JPH0459638B2 - - Google Patents

Description

[Detailed Description of the Invention] (a) Industrial Application Field The present invention relates to a speech recognition device capable of recognizing speech.

(b) Conventional example A conventional speech recognition device predetermines a speech pattern A=(a ₍₁₎ , a ₍₂₎ , ...) consisting of a time series of feature parameters indicating the speech characteristics from the speech signal. Multiple voices are extracted, and an unknown pattern X = (x ₍₁₎ , x ₍₂₎ , ...) consisting of a time series of feature parameters for unknown voices is extracted for each time series A =
(a ₍₁₎ , a ₍₂₎ , ...), when the distance between these two patterns D = |A-X|= 〓 ¹ |a(i)-x(i)| is the minimum The audio of series A was determined to be the unknown audio at this time.

However, although time series of voice spectral values, time series of autocorrelation coefficients, etc. are used as the above-mentioned voice characteristic parameters, these parameters vary to some extent depending on the speaking situation of the speaker. For this reason, even if the voices are the same, there is a large distance between the patterns of the pre-registered voice and the unknown input voice, resulting in a drawback of erroneous recognition.

(c) Object of the Invention The present invention provides a speech recognition device that reduces the occurrence of recognition errors, that is, improves the recognition rate.

(d) Structure of the Invention The speech recognition device of the present invention recognizes unknown input speech using a time series of average value parameters indicating the average value of feature parameters of registered speech stored in advance and a time series of its standard deviation. Statistical processing is applied to the time series of feature parameters, and the feature parameters of this input voice are set to a binary state of "1", which indicates when the difference from the average value parameter is greater than a standard deviation times a constant, and when it is smaller. ”, “0” are converted into a time series of binary signals to determine the degree of similarity, and the registered voice with the highest degree of similarity is determined to be the input voice at that time.

(E) Embodiment FIG. 1 shows an embodiment of the speech recognition device of the present invention. In the figure, 1 is a microphone that converts audio into an electrical audio signal, and 2 is a parameter that extracts a frequency spectrum value, which is a feature parameter indicating the characteristics of the audio, from the audio signal obtained from the microphone 1. This is an extraction circuit that uses, for example, an 8-channel bandpass filter to generate a time series consisting of 8 samples each of frequency spectrum values f ₁ , f ₂ , f ₃ , ... f ₈ that are obtained by dividing the audio band (100 to 4000 Hz) into 8. A voice pattern expressed as is obtained.
In other words, when the filter number is n and the sample number is t, the characteristic parameter is expressed as fn(t), and the voice pattern F is becomes. 3 is a mode selection switch for switching between registration mode and recognition mode; when connected to the Q side, the mode is set to registration; when connected to the P side, the switch is set to recognition mode. 4 is a statistical processing circuit to which the voice pattern from the parameter extraction circuit 2 is input in the registration mode when the mode selection switch 3 is connected to the Q side; Based on a plurality of voice patterns, each feature parameter fn(t) forms a normal distribution as shown in FIG. 2, and an average value pattern _{1(1) 1(2)} ... consisting of average value parameters fn(t) is obtained. _{1(8) 2(1) 2(2)} … ₂₍₈₎ … … … _{8(1) 8(2)} … ₈₍₈₎ and the standard deviation pattern consisting of its standard deviation △ fn(t) △ Ff ₁₍₁₎ f ₁₍₂₎ … f ₁₍₈₎ f ₂₍₁₎ f ₂₍₂₎ … f ₂₍₈₎ … … … f ₈₍₁₎ f ₈₍₂₎ … f _{8( 8)} . Reference numeral 5 denotes a memory circuit, and an average value pattern memory section 51 stores average value patterns A, . and a mellow deviation memory unit 5 that stores standard deviation patterns △ A, △ B, △ C, △ D, △ E in association with this.
It consists of 2 and. On the other hand, 6 is a buffer memory for temporarily storing the voice pattern X for the unknown input voice X obtained from the parameter extraction circuit 2 during the recognition mode in which the mode selection switch 3 is connected to the P side. 7 is a comparison means, which compares the input audio pattern X of the buffer memory 6.
From each parameter xn(t), each average value pattern A of the average value pattern memory section 51 of the memory circuit 5, . . . respective parameters n(t), n
(t), . . . and the standard deviation patterns ΔA, ΔB, . . . of the standard deviation memory section 52 of the memory circuit 5.
A multiplier 72 that multiplies (t), ... by a constant K, for example, 1 or 2, and a subtracted value from the subtracter 71.
xn(t)-n(t), xn(t)-n(t), ... are the multiplied values K△ an(t), K from the multiplier 72, respectively.
△ Compare with bn(t), ...｜xn(t)-n(t)｜
Outputs “1” when KΔ an(t) |xn(t)−
The comparator 73 outputs "0" when n(t)|>KΔan(t). That is, when K=1, for example, |xn(t) for the registered voice of A
-n(t) | K△ an(t), xn(t) becomes an(t) based on the normal distribution as shown in Figure 2.
Since they are considered similar with a probability of 68.3%, "1" is given, and in the opposite case, "0" is given, and the unknown pattern Xx ₁₍₁₎ x ₁₍₂₎ … x ₁₍₈₎ x ₂₍₁₎ x ₂₍₂₎ … x ₂₍₈₎ … … … x ₈₍₁₎ x ₈₍₂₎ … x ₈₍₈₎ is for each registered voice “ The matrix pattern represented by the binary signal δ of "1" and "0" is converted into a matrix pattern Δ=δ ₁₁ δ ₁₂ ... δ ₁₈ δ ₂₁ δ ₂₂ ... δ ₂₈ ... ... δ ₈₁ δ ₈₂ ... δ ₈₈ .

