JPH0376475B2 - - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION [Technical Field of the Invention] The present invention relates to a speech recognition device that can effectively recognize discretely uttered word sounds.

[Technical background of the invention and its problems]

When recognizing discretely uttered word sounds,
It is very important as preprocessing to detect the speech section in which the word speech exists from the input speech signal sequence. Conventionally, however, the above-mentioned voice section has generally been detected using energy changes in the input voice. This method of detecting speech intervals has the advantage of being very simple, but it is difficult to remove the noise when the speech input environment is such that a lot of noise is added to the uttered word sound. It was extremely difficult to stably recognize the above word sounds. This is because if noise is added near the area where the speech to be recognized exists, the above-mentioned energy alone cannot distinguish the word speech from the noise, and the noise is considered part of the speech and incorporated into the recognition process. This is to put it away. Various processing methods, such as endpoint-free DP matching methods, have been considered to overcome these problems, but these methods have had problems such as an enormous amount of recognition processing, making them impractical.

[Purpose of the invention]

The present invention has been made in consideration of these circumstances, and its purpose is to provide a simple method that can stably recognize discretely uttered word sounds without being disturbed by noise. The objective is to provide a highly practical speech recognition device.

[Summary of the invention]

The present invention acoustically analyzes an input speech signal, extracts, for example, its energy change and phonological features from the acoustic analysis results, and searches for starting and ending candidate points of the speech signal according to these features. Detect it. Then, the features necessary for the recognition processing of the speech signal in the plurality of speech candidate sections obtained from all possible combinations of these start end candidate points and end end candidate points are resampled and extracted, and these resampled and extracted features are Speech recognition is performed according to the characteristics.

〔Effect of the invention〕

Therefore, according to the present invention, by performing speech recognition processing using features resampled in each of a plurality of speech candidate sections, and extracting the most reliable recognition result among them, noise components are removed. It becomes possible to remove the information of the speech candidate section detected including the speech candidate section, and easily perform stable speech recognition there.

[Embodiments of the invention]

Hereinafter, an embodiment of the present invention will be described with reference to the drawings.

FIG. 1 is a schematic configuration diagram of an embodiment device. Discretely generated word sounds are acoustoelectrically converted in the audio input section 1, amplified to an appropriate signal level, and then A/D converted and input. The acoustic analysis processing unit 2 spectrally decomposes the audio signal inputted through the audio input unit 1 into each predetermined frame period through a digital bandpass filter, and performs an acoustic analysis on the audio signal. However,
The feature extraction unit 3 receives the spectrally decomposed audio signal data and processes it in units of frames to extract features of the audio signal. That is, the feature extraction unit 3 extracts the phonetic features of each frame, and for example, attaches a label indicating the vowel type to a vowel frame, and attaches a label indicating the nasal type to a nasal frame. , and other consonants are labeled as plosive, voiceless, voiceless, fricative, etc. As a result, an audio feature time series consisting of a series of labels attached to each frame of the input audio signal is obtained. Note that the above-mentioned labeling of vowel/nasal sounds types is done by calculating the degree of similarity between the spectral pattern data obtained on a frame-by-frame basis and the standard spectral pattern for each type of vowel/nasal sound registered in advance in the dictionary. This is done by determining its label. Further, the above-described consonant type classification is performed by, for example, examining the approximate shape of the spectral pattern obtained for each frame. In other words, if the spectral pattern increases monotonically along the frequency, it is determined to be frictional, and if the spectral pattern has a so-called bell shape with a high center, it is determined to be a rupture. This is done by determining the gender of the person.

The speech section detection circuit 4 extracts the feature time series of the speech signal consisting of the label sequence obtained by the feature extraction section 3 as described above, the spectrum data of the speech signal obtained by the acoustic analysis processing section 2, and its energy data. is input, and the starting end candidate point S and end end candidate point E of the input audio signal are determined according to this information, and multiple audio candidate sections are created by possible combinations of these starting end candidate points S and end end candidate points E. is being detected. In other words, if the number "1" is generated as "ichi" and is input with noise before and after it,
The audio signal waveform is as shown in FIG. 2, for example. For such an input audio signal, start point candidate points S ₁ , S ₂ , S ₃ are determined according to the above-mentioned characteristic time series, spectrum data, and energy data, and end point candidate points E ₁ , E ₂ , E ₃ are determined. seek. The speech candidate section found from the combination of these start and end candidate points is determined as follows in the example shown in Figure 2, since the start end candidate point always exists temporally before the end candidate point. It will be done.

[S ₁ , E ₁ ], [S ₁ , E ₂ ], [S ₁ , E ₃ ] [S ₂ , E ₁ ], [S ₂ , E ₂ ], [S ₂ , E ₃ ) [S ₃ , E ₃ ] Of the above voice candidate sections, [S ₂ , E ₁ ], [S ₃ ,
E ₃ ] has a length less than one audio frame, so it may be excluded from the candidate sections to be processed.

