JPH02165199A - voice recognition device - Google Patents

Abstract

(57) [Summary] This bulletin contains application data before electronic filing, so abstract data is not recorded.

Description

[Detailed description of the invention] Separation of technical supervision The present invention relates to a speech recognition device.

In general, speech recognition devices often use BPF and DET to detect the power distribution in the frequency direction, because the amount of calculation is too large to perform FFT.
This is because data cannot be processed in real time. Furthermore, with regard to digital filters, increasing the number of channels is not practical because the scale of the hardware becomes too large.The number of channels of this BPF differs mainly depending on the band they cover. There are various ways of thinking regarding the sounds targeted by speech recognition.

In most cases, the BPF is set to cover a frequency range of about 200Hz to 7KHz. This band is covered by an n-channel BPF. However, when the speech recognition device is used, the input band may be narrow. For example, for telephone audio, the highest frequency is 3,4
The range is up to K I - I z, and there is no power in the band beyond that. Also, in the case of wireless audio, the frequency is often similarly narrow.

In such cases, conventionally B
A voice recognition device with the same number of channels as PF was used. However, no countermeasures could be taken if the user did not know what kind of environment the device would be used in. For example, voice for telephone! When the second recognition device is used with a normal microphone, the recognition rate may decrease because information of 3 or 4 K Fl z or more is not used. On the other hand, when a wide-band audio device is used in the telephone band, it samples inputs in unnecessary bands and generates dictionaries in unnecessary bands, resulting in long U and νurJ. It has the disadvantage that it may be difficult to use, and it requires a large amount of memory.

1-1 The present invention was made in view of the above-mentioned circumstances,
In particular, this was done for the purpose of reducing recognition time and memory used by dictionaries by automatically examining data in bands that are not needed under the usage conditions of speech recognition devices. .

In order to achieve the above object, the present invention also provides means for amplifying input audio to an appropriate level.

An nch bandpass filter (BPF), an nch rectifier circuit (DET), means for time-divisioning their outputs into serial data, an A/D conversion means, a means for extracting features, and an audio means for cutting out sections, means for storing a dictionary, means for recognition, and means for storing power data from Ich to nch for each sampled frame.

A speech recognition device having means for calculating channel power from the data, means for storing the calculated value, means for comparing the value with a certain threshold value, and means for indicating whether each channel is valid or invalid based on the result. When the device is in use, the user makes a predetermined utterance, detects the power of each bandpass filter, and detects the bandpass filter in the band that has no power in the usage condition. In this case, the input data is reduced without performing A/D conversion of the output of the bandpass filter, or the band is not registered in the dictionary. Hereinafter, the present invention will be explained based on examples.

FIG. 1 is an elliptic diagram for explaining one embodiment of the present invention, in which 1 is an amplifier, 2 is a group of band-pass filters (B
PF), 3 is an A/D converter, 4 is a speech section detector, 5 is a feature extraction unit, 6 is a recognition unit, 7 is a dictionary storage unit, 8 is a power storage unit, 9 is a channel power calculation unit, 10 is a Channel power storage unit, 11 is a comparator, 12 is a channel enable flag, and 13 is a test mode switch.

Before performing dictionary registration or recognition using the speech recognition device, the user uses the test mode switch 13 to put the device into a test mode in order to inform the recognition device of the state of the input speech. The switch 13 may be one that can be pressed directly by the user; for example, in the case of a telephone or the like, the switch 13 may be set to a test mode by pressing a button on a telephone, or the mode may be switched wirelessly or the like.

When the device enters the test mode, it detects a voice section from the input voice and stores data of the detected voice section in the power storage section. The stored data is from Ich to nch
The output of DET up to, that is, the power value of each BPF is
It is a D-converted value, and the sample time is registered in a dictionary or P! Make it equal to the time of awareness.

The audio used for this test must have components ranging from low frequencies to high frequencies.

For example, if the frequency is high, it may include the consonant 191, and if the frequency is low, it may include 'n, 11'. The test word to be uttered may be specified by the user in advance, and may be any word that includes the above-mentioned sounds. The power data of this test word is added for each channel and for all frames within the speech section. This added power value is stored in the channel power storage section 1o.

The range of B P F to be used is determined by this power value.

FIG. 2 is a diagram for explaining an example of means for determining the range of BPF to be used, and the channel power storage section 10 stores the power within a word for each channel.

