer in the order in which they were entered. Next, the voice section detection unit 13 detects the start and end points of voice using the detected power value. The voice section detection unit 13 compares the power with a predetermined threshold, and regards the point beyond this threshold as the start and end points.
In the present invention, this voice section is one frame.

Next, the sampling section 14 samples the pitch and the LSP parameter within the section from the start to the end of utterance detected by the voice section detecting section 13 at a predetermined sampling period. This takes into account that the length and speed of utterance differ depending on the individual (for example, even when the same "1 (ichi)" is pronounced, the length and speed of utterance differ depending on the individual or case). This is done to minimize the complexity of calculation due to individual differences. That is, fix the framed parameters in the start / end points to a predetermined number,
Variations are reduced by making the number of samples of this framed parameter the same as the number of samples in the dictionary unit described later. (This process is referred to as resampling) Next, when the mode is the training mode due to an input from the outside, the training switch 19 is set to the state shown in FIG.
MODE = 1 side, that is, the dictionary section side, and the pitch and the LSP parameter are input to the first dictionary section 16 and the second dictionary section 18, respectively, and accumulated as reference data. On the other hand, when the mode is the recognition mode, the training switch 19 is set to MODE in FIG.
= 2 side, that is, it switches to the comparison unit side. The first matching unit 15 compares the input pitch with the reference data stored in the first dictionary unit 16 for similarity. This comparison can be made, for example, by taking an exclusive OR of both data. Alternatively, they can be compared by calculating the Hamming distance between the two. In this case, the one with the smallest Hamming distance is considered the most likely candidate.

As described above, the pitch value can be expressed by a binary value depending on the difference between voiced sound and unvoiced sound. FIG. 2 shows an example thereof. In FIG. 2, 16 points from the start point to the end point of utterance are sampled, and are represented by 1 for voiced sound and 0 for unvoiced sound. In this example, the voice value is set to 1 in the case of voiced sound, but the pitch value may not necessarily be 1 and may be a multiple value. If the pitch value is small, variations are allowed, so there are many candidates to be compared. On the other hand, when the value of the pitch is large, the number of candidates to be compared is small because the preliminary selection is performed according to that value.

After the comparison by the first matching unit 15, the second matching unit 17 compares the LSP parameters. For example, taking the recognition of a single digit as an example, the first matching unit 15 performs comparison with 10 references (0 to 9) stored in the first dictionary unit 16, and The matching unit 17 makes a comparison with the more narrowed-down, for example, about 3 references stored in the second dictionary unit 18, and obtains the final recognition result.

As described above, in the unspecified speaker word speech recognition system of the present invention, the pitch of the input speech and the LSP parameter are used for speech recognition. Then, the words are preselected by comparing the first pitches to reduce variations in utterance. The next LSP
Since it is possible to reduce the number of times of matching in comparison by parameters, it is possible to recognize voice with higher accuracy and with a small amount of calculation and memory size.

The present invention is not limited to the above-described embodiment, but various modifications can be made depending on the application field such as the sampling cycle in resampling and the pitch value.

[0015]

As described above, according to the present invention, a voice can be recognized with high accuracy and with a small amount of calculation and memory size.

[Brief description of drawings]

FIG. 1 is a block diagram showing a configuration of an embodiment of an unspecified speaker word speech recognition system of the present invention.

FIG. 2 is a diagram showing a pitch value in the embodiment of FIG.

FIG. 3 is a block diagram showing an example of a conventional unspecified speaker word speech recognition system.

[Explanation of symbols]

10 ... Analog / digital converter, 11 ... Pre-processing unit, 1
2 ... power / LSP / pitch detection unit, 13 ... voice section detection unit, 14 ... resampling unit, 15 ... first matching unit, 16 ... first dictionary unit, 17 ... second matching unit, 18 ... 2 dictionary section, 19 ... Mode switching section.

Claims (4)

[Claims] 1. The power of an input voice digitized, LS
Power / LSP to detect P parameter and pitch
/ A pitch detection section, a voice section detection section for detecting the start and end points of the detected utterance of the input voice, and a voice section detected in the section from the start to the end of the utterance detected by the voice section detection section. A sampling unit that samples the power, the LSP parameter, and the pitch at a predetermined sampling period, a first dictionary unit that stores the pitch sampled by the sampling unit, and the LSP parameter that is sampled by the sampling unit. The second dictionary unit for storing, the first matching unit for recognizing a word by comparing the output from the sampling unit and the pitch stored in the first dictionary unit, and the first matching unit for the first matching unit. Recognition result and L stored in the second dictionary section
An unspecified speaker word speech recognition system comprising a second matching unit that recognizes a word by comparing it with an SP parameter. 2. The value of the pitch is a binary value "1" when the input is voiced and is a binary value "0" when the input is silent.
The unspecified speaker word speech recognition system according to claim 1. 3. The first matching unit compares the output value from the sampling unit with the pitch value stored in the first dictionary unit based on a Hamming distance calculation result. Item 1. An unspecified speaker word speech recognition system according to item 1. 4. The power of the digitized input voice, LS
Detecting the P parameter and the pitch, detecting the start and end points of the detected utterance of the input voice, and detecting in the section from the start to the end of the utterance detected by the voice section detector. Sampling the sampled power, LSP parameter, and pitch at a predetermined sampling period, storing the sampled pitch in a first dictionary unit, and sampling the sampled LSP parameter in a second dictionary unit. To recognize a word by comparing the sampling result with the pitch stored in the first dictionary section, and the recognition result and the LSP parameter stored in the second dictionary section. And a method for recognizing a word by comparing