JPH0341840B2 - - Google Patents

Description

[Detailed Description of the Invention] (a) Field of Industrial Application The present invention relates to a speech spectral parameter extraction device for extracting spectral parameters indicating the frequency characteristics of speech handled in speech recognition processing and analysis processing for speech synthesis. .

(b) Prior art In this type of speech spectrum parameter extraction device, it is considered important to extract speech spectrum parameters that are not affected by noise and faithfully represent the characteristics of the speech. There are background noises that are input at the same time when inputting a signal, that is, external noises, and internal noises generated by the device itself.
In particular, it has been difficult to completely eliminate internal noise.

(c) Object of the Invention The present invention has been made in view of the above-mentioned points, and provides a speech spectrum parameter extraction device capable of extracting accurate speech spectrum parameters free of noise components.

(d) Structure of the Invention The audio spectrum parameter extraction device of the present invention includes a spectrum analysis circuit for extracting spectrum parameters indicating frequency characteristics based on an electrical signal from a microphone, and a spectrum analysis circuit for extracting spectrum parameters indicating the frequency characteristics of the electrical signal from the microphone. a first subtracter that subtracts a value of a spectral parameter in a second memory from a spectral parameter in the first memory; and a subtracted value from the first subtractor. a third memory; an adder for adding the spectral parameter value in the second memory to the subtracted value in the third memory; and subtracting the added value from the adder from the spectral parameter value in the first memory; and a second subtractor for inputting a voice input by separately storing the internal noise and external noise spectral parameters in the second and third memories during the inhibit period when there is no voice input. During the permissible period, the subtracted value obtained from the second subtracter is extracted as a voice spectrum parameter.

(E) Embodiment FIG. 1 shows an embodiment of the audio spectrum parameter extraction device of the present invention. In the figure, 1 is a microphone for converting audio into electrical signals, 2
is a spectrum analysis circuit that extracts spectral parameters indicating the frequency characteristics based on the electrical signal obtained from the microphone 1 via the open/close switch SW1. The spectrum analysis circuit 2 includes eight parallel-connected bandpass filters BPF1,...BPF8 that separate and filter an audio electrical signal into, for example, eight different frequency bands, and shortens the output signal value from each of these bandpass filters. Time (about 10m
sec) eight averagers A1, ..., A8 for averaging,
A multiplexer MP that samples the averaged signals obtained from these averagers A1, ..., A8 at a period of approximately 10 msec and obtains eight spectral parameters A/D converted at each sampling period;
Consisting of More specifically, each bandpass filter BPF1, ..., BPF8 of this analysis circuit 2 employs a switch capacitance filter.
Depending on the clock frequency obtained from the clock generator CLK, each bandpass filter BPF1,
..., the filtering band of each BPF8 is set. 3 and 4 are the first ones connected from the spectrum analysis circuit 2 through the changeover switch SW2.
and a second memory, in which the sample values of the spectral parameters sequentially obtained from the spectral analysis circuit 2 are transferred to the changeover switch SW2.
By connecting the switch SW2 to the a side, the data is stored in the first memory 3, and by connecting the changeover switch SW2 to the b side, the data is stored in the second memory 4. 5 is a first subtractor that subtracts the value of the spectral parameter in the second memory 4 from the value of the spectral parameter in the first memory 3; 6 is a third memory that stores the value subtracted from the first subtractor 5; be. 7 is the subtracted value of the third memory 6 from the second memory 4.
An adder 8 adds the sample values of the spectral parameters of the adder 7, and an adder 8 adds the added value from the adder 7.
a second subtractor for subtracting from the sample value of the spectral parameter in the memory of . Reference numeral 9 denotes a storage means for storing the subtracted value obtained from the second subtractor 8 as a voice spectrum parameter.

Furthermore, although not shown, the first switch described above
Opening/closing and switching operations of SW1 and second switch SW2 can be realized by general control means, for example, microcomputer control. An example of such switch control is shown in FIG.

Next, we will take as an example the case where the device according to the embodiment of the present invention sequentially extracts the spectral parameters of each of the word sounds sequentially input to the microphone 1, and the timing diagram in FIG. 2 and the waveform diagram in FIG. 3 will be explained. The operation will be explained using . In this case, the spectral parameters of the voice obtained by such a device are stored in the storage means 9, and then transferred from the storage means 9 to a data processing circuit (not shown) such as a pattern matching circuit for voice recognition. It will be transmitted to
On the contrary, during the reading period from the storage means 9, writing is no longer possible, so that it is impossible to extract the audio spectrum parameters. Therefore, for any reason including this condition, the second
As shown in the figure, a prohibition period T1 for prohibiting writing of audio spectral parameters to the storage means 9;
T3, . . . are provided, and audio spectral parameters can be extracted in permissible periods T2, T4 between the prohibited periods T1, T3, . . . and the next prohibited period T3, .

