JPS6334479B2 - - Google Patents

Description

[Detailed description of the invention]

<Technical Field> The present invention relates to a simple speech recognition device. <Prior art> A characteristic of Japanese speech is that a vowel is added to every syllable, so in a limited number of sentences and words,
Even when consonants are ignored and the vowel sequence is used, a large portion can still be deciphered. Vowels have a lot of energy and last a long time, so they are less affected by noise. Therefore, it is extremely important to accurately recognize vowels in Japanese speech. However, the formant, which is important for vowel recognition, differs depending on gender and age, and the vowels in continuous speech also differ depending on the preceding and following phonemes and the rate of speech, and the range of five vowels has expanded considerably. Taking this into consideration, it is necessary to consider the following two methods. (i) Increase frequency resolution by using multiple filters. Then, the formant frequency is normalized. Although this method has the disadvantage of complicating subsequent processing, it can be applied even if the range of variation is quite wide. (ii) Consider expanding the frequency range of the filter to absorb variations. Results of statistical experiments in method (ii) above, 5
Five filters were found to be sufficient for vowel identification. An example is shown below.

[Table] Filter No. 1 and No. 2 are in the first formant region,
Filter No. 3 is the coexistence region of the first or second formant, which differs depending on the vowel. Filter No.4,
No. 5 is the second formant region. After appropriately weighting the outputs of these five filters, the vowels are identified by comparing the filters with each other. This is a method that can identify five vowels over a fairly wide range. However, in continuous speech, such five vowels alone are insufficient, and other information is required. <Object of the Invention> In view of the above points, the present invention provides a simple speech recognition device with a simple configuration that can be effectively realized using the filter in the method (ii) above. <Example> An example of the present invention will be described below with reference to the drawings. FIG. 1 is a block diagram showing an embodiment of the speech recognition device of the present invention. An audio signal is input through a microphone 1 and enters an amplifier 2. Amplifier 2 is equipped with an automatic volume adjustment circuit and can be used as required. Also, the frequency characteristics of this amplifier 2 are high frequency emphasized (+6dB/octave) starting from about 1KHz. The output of amplifier 2 is input to filter group 3. This is composed of the five filters mentioned above. Each filter No.1
The output of No. 5 is input to the envelope detection circuit 4.
In addition, the envelope detection circuit 4 includes a circuit that receives direct input from the preamplifier 2, and consists of six circuits in total. This circuit has a low pass filter circuit and has a 20
The cutoff frequency is set to ~50Hz. The output of the envelope detection circuit 4 is input to a vowel/voiceless consonant/silence discrimination circuit 5 and a filter selection circuit 6. The vowel/voiceless consonant/silence discrimination circuit 5 is a circuit that discriminates five vowels, voiceless consonants, and silence (including voiced consonants as a result) by mutually comparing each output of the envelope detection circuit 4, and the discrimination state is as follows. It is output as a 3-bit digital signal. On the other hand, the filter selection circuit 6 divides each output of the envelope detection circuit 4 corresponding to the five filter outputs into two groups, that is, filter numbers, as described above.
1, No.2, No.3 and filter (No.3), No.4, No.
5 (filter No. 3 of the latter group is included or excluded depending on the selection of the former group), selects the filter with the maximum output in the two groups, and controls the switch circuit 7. As a result, the two input signals are inputted to the formant frequency transition portion detection circuit 8 as two inputs. That is, filters No. 1, No. 2, and No. 3 are divided into two groups, with filters No. 1, No. 2, and No. 3 as a first formant region, and filters No. 3, No. 4, and No. 5 as a second formant region. Select the filter with the maximum output among the three. If the maximum output is not filter No. 3, then the second
Select the filter that produces the maximum output among the three filters in the formant region. If No. 3 of the filters in the first formant region shows the maximum value, the filter with the larger output is selected from No. 4 and No. 5 for the second formant region. By the above method, it is possible to automatically select a filter in which the first and second formants are effectively present. The output of this filter is input to a zero-crossing counting circuit, and the temporal change in the output is observed. It is also known experimentally that this output estimates the first and second formant frequencies. Therefore, since this temporal change in the number of zero crossings expresses the transient part of the formant, it takes the characteristics of the temporal change from the stationary vowel, and the transient part of the formant before entering the stationary vowel is ``crossed'' and the stationary vowel is The “emergence” of formant transients after vowels
