JP3008404B2 - Voice recognition device - Google Patents

Description

Description: BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech recognition apparatus, and more particularly to a speech recognition apparatus intended to obtain phoneme feature information useful for phoneme recognition from temporal changes in the power of input speech. .

[Conventional technology]

Conventionally, in phoneme recognition and speech recognition, a temporal change in the power level of an input speech has been used as one of parameters for discriminating a voiced / unvoiced section or for judging the presence or absence of a phoneme.

[Problems to be solved by the invention]

The above determination has been made by comparing the power level of the input voice with a predetermined threshold. For example, when the power level of the input voice exceeds the threshold, it is determined that the voice or phoneme rises, and when the power level of the input voice falls below the threshold, it is determined that the voice or phoneme falls. In other words, in the conventional technology,
Since the power of the input voice has been used only for detecting the presence or absence of voice or phoneme, it has been desired to utilize the power more effectively.

Accordingly, an object of the present invention is to provide a speech recognition apparatus that can obtain useful information for phoneme recognition from a temporal change in power of an input speech.

[Means for solving the problem]

The present invention detects a peak of a temporal change in the power of the input voice, detects a peak edge in which the power of the input voice is smaller than the power at the peak of the input voice by a predetermined amount, and detects the peak edge based on the input voice. Generating a parameter representing an acoustic element, generating phoneme boundary information based on the parameter, referring to the information of the energy concentration section determined by the peak edge and the phoneme boundary information, and based on the parameter, It is configured to recognize phonemes.

[Action]

The power of the input voice is divided by a predetermined time unit, a peak is detected by calculating a temporal change in the power level within the predetermined time unit, and a point at which the power is smaller than the peak by a predetermined amount, That is, a peak edge is detected. Based on the peak edge, phoneme feature information useful for phoneme recognition,
For example, an energy concentration section, a burst section, and the like can be obtained, and the accuracy of phoneme recognition can be improved.

〔Example〕

An embodiment of the present invention will be described below with reference to FIGS.

 FIG. 1 shows an example of a speech recognition apparatus according to the present invention.

An audio signal from the microphone 1 is supplied to the A / D conversion circuit 4 via the amplifier 2 and the low-pass filter 3. The above-mentioned audio signal is, for example, 1
It is converted to a 12-bit digital audio signal at a sampling frequency of 2.5 KHz. This digital audio signal is supplied to the acoustic analysis means 5.

The acoustic analysis unit 5 includes a transient detection parameter generation unit 51 having a band-pass filter bank, a logarithmic power detection unit 52 for detecting audio power, a zero-cross rate calculation unit 53, and a primary for checking the correlation between adjacent samples. And a means 55 for calculating the slope of the power spectrum, and a means 56 for detecting the fundamental period of the voice.

The transient detection parameter is for detecting the transient and steadiness of the input voice.
The amount of change in the audio spectrum is defined as the sum of the variances in the block in the time direction of each channel (frequency). That is, the gain of the audio spectrum Si (n) is normalized by an average value Savg (n) shown below in the frequency direction.

Here, i indicates the channel number, and q indicates the number of channels (the number of bandpass filters). The information of each channel of the q channel is sampled in the time direction,
A block of q-channel information at the same time is called a frame, and n indicates the number of a frame used for recognition.

Speech spectrum i (n) after gain normalization
Is And the transient detection parameter T (n) is the sum (2M + 1) of M frames before and after that frame [n−M, n
+ M] is defined as the sum of variances in the time direction of each channel in the block.

here, , Which is the average value in the time direction within the block of each channel.

In practice, the change near the [n−M, n + M] block center is easy to pick up sound fluctuations or noises, so that it should be removed from the calculation of the transient detection parameter T (n). It is transformed as follows.

Then, in the equation (5), a = 1, M = 28, m =
3. Assuming q = 32, the transient detection parameter T (n) is obtained. For example, in the case of the input voice “asa”, a transient detection parameter T (n) as shown in FIG. 2A is obtained.

Other parameters, such as the log power shown in FIG. 2B, the zero cross rate shown in FIG.
, The slope of the power spectrum shown in FIG. 2E, and the calculation of parameters such as the fundamental period shown in FIG. 2F, at a certain point in time, similarly to the transient detection parameter T (n). Consider a window having a time width of M frames before and after (frame), and sequentially move the window in the time direction by one sample point,
It is obtained by performing calculations in each window.
FIG. 2G shows an example of a speech waveform of the input speech “asa” and phoneme boundary candidates.