As mentioned above, during the prohibition period T1 during which voice input to the microphone 1 is prohibited, the open/close switch SW1 and the changeover switch SW2 are operated as shown in FIG.
As shown in the figure, opening/closing and switching are repeated in synchronization. As a result, when the open/close switch SW1 is opened and the changeover switch SW2 is connected to the b side, the spectrum analysis circuit 2
will analyze the no signal, but each switch has a bandpass filter consisting of a capacitance filter.
Because BPF1, ..., BPF8 are driven by receiving each clock from the clock generator CLK, and furthermore, because the reference output voltage varies depending on the temperature characteristics of these filters BPF1, ..., BPF8, each band The outputs from the pass filters BPF1, . . . , BPF8 are not zero, but some noise components are output. Also, each bandpass filter
The output values of noise components from BPF1, ..., BPF8 are further averaged for a short time by each averager A1, ..., A8.
Noise components due to temperature characteristics etc. will be superimposed. The output from multiplexer MP is thus
In other words, the output from the spectrum analysis circuit 2 in a no-signal state is as shown in FIG. 3, which is generated by the analysis circuit 2 itself.
As shown by Ni, the internal noise spectrum parameter has a relatively large variation, and the contents of the second memory 4 are rewritten with the value of one sample of this internal noise spectrum parameter. Therefore, the latest internal noise spectrum parameters are always stored in the second memory 4 during the prohibited period T1 whose time length is not necessarily constant as shown in FIG. On the contrary, the above opening/closing switch SW1
is closed and switching switch SW2 is closed.
When connected to the side, the microphone 1 picks up background noise, ie external noise, and the electrical signal of this external noise is analyzed by the spectrum analysis circuit 2.
However, the spectral parameters obtained as a result of this analysis are an external noise spectral parameter with a relatively small variation as shown in No. of FIG. The content of the first memory 3 is updated with the value of one sample of this parameter as shown in FIG. Therefore, the first
The subtracter 5 sequentially outputs the latest sample values of external noise spectrum parameters obtained by subtracting the sample values of the internal noise spectrum parameters of the second memory from the sample values of the synthetic noise spectrum parameters of the first memory. It happens.

During this prohibition period T1, when the signal C indicating that the external noise spectrum parameter is to be taken in, that is, the external environment of the device has changed and the background noise seems to have changed, the user presses, for example, the key input section. When there is an external noise acquisition command obtained by operating C, the third memory 6 receives this signal C.
As shown in FIG. 2, the first subtractor 5
The external noise spectrum parameters obtained at this time are stored and retained.

Thus, as shown in FIG. 3, the third memory 6 stores the sample value of the spectral parameter at point X of the external noise No. at time t when the above-mentioned captured signal C is received during the prohibition period T1. This value is stored in the external noise that can be considered steady state for a relatively long time.
It will not be updated until it is determined that No has changed. On the other hand, at the end of the prohibition period T1, that is, at the start of the following permissible period T2, the second memory 4 contains internal noise Ni that can be considered to be in a steady state for a relatively short period of time, as shown in FIG. The sample value of the spectrum parameter at point Y at this time is stored, and this value is updated to the value at point Z at the end of the next inhibited period _T3 .

During the permissible period T2 during which audio input to the microphone 1 is permitted, the open/close switch SW1 and the changeover switch SW2 are closed and connected to the a side as shown in FIG. Since the voice is input, the spectrum analysis circuit 2 outputs the voice S, external noise No., as shown in FIG.
Sample values of the synthesized spectral parameters consisting of the internal noise Ni and the internal noise Ni are sequentially output, and each sample value is sequentially stored in the first memory 3. on the other hand,
The adder 7 outputs a composite noise obtained by adding the sample value of the internal noise spectrum parameter at the latest Y point in the second memory 4 and the sample value of the external noise spectrum parameter at the X point in the third memory 6. Sample values of spectral parameters are obtained. Therefore, the second subtracter 8 outputs a true speech spectrum parameter obtained by sequentially subtracting one sample of the synthesized noise spectrum parameter from the adder 7 from each sample value of the synthesized spectrum parameter in the first memory 3. Each sample value is obtained and sequentially written into the storage means 9.

When the input of one word's worth of speech is thus completed, and this completion is determined by, for example, detecting the state in which the sample values of each parameter in the storage means 9 become zero, the permissible period T2 Then, the process moves to a prohibition period T3 for reading out the voice spectrum parameters in the storage means 9, for example, to a pattern matching circuit for voice recognition. Then, during this new inhibition period T3, the internal noise spectrum parameter in the second memory 4 is updated to the sample value at the new point Z. Therefore, in the next tolerance period T4 as well, by using the internal noise spectral parameter corresponding to the change in the internal noise Ni, the spectral parameter of speech without noise components can be obtained as in the case of the tolerance period T1.

