JPH0520760B2 - - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to a voice pattern creation device that creates a voice pattern representing the characteristics of a voice based on a voice signal.

This type of speech pattern generation device is installed in a speech recognition device, and the input speech is pattern-recognized using the speech pattern obtained from the device. Therefore, this sound pattern is
It must accurately reflect the characteristics of the voice, and for this reason, research and development of this type of voice pattern creation device is being actively conducted.

FIG. 1 shows a conventional voice pattern creation device.
In the figure, 1 is an input terminal into which the audio signal is input, and 2 is a group of bandpass filters for extracting the frequency spectrum as a characteristic parameter of the audio.
It consists of several bandpass filters. Reference numeral 3 denotes a sampling circuit for sampling eight frequency spectrum values obtained from the bandpass filter group 2, and the sampling period thereof is set to 6 msec by the clock generator 4.
Reference numeral 5 denotes a parameter editing circuit for editing the frequency spectrum values from the sampling circuit 3 to obtain an audio pattern, and a buffer memory 6 for temporarily storing eight frequency spectrum value sequences obtained in time series from the sampling circuit 3. and a readout circuit 7 for selecting eight frequency spectrum values of 16 samples at equal sample intervals from the eight frequency spectrum value sequences stored in the buffer memory 6.
It consists of and. Reference numeral 8 denotes an audio pattern memory in which frequency spectrum values read out by the reading circuit 7 from the buffer memory 6 of the parameter editing circuit 5 are stored as an audio pattern of 16 rows and 8 columns.

According to the voice pattern creation device having such a configuration, the voice pattern of 16 rows and 8 columns stored in the voice pattern memory becomes a sequence of frequency spectrum values at equal time intervals with respect to the time axis.

However, when attempting to recognize speech using the above-mentioned speech patterns, for example, taking numeric speech as an example, 1 (ichi), 7 (hichi), 8
(hachi), and 2 (ni) and 4 (shi), and 9 (kyu)
In many cases, words can be distinguished only by the difference in the sound at the beginning of the word, such as ``10 (jyu)'' and 10 (jyu). Therefore, there was an inconvenience in erroneous recognition between these recognized words. Such inconveniences tend to occur when attempting to recognize not only numeric sounds as described above but also word sounds in the same field, such as alphabets and color names, and this has led to a drop in recognition rate.

The present invention has been made for the purpose of solving the above-mentioned disadvantages, and will be described in detail below.

The voice pattern creation device of the present invention shown in FIG. 2 is shown. In the same figure, 1, 2, 3, 6, 7, and 8 indicate input terminals, a group of bandpass filters, a sampling circuit, a buffer memory, a readout circuit, and an audio pattern memory, similar to the conventional device shown in FIG. The feature of the present invention lies in the parameter editing circuit 5'. That is, in this editing circuit 5', 9 is the first register, in which eight frequency spectrum values of the voice extracted by the bandpass filter group 2 and sampled by the sampling circuit 3 are stored for one sample. It will be done. A second register 10 stores eight frequency spectrum values for one sample transferred from the first register. Reference numeral 11 denotes a rising edge detection circuit, which calculates the sum of the eight frequency spectrum values of the second register 10 from the sum of the eight frequency spectrum values of the first register 9 [corresponding to the sound pressure (energy) of the audio signal]. An operation to subtract is performed,
When the result of this calculation is positive, that is, when the sum of the frequency spectrum values of the voice is in an increasing state, a rising edge detection signal is output. Reference numeral 12 denotes a variable clock generator, which normally generates a clock pulse with a period of 6 msec, but when it receives the rising edge detection signal from the rising edge detection circuit 11, it generates a clock pulse with a period shorter than the normal period, in this case half of 6 msec. Generates a clock with a period of 3 msec. The output from the variable clock generator 12 is sent to the sampling circuit 3 and the first register 9, and the period of this clock becomes the sampling period of the sampling circuit 3 and the rewriting period of the first register 9. becomes. 13 is a parameter change rate detection circuit, which subtracts each of the eight frequency spectrum values of the second register 10 from each of the eight frequency spectra of the first register 9, and calculates the sum of these eight subtraction results, that is, An operation is performed to calculate the amount of variation in the frequency spectrum value. 14 is a comparison circuit which subtracts the threshold value set by the threshold value setting circuit 15 from the calculation result obtained from the parameter change rate detection circuit 13, and when this subtraction result is positive, that is, the variation of the frequency spectrum value is When a predetermined threshold is exceeded,
A spectral variation detection signal is sent to the second register 10. The second register 10 receives this spectrum fluctuation detection signal and writes the eight frequency spectrum values of the second register 10 into the buffer memory 6, and subsequently the contents of this register 10 are transferred to the first register 9. It is rewritten into eight frequency spectrum values. On the contrary, when there is no spectrum fluctuation detection signal, the contents of the second register 10 are held.