It is characterized by syllables, consonants, vowels, and consonant chains. Figure 2 shows the formant frequency transition part detection circuit 8.
(the same applies to the first formant part and the second formant part), and the selected filter output is first input to the zero crossing counting circuit 8a. This is a circuit that counts how many times the audio waveform crosses zero within a certain period of time (approximately 10ms). At the next stage, there is a circuit 8b that temporarily stores the counting results (approximately 200ms), and a judgment circuit 8c connected to this judges whether the vowel is a stationary vowel or a transitional part based on the degree of change in the counting within that range. Then, by extracting features such as the degree of change, direction, and time of change, the characteristics of the "crossing" between the consonant and vowel parts are extracted. The output of the formant frequency transition portion detection circuit 8 and the output of the vowel/voiceless consonant/silence discrimination circuit 5 are input to a discrimination unit determination circuit 9. In this way, identification units are selected and determined from vowels, voiceless consonants, voiceless and voiced consonants, formant transition parts, etc. here,
Discrimination units are a type of unit that constitutes sounds that indicate various Japanese vowels, consonants, and transitional parts (watari) that exist between vowels and consonants. For example, in the case of the Japanese word "ta", it is pronounced as "ta", but there is an abrupt change in the spectrum (particularly formant) between the plosive part of "t" and the vowel stationary part of "a". Although there is an unsteady section where the value changes, this section is also one identification unit. The identification unit standard pattern storage circuit 10 stores the characteristic patterns of these identification units in advance, and stores the output of the formant frequency transition part detection circuit 8 and vowels, voiceless consonants,
The identification unit is selected by comparing with the output of the silence identification circuit 5. This selection determines the output that corresponds to the identification unit. The output of the identification unit determination circuit 9 is sent to the identification system matching circuit 1.
1 is input. Standard patterns for each identification unit to be compared are statistically determined from a large number of samples and are stored in the memory circuit 12. The identification series matching circuit 11 compares the data corresponding to the identification unit from the identification unit determination circuit 9 with the standard pattern for each identification unit from the memory circuit, and determines which identification unit the current data corresponds to. do. Then, the discriminating circuit 13 removes the recognition information of the erroneously identified equals, selects the most similar discriminating series by comparing the discriminating series of the speech to be recognized, and recognizes the five vowels in the continuous speech. Incidentally, the above-mentioned device can also be constructed from a general-purpose 8-bit microcomputer or the like. <Effects of the Invention> As described above, the present invention can effectively recognize five vowels in continuous speech using five filter circuits, and is a simple device. In addition, by mutually comparing filter outputs and selecting in advance a filter in which the first and second formants exist, and connecting it to a zero-crossing counting circuit, it is possible to estimate formants with noise resistance, and also to It is a useful combination of methods with excellent resolution.

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing one embodiment of the present invention;
FIG. 2 is a block diagram showing the main parts of FIG. 1 in detail. 3... Filter group, 4... Envelope detection circuit,
5... Vowel/vowel/voiceless consonant/silence identification circuit, 6... Filter selection circuit, 7... Switch circuit, 8...
Formant frequency transition portion detection circuit, 9...Identification unit determination circuit, 10...Identification unit standard pattern storage circuit, 11...Identification series matching circuit, 12...Memory circuit, 13...Identification circuit.

Claims (1)

[Claims] 1. Five filter circuits including a first formant region and a second formant region are provided, and five vowels, voiceless consonants, silence, etc. are identified by mutual comparison of the outputs of the five filter circuits. means for selecting the outputs of the two filter circuits with the maximum outputs in which the first formant and the second formant are present from among the five filter circuits; and the outputs of the two selected filter circuits. means for detecting a formant frequency transition part by zero crossing counting based on the discrimination output of the five vowels, voiceless consonants, silence, etc. and means for determining a discrimination unit based on the detection output of the formant frequency transition part; and the determined discrimination. A speech recognition device comprising: means for performing judgment recognition of continuous speech from a time series of units;