Each parameter obtained from the acoustic analysis means 5 is supplied to the phoneme recognition means 8 as a parameter for recognition processing. Each parameter output from the means 51 to 55 is supplied to the feature point extracting means 61 of the first segmentation means 6 as a segmentation parameter. And means
The output from the bandpass filter bank at 51 is supplied to the peaking processing circuit 11.

The peaking processing circuit 11, as shown in FIG.
It comprises a power calculation circuit 14, a peak detection circuit 12, and a peak edge detection circuit 13.

When the digital output of the power of the input sound supplied from the band-pass filter bank 51 is sequentially supplied to the power calculation circuit 14, an average value of the digital output is obtained for each frame. Then, when this average value is supplied to the peak detection circuit 12 in time series, a peak of the power level is detected in the time axis direction. When the obtained temporal change of the power level is, for example, as shown in FIG.
Peaks PP1 and PP2 are detected.

When the peaks PP1 and PP2 and the temporal change of the power level are supplied to the next-stage peak edge detection circuit 13,
On both sides of the PP, the point where the power level has decreased by a predetermined level, for example, the point where the power level has decreased by 3 dB is the peak edge PE1, PE
2. Detected as PE3, phoneme recognition means 8 via terminal 15
Supplied to

By setting the peak edge PE, phonemic feature information can be obtained corresponding to the power level waveform of the input voice.
The above-described detection of the peak PP and the peak edge PE is performed in the fourth step.
As shown, the digital output is converted to a predetermined time unit T
And is performed within each time unit T.

For example, in the waveform shown in FIG. 5, points that are 3 dB lower than the peak PP6 are detected as the peak edges PE61 and PE62. Thus, the energy concentration section TE is defined between the two peak edges PE61 to PE62. This is a section in which the energy of the power of the input voice is concentrated at a high level, and is determined to be a phoneme section in which a phoneme occurs. In this case, the area outside the energy concentration section TE can be determined to be a silent section or a pause section. Thus, information that can determine the presence or absence of a phoneme is obtained.

Next, in the waveform shown in FIG. 6, peak edges PE71 and PE72 are detected with respect to the peak PP7. A section between the peak edges PE71 to PE727 is detected as a section shorter than a predetermined section [time length] TC. As described above, when the peak PP and the peak edge PE are detected in a time shorter than the predetermined time length TC, the interval between the peak edges PE71 to PE727 is determined to be a burst section TPE. From now on, for example,
The information which can judge the plosive sound and the affricate is obtained.

In this example, an example in which a point 3 dB lower than the peak PP is detected as the peak edge PE is shown.However, the present invention is not limited to this example. It may be detected as an edge PE.
Further, in this example, an example in which the band-pass filter bank 51 is used is described, but the present invention is not limited to this example.
FFT may be used.

As described above, the peak edge PE whose power is 3 dB smaller than the peak PP obtained from the temporal change of the power level of the input voice is detected, and based on the peak edge PE, the phonological features such as the energy concentration section TE and the burst section TPE are detected. Information can be obtained. Information on peak PP and peak edge PE,
Information such as the energy concentration section TE and the burst section TPE is supplied to the phoneme recognition means 8 as phoneme feature information.
Thereby, the accuracy of phoneme recognition can be improved.

The first segmentation means 6 extracts general feature points in order to obtain phoneme boundary candidates from the segmentation parameters. In this example, the following seven types are used as feature points.

Rise point-point where the area changes from a flat part in the direction of increase Fall point-point where the area changes in the direction of decrease and then becomes flat Increase point of change-point where the rate of change changes Decrease point-point where the rate of decrease changes Peak point-peak position Positive zero cross point-point crossing zero level in increasing direction Negative zero cross point-point crossing zero level in decreasing direction Feature point extraction means 61 has a feature from feature point information storage means 62 A feature point is extracted for each parameter with reference to the point information. In each of the parameters in FIGS. 2A to 2E, the position indicated by a vertical line in the time axis direction is the position of each feature point. Each parameter obtained from the first segmentation means 6 and provided with a feature point is supplied to a feature point integration processing means 71 of the second segmentation means 7.