For each of the voice candidate sections obtained in this way, the recognition unit 5 inputs, for example, spectrum information that is a characteristic of the voice signal. Then, the audio signal spectrum information sequence of each audio candidate section is resampled, its characteristic pattern vector is determined, and the similarity is calculated with each standard pattern vector of multiple audio categories registered in advance as a dictionary. Then, voice recognition processing is performed. Recognition unit 5
performs the above-mentioned recognition processing on each of the plurality of speech candidate sections obtained as described above, and outputs the recognition results to the control section 6. This control section 6
While controlling each of the processing units 2, 3, 4, and 5 described above, the recognition unit 5 inputs the recognition results for each speech candidate section obtained by the recognition unit 5 and makes a comprehensive judgment. Then, if the speech candidate section includes noise, the above recognition result (similarity value) is naturally degraded due to this, so this is removed, the most reliable recognition result is extracted, and the above recognition result (similarity value) is removed. This is output as a correct recognition result for the audio signal. In this way, recognition results obtained from highly reliable voice candidate sections can be obtained.
Recognition results obtained from information containing noise added before and after speech are effectively eliminated.
In other words, from among the plurality of speech candidate sections, only information in the speech candidate sections that do not include noise is effectively extracted and recognized.

Note that the speech recognition process for each speech candidate section in the recognition section 5 may be performed using various conventionally proposed methods as appropriate. It goes without saying that various characteristics of the voice used in this recognition process can be adopted.

By the way, the most distinctive feature of this device is
Detection of starting end candidate point S and end end candidate point E for the audio signal, and detection of these starting end candidate point S and end end candidate point S,
The process of detecting voice candidate sections obtained from the combinations of E is performed by the voice section extraction circuit 4 as follows. FIG. 3 is a flowchart showing an example of the processing process. This processing is performed by first initializing the processing control counter value and then inputting the audio signal energy of the nth frame. Afterwards,
For example, divide an audio signal into a silent class and a voice class according to a temporarily set threshold, find the interclass variance between each class, and set the optimal threshold to maximize the value.
Set E _th and calculate its threshold E _th and input audio energy.
Compare E _(o) . Then the input audio energy E _(o)
The time point at which E _{th exceeds the threshold value E th} is defined as S' _(i) and is the first candidate point for the starting point. Then, the point in time when the input audio energy E _(o) falls below the threshold E _th is detected, and this is set as the first candidate point E' _(i) of the end. Afterwards,
The interval between the starting and ending candidate points obtained in this way is determined as T _k = |S′ _(k) −E′ _(k) |, and it is determined whether or not it exceeds a predetermined interval T _th . Find voice candidate sections. As a result, erroneous speech candidate sections obtained in fragments are removed. Then, the above-mentioned voice candidate section detection is sequentially inputted over all frames of the input voice signal, and all possible combinations thereof are checked to obtain all voice candidate sections.

Thereafter, the speech characteristics of the speech candidate section are examined, it is determined whether or not the section contains a noisy component, and the speech candidate section containing the noisy component is removed from the recognition target. Then, the end of the utterance is determined by e.g.
The section where E _(o) < E _th is detected because it continues for a predetermined period M _th , and the speech features in the speech candidate section detected so far are compared with the utterance shape of the word registered in advance. , determine the final voice candidate section. For all of the speech candidate sections determined in this way, for example, the degree of similarity with the word dictionary registered in the dictionary is determined using the composite similarity method, and the similarity values are compared with each other to find the most reliable one. The results with high results are output as recognition results.

According to the recognition process described above, the recognition results obtained in each of the candidate sections containing word sounds are compared with each other and the most reliable one is extracted, so that the final utterances are discretely uttered. The recognition results obtained from the characteristics of the voice itself can be obtained with high reliability. In other words, since the detection of the voice section and the recognition process are performed in conjunction with each other, a great practical effect is achieved in that the recognition process can be performed stably.

Note that the present invention is not limited to the above embodiments. For example, various methods can be used as appropriate for the process of detecting the start end candidate point and the end candidate point, the process of extracting the speech candidate section obtained by a combination of these, and the recognition process. Furthermore, there are no particular limitations on the characteristics of the audio used in the above-described processing. In short, the present invention can be implemented with various modifications without departing from the gist thereof.

[Brief explanation of drawings]

FIG. 1 is a schematic diagram of an apparatus according to an embodiment of the present invention;
FIG. 2 is a diagram showing an audio signal waveform, starting and ending candidate points, and their audio candidate sections, showing the processing of the embodiment device, and FIG. 3 is a diagram showing the flow of recognition processing in the embodiment device. 1...Audio input section, 2...Acoustic analysis processing section, 3
...Feature extraction unit, 4...Speech section detection circuit, 5...
...Recognition unit, 6...Control unit.

Claims (1)

[Claims] 1. Means for inputting an audio signal and performing acoustic analysis at each predetermined frame period, and from this acoustic analysis result, labeling vowels and nasals for each frame based on phoneme similarity, and labeling from the outline of the spectrum. means for labeling consonants to extract features of a speech signal; means for detecting a starting point and an end candidate point of the speech signal according to the features of the speech signal and the acoustic analysis results; The vowel/nasal labels and consonant labels obtained for each frame are used to determine whether the segment is a voice or not for multiple speech candidate sections obtained by all possible combinations of the above starting point candidate points and end point candidate points. means for determining if it is noise; means for resampling and extracting the features of the speech signal in the speech candidate sections detected as speech; What is claimed is: 1. A speech recognition device comprising: recognition means.