8 is a recognition processing circuit that calculates the sum of its 16 constituent elements 〓 ⁱ 〓 ^j δij, that is, the number of existing "1"s, as a class degree, based on the matrix pattern △ of the binary signal obtained from the comparison means 7. Yes, A, B, C, D, E
For each registered voice, this level or higher is, for example,
11, 2, 8, 7, 3, it is determined that the input voice at this time was A.

In the registration mode connected to the mode selection switch 3Q, each of a plurality of predetermined voices is uttered several times, for example, three times each, and the average value pattern of each voice is stored in the memory circuit 5. , . . . and their standard deviation patterns △ A, △ B, . . . are stored. Then, in the recognition mode in which the mode selection switch 3 is connected to P, an unknown voice is input, and the voice pattern The patterns _ΔA , _ΔB , . At this time, K is used to determine the permissible variation in audio.
By setting the value of 0.5 to 2, the binary signal pattern optimally indicates the similarity between the unknown speech pattern and the registered speech pattern, and highly reliable pattern recognition can be performed based on this. .

In the following explanation, a speech pattern consisting of a time series of frequency spectrum values is used as a time series of feature parameters, but in addition to this, time series of various feature parameters such as the Percoll coefficient, which is an autocorrelation coefficient, can be used. is also possible.

(C) Invention and Effects As is clear from the above description, the speech recognition device of the present invention uses a time series of an average value parameter indicating the average value of feature parameters of registered speech stored in advance and a time series of its standard deviation. is used to perform statistical processing on the time series of the feature parameters of unknown input speech, converting it into a time series of binary signals of "1" and "0" to find the similarity, and calculate the similarity. Since it is determined that the registered voice is the one at the maximum, it is possible to remove variations in each feature parameter depending on the voice utterance situation, and it is possible to derive the optimal degree of similarity, greatly improving the recognition rate. can be expected.

[Brief explanation of drawings]

FIG. 1 is a block diagram showing an embodiment of the speech recognition device of the present invention, and FIG. 2 is a normal distribution diagram, where 1 is a microphone, 2 is a parameter extraction circuit, 4 is a statistical processing circuit, and 5 is a memory circuit. , 7 is a means of comparison,
Reference numeral 8 indicates a recognition processing circuit.

Claims (1)

[Claims] 1. Audio input means for converting audio into electrical signals;
a parameter extracting means for extracting characteristic parameters representing the characteristics of the voice from the electric signal of the input voice; A storage means for storing a time series of standard deviations corresponding to a time series of average value parameters, and a time series for each registered voice from the storage means in response to a time series of characteristic parameters of an unknown input voice obtained from the parameter extraction means. Read out the time series of the average value parameter and the time series of its standard deviation, compare it with the standard deviation obtained by multiplying the error value between the feature parameter of the input voice and the average value parameter by a constant, and based on the result of this comparison, determine that the above error value is a constant. “1” and “0” indicate the binary state when it is larger and smaller than the multiplied standard deviation.
a comparison means for converting the time series of the characteristic parameters of the unknown input voice into a time series of binary signals by outputting a binary signal; , a recognition processing means that calculates the degree of similarity of each registered voice to the unknown voice and determines that the registered voice with the highest degree of similarity is the currently input unknown voice.