Next, the operation of the voice pattern creation device of the present invention will be explained. FIG. 3 shows the time change curve of the sound pressure (energy) of the sound signal when the number sound 42 (Shijuni) is uttered and its sampling period t. When an audio signal is input to input terminal 1, bandpass filter group 2 extracts eight frequency spectrum values. These eight frequency spectrum values are sampled by the sampling circuit 3 and sequentially stored one sample each in the first register 9 and the second register 10, and the sampling period t is as shown in FIG. 3 msec (=
t ₁ ), and 6 msec (=
_t2 ). That is, in the rise detection circuit 11, the rise areas a, b, and c of the sound pressure of the audio signal are detected based on the frequency spectrum values of both the first register 9 and the second register 10, and these areas, The rising edge detection signal output at the time , is sent to the variable clock generator 12 . The variable clock generator 12 normally supplies a clock with a period of 6 msec (=t ₂ ) to the sampling circuit 3, but when it receives the rise detection signal, it supplies a clock with a period of 3 msec (=t 2 ).
It supplies a clock with a period of t ₁ ). Also, the clock from this variable clock generator 12 is
The first register 9 is also supplied to the first register 9 when the sampling circuit 3 is connected to the normal 6 m
While sampling is performed at a sec period, the contents of this register 9 are rewritten to the frequency spectrum value obtained from the sampling circuit at a 6 msec period synchronized with this, while when the sampling circuit 3 is sampling at a 3 msec period. teeth,
The contents of this register 9 are also rewritten every 3 msec. On the other hand, the parameter change rate detection circuit 13 of the parameter editing circuit 5' calculates the variation of the frequency spectrum value based on the frequency spectrum values of both the first register 9 and the second register 10, and A spectral variation detection signal obtained from the comparator circuit 14 is sent to the second register 10 when the spectral variation detection signal is larger than the threshold value of the threshold value setting circuit 15. Upon receiving this spectrum fluctuation detection signal, the second register 10 writes the eight frequency spectrum values of the second register 10 into the buffer memory 6, and subsequently the contents of this second register 10 are transferred to the first register. It is rewritten into 8 frequency spectrum values of 9. Conversely, when there is no spectrum fluctuation detection signal, the second register 1
The contents of 0 are retained. The threshold value of the threshold value setting circuit 15 may be set to a constant value, but as shown by the dotted line in FIG. If the threshold value setting circuit 15 sets a threshold value proportional to the frequency, the optimum threshold value will be obtained in accordance with the rate of change in the frequency spectrum value.

In this way, the data 8 are sequentially written into the buffer memory 6.
The two frequency spectrum value sequences are sampled at 3 msec intervals when the sound pressure of the audio signal is in a rising state, and at 6 msec intervals when it is not in a rising state, and the number of samples in the rising part, which is important for voice identification. occupies a large proportion of the total number of samples, and in addition, among these samples, unnecessary samples whose frequency spectrum values have been found to have almost no fluctuation compared to the immediately preceding sample have been removed. Become. Next, the eight frequency spectrum value strings in the buffer memory 6 are processed by the readout circuit 7, which reads out 16 equally spaced samples from these samples and stores them in the audio pattern memory 8 so that they are arranged in 16 rows and 8 columns. The voice pattern is normalized to .

In the above explanation, an embodiment has been described in which the rising edge detection circuit 11 detects the increasing state of the sound pressure of the audio signal based on the frequency spectrum value which is a characteristic parameter of the audio. The rising edge detection circuit 11 may be configured by a differentiating circuit that extracts the signal and differentiates this signal.

As is clear from the above description, the audio pattern creation device of the present invention includes a rising edge detection circuit that detects the rising state of the sound pressure of the audio signal, and this circuit detects when the sound pressure of the audio signal at this time rises. When detecting that the voice is in the rising state, the sampling period of the voice characteristic parameters is set shorter than when detecting that the voice is not in the rising state. The number of samples in the rising region occupies a higher proportion of the total number of samples, and it is possible to obtain a voice pattern that more accurately reflects the characteristics of the voice. Furthermore, it is equipped with a parameter change rate detection circuit that detects the rate of change of the characteristic parameter of the voice, and creates a voice pattern using the characteristic parameter when the parameter change rate obtained by this circuit is larger than a predetermined value. Therefore, it is possible to exclude unnecessary samples that have almost no variation compared to the sample immediately before the feature parameter, and it is possible to obtain a speech pattern consisting only of important samples that represent the state of variation in the feature parameter.

[Brief explanation of the drawing]

FIG. 1 is a block diagram of a conventional voice pattern creation device, FIG. 2 is a block diagram of a voice pattern creation device of the present invention, and FIG. 3 is a signal waveform diagram of voice sound pressure showing the sampling period according to the device of the present invention. and
2 is a group of pass filters, 3 is a sampling circuit, 5 and 5' are parameter editing circuits, 6 is a buffer memory, 8 is an audio pattern memory, 9 is a first register, 10 is a second register, 11 is a rising edge detection circuit , 12 is a variable clock generator, 13 is a parameter change rate detection circuit, 14 is a comparator, and 15 is a threshold value setting circuit, respectively.

Claims (1)

[Scope of Claims] 1. A feature extraction circuit that extracts voice feature parameters from an audio signal, a sampling circuit that samples the feature parameters obtained from the feature extraction circuit, and a feature parameter that samples the feature parameters sampled by the sampling circuit. The rise detection circuit comprises a parameter editing circuit that edits in time series to obtain an audio pattern, and a rise detection circuit that detects a rise area of a change in sound pressure of an audio signal based on the audio signal or the characteristic parameter. When it is detected that the sound pressure change of the audio signal is in a rising state, the sampling period of the sampling circuit is set shorter than when it is detected that the sound pressure change of the audio signal is not in a rising state. A voice pattern creation device. 2. The parameter editing circuit includes a parameter change rate detection circuit that detects the rate of change of the characteristic parameters sequentially obtained from the sampling circuit, and the parameter change rate detection circuit detects the rate of change of the characteristic parameters obtained from the parameter change rate detection circuit. 2. The voice pattern creation device according to claim 1, wherein the voice pattern is obtained by editing the feature parameters when the value is larger than the value.