The second segmentation means 7 comprises a feature point integration processing means 71, a phoneme boundary feature detection means 72, a feature point integration information storage means 73, and a phoneme boundary feature information storage means 74.

Since the feature points obtained by the first segmentation means 6 have positional deviation, undetection, and the like for each parameter, the feature point integration processing means 71 refers to the feature point integrated information from the feature point integrated information storage means 73, and refers to each parameter. And the phonetic boundary candidates are determined. The feature point integrated information is information on which parameter feature point has priority.

The phoneme boundary feature detecting means 72 obtains a phoneme boundary feature of each phoneme boundary candidate. In this example, eight types of phoneme boundary features are used.

Rise from silence (SR) Consonant → vowel (CV) Consonant → consonant (CC) Vowel → vowel (VV) Fall to vowel (VF) Vowel → consonant (VC) Fall to silence (FS) Voice → silent (SS) The phoneme boundary feature information storage means 74 stores these eight types of phoneme boundary feature information. The phoneme boundary feature detecting means 72 refers to the information from the phoneme boundary feature information storage means 74 and detects phoneme boundary features of each phoneme boundary candidate. As a result, as shown in FIG. 2G, the phoneme boundary feature is shown near the vertical line of the phoneme boundary candidate.

From the second segmentation means 7, phoneme boundary candidate information and its phoneme boundary feature information are obtained as phoneme section information, and this phoneme section information is supplied to the phoneme recognition means 8.

The phoneme recognition means 8 uses the parameters supplied from the acoustic analysis means 5 as parameters for recognition processing, and performs phoneme recognition while referring to phoneme feature information from the peaking processing circuit 11 and phoneme section information from the second segmentation means 7. Execute. Then, the recognized phoneme symbols are obtained from the phoneme recognition means 8 and supplied to the subsequent continuous speech and large vocabulary speech recognition means.

In this embodiment, an example in which hardware is used has been described. However, the first and second segmentation units 6 and 7, the calculation unit of the acoustic analysis unit 5, the peaking processing circuit 11, the phoneme recognition unit 8, and the like are implemented by a computer. It may be realized.

〔The invention's effect〕

According to the present invention, it is possible to newly extract a peak and a peak edge in the temporal change of the power of the input voice, and to obtain phonemic feature information of the input voice, such as an energy concentration section and a burst section, from the peak edge. effective. In addition, since phoneme feature information can be used for phoneme recognition, the power of the input speech can contribute not only to discrimination between voiced and unvoiced, detection of the presence or absence of phonemes, but also to the improvement of phoneme recognition accuracy. There is.

[Brief description of the drawings]

FIG. 1 is a block diagram showing an embodiment of the present invention, FIG. 2 is a waveform diagram for explaining each embodiment, FIG. 3 is a block diagram showing a peaking processing circuit, and FIGS. It is explanatory drawing for demonstrating an Example, respectively. Explanation of main symbols in the drawings 8: phoneme recognition means, 11: peaking processing circuit, 12: peak detection circuit, 13: peak edge detection circuit, 14: power calculation circuit.

──────────────────────────────────────────────────続 き Continuation of front page (56) References JP-A-62-141595 (JP, A) JP-A-59-111700 (JP, A) JP-A-60-39691 (JP, A) JP-A 63-141 168700 (JP, A) JP-A-59-123893 (JP, A) JP-A-61-185799 (JP, A) JP-A-58-130395 (JP, A) JP-A-2-232699 (JP, A) JP-A-1-170998 (JP, A) JP-B-62-58518 (JP, B2) JP-B-5-88840 (JP, B2) (58) Fields investigated (Int. Cl. ⁷ , DB name) G10L 3/00-9/20 JICST file (JOIS)

Claims (1)

(57) [Claims] 1. A peak detecting means for detecting a peak of a temporal change in power of an input voice, and a peak edge detecting a peak edge in which the power of the input voice is smaller than the power of the input voice by a predetermined amount. Detecting means; sound analyzing means for generating a parameter representing an acoustic element based on the input voice; phoneme boundary information generating means for generating phoneme boundary information based on the parameter; energy concentration determined by the peak edge A speech recognition device, comprising: phoneme recognition means for recognizing a phoneme of the input speech based on the parameter with reference to section information and the phoneme boundary information.