Here, an explanation will be given assuming that the input voice is from a telephone. Telephone audio generally has no power in the high frequency range. Therefore, when examining the power of the test word, there is almost no power at high frequencies, and the output data of the BPF at high frequencies is close to O. In other words, the BP that does not produce this power
Regarding F, it can be said that there is no audio information at all. Therefore,
A certain threshold value m is compared with the value in the channel power storage unit 10, and if it is larger than m, the channel is determined to be valid, and conversely, if it is smaller, it is determined to be invalid. The threshold value is used to remove the influence of noise and may be determined experimentally, and even if the value is changed by that much, there is no effect. Information on whether the channel is valid or invalid, which is obtained in this way, is indicated by the channel enable flag 12. In this embodiment, for each channel, if it is valid, it is set to 1; if it is invalid, it is set to 0, but other methods may be used. Based on the channel information obtained in this way, the device performs dictionary registration and recognition.

Next, the operation during recognition will be explained. At the time of recognition, the power of BIF is A/D converted at the same time as the voice is input, and pattern matching is performed while taking in the value. For example, when a microprocessor is used, an interconnect is used to capture data and perform feature extraction and matching operations at other times. Normally, it is necessary to input the outputs of all BPFs, but in the present invention, since the validity and invalidity of each channel is already known, it is only necessary to perform feature extraction and pattern matching on the data of the necessary channels. Inputting the data of each channel requires conversion time for A/D conversion and time to input the data, so reducing the number of input channels increases the time for the microprocessor to perform feature extraction and pattern matching. , the recognition time will be shortened.

Next, the operation at the time of dictionary registration will be explained. For example, B
T' S P (+3inary Time-5pe
ctrum Pattern) method dictionary registration will be explained. Generally, information on frequency components is required,
In many cases, the dictionary has the same number of channels as "B PF". The dictionary shown in Figure 3 (a) is a dictionary when n channels are used. In this case, the number of channels rl and frame yJ, 1 and one channel. A memory of a size equal to the number of bits of data (in this case, 2 bits 1~) is required.In contrast, the dictionary of this device (Figure 3 (b), which knows whether each channel is valid or invalid) is required. )) reduces the channel data to rn, reducing the memory required by the dictionary.

By storing the above-mentioned channel enable flag information as part of the data when storing the dictionary on a floppy disk or the like, only a -degree test is required.

肱 - As is clear from the above description, in the speech recognition device according to the first item of the claims, high-speed recognition is possible because the input from the I3PF is reduced. Furthermore, in the speech recognition device according to the second aspect of the present invention, since bands with no power are not registered in the dictionary, the memory of the dictionary can be reduced.

[Brief explanation of the drawing]

FIG. 1 is a diagram for explaining one embodiment of the speech recognition device according to the present invention, FIG. 2 is a diagram for explaining an example of means for determining the range of BPF to be used, and FIG. , B.T.
FIG. 3 is a diagram for explaining an example of dictionary registration using the SP method. 1... Amplifier, 2... Bandpass filter group, 3
... A/D converter, 4... Voice section detector, 5...
・Feature extraction unit, 6... Recognition unit, 7... Dictionary storage unit,
8... Power storage unit, 9... Channel power calculation unit, 10... Channel power storage unit, 11... Comparator. 12...Channel enable flag, 13...Test mode switch. Patent applicant Ricoh Co., Ltd.

Claims (1)

[Claims] 1. A means for amplifying input audio to an appropriate level, an nch bandpass filter (BPF), and an nch bandpass filter (BPF);
rectifier circuit (DET), means for time-dividing their outputs into serial data, A/D conversion means, means for extracting features, means for cutting out voice sections, and means for storing a dictionary. , means for recognizing, means for storing power data from Ich to nch for each sampled frame, means for calculating channel power from the data, means for storing the calculated value, and means for storing the calculated value. In a speech recognition device having a means for comparing with a certain threshold value and a means for indicating whether each channel is valid or invalid based on the result, when the speech recognition device is used, a voice predetermined by the user is made. Then, detect the power of each bandpass filter, detect the bandpass filter in the band that has no power in the usage state, and use the A/D of the output of the bandpass filter for that band.
A speech recognition device characterized by reducing input data without performing conversion. 2. A means for amplifying input audio to an appropriate level, an nch bandpass filter (BPF), and an nch
rectifier circuit (DET), means for time-dividing their outputs into serial data, A/D conversion means, means for extracting features, means for cutting out voice sections, and means for storing a dictionary. , means for recognizing, means for storing power data from Ich to nch for each sampled frame, means for calculating channel power from the data, means for storing the calculated value, and means for storing the calculated value. In a speech recognition device having a means for comparing with a certain threshold value and a means for indicating whether each channel is valid or invalid based on the result, when the speech recognition device is used, a voice predetermined by the user is made. A speech recognition device characterized in that the power of each bandpass filter is detected to detect a bandpass filter having no power in a used state, and that band is not registered in a dictionary.