In the embodiment described above, each memory 3,
4 and 6 are used to store one sample value of the parameter, but they may be used to store a plurality of consecutive samples. In this case, the second memory 4 and the third memory 6 are used. By adding a function that can calculate the average value of an appropriate number of consecutive sample values, it is possible to obtain the average internal noise spectrum parameter and external noise spectrum parameter at that time, making it possible to extract more accurate speech spectrum parameters. becomes.

(f) Effects of the Invention As is clear from the above description, the audio spectrum parameter extraction device of the present invention stores the latest internal noise spectrum parameters in the second memory at the end of the prohibition period without audio input; Separately, external noise spectrum parameters are stored in advance in a third memory within the prohibited period, and in the next permissible period when audio input is present, a composite noise spectrum parameter is obtained by adding the values of the parameters in both memories. , the subtracted value obtained by subtracting the value of this parameter from the value of the synthesized speech spectrum parameter consisting of the current speech, internal noise, and external noise in the first memory is extracted as the speech spectrum parameter, so it takes a relatively long time. For external noise that can be considered to be in a steady state, the operator can select a time when there is no voice input and obtain the external noise spectral parameters accordingly, while for internal noise that can be considered to be in a steady state for a relatively short period of time. In contrast, it is possible to obtain the internal noise spectrum parameters immediately before the voice input. Therefore, by performing correction using these external noise spectrum parameters and internal noise spectrum parameters, it becomes possible to extract accurate speech spectrum parameters free of noise components.

[Brief explanation of the drawing]

FIG. 1 is a block diagram of an embodiment of the audio spectrum parameter extraction device of the present invention, FIG. 2 is a timing diagram, and FIG. 3 is a signal waveform diagram, in which 1 is a microphone, 2 is a spectrum analysis circuit, 3, 4 and 6 are memories, 5 and 8 are subtracters, 7 is an adder, and 9 is a storage means, respectively.

Claims (1)

[Scope of Claims] 1. A microphone for converting sound into an electrical signal, a spectrum analysis circuit for extracting a spectrum parameter indicating the frequency characteristics based on the electrical signal from the microphone, and a spectrum analysis circuit for extracting a spectrum parameter indicating the frequency characteristic of the electrical signal from the microphone. first and second memories for storing spectrum parameters to be analyzed; an on/off switch for connecting or disconnecting the microphone and the spectrum analysis circuit;
or a changeover switch that switches the connection to the second memory; a first subtracter that subtracts the value of the spectral parameter of the second memory from the value of the spectral parameter of the first memory; and a subtraction from the subtracter. a third memory for storing a value; an adder for adding the value of the spectral parameter in the second memory to the subtracted value in the third memory; and an adder for adding the value of the spectral parameter in the second memory to the subtracted value in the third memory; a second subtracter for subtracting from the value of the parameter; a storage means for storing the subtracted value obtained from the second subtracter as a voice spectrum parameter; and a control means for controlling switch operations of the opening/closing switch and the changeover switch. and the opening/closing operation is performed by the control means during a prohibition period in which writing of the audio spectral parameters to the storage means is prohibited, which includes at least a period for reading and outputting the audio spectral parameters from the storage means. The switch and the changeover switch are opened/closed and switched at a synchronized period, and the contents of the second memory are rewritten to the latest internal noise spectrum parameters generated in the spectrum analysis circuit itself, and the update process of the first memory is performed. Alternately performing update processing to rewrite the contents to the latest synthetic noise spectrum parameters of the internal noise spectrum parameters generated by the spectrum analysis circuit itself and the external noise spectrum parameters obtained by spectrally analyzing the background noise input to the microphone, and further During this prohibition period, when a signal indicating that external noise is to be taken in is received, the value of the internal noise spectrum parameter in the second memory is calculated from the value of the synthesized noise spectrum parameter in the first memory obtained from the first subtracter. The reduced external noise spectral parameter is stored in a third memory, and the control means closes the on/off switch during a permissible period in which writing of the audio spectral parameter to the storage means is permitted. The changeover switch is connected to a first memory and combines the internal spectral parameters immediately before this period stored in the second memory obtained from the adder with the external noise spectral parameters in the third memory. The value of the noise spectrum parameter is subtracted by the second subtractor from the value of the synthesized spectrum parameter including the voice, external noise, and internal noise sequentially obtained from the first memory, and this subtracted value is used as the voice spectrum parameter. An audio spectrum parameter extraction device characterized in that the audio spectrum parameters are stored in the storage means as follows.