JPH02240700A - Voice recognizing device - Google Patents

Abstract

PURPOSE:To execute phoneme recognition with high accuracy and to recognize continuous voices of a large vocabulary by comparing and collating the characteristics of the phoneme segments extracted from an input voice signal and the characteristics of the phoneme segments described in a phoneme knowledge base, thereby recognizing the phoneme. CONSTITUTION:The respective parameters obtd. in an acoustic analyzing circuit 5 are supplied as parameters for recognition processing to a phoneme recognizing circuit 8. The respective parameters outputted from circuits 51 to 56 are supplied as the parameters for segmentation to a characteristic point extracting circuit 61 of a 1st segmentation circuit 6. The respective parameters attached with the characteristic points are supplied to a 2nd segmentation circuit 7. The phoneme characteristics of the respective phoneme segments extracted from the parameters for the recognition processing are compared and contrasted with the phoneme characteristics of the phoneme segments stored in the phoneme knowledge base and the phoneme candidate arrays are outputted in accordance with the results thereof in the phoneme recognizing circuit 8. The phoneme recognition with the high accuracy is executed in this way and the continuous voices of the large vocabulary are recognized.

Description

DETAILED DESCRIPTION OF THE INVENTION [Field of Industrial Application] The present invention relates to a speech recognition device, and particularly to a speech recognition device equipped with a phonological knowledge base and inference means.

[Summary of the invention]

The present invention provides a speech UW recognition device for vowels, consonants,
Speech is divided into phonological segments at stationary parts, transitional parts, etc. of formants, and a phonological recognition base described in units of segments is provided, and features of the phonological segments extracted from the input speech signal and phonological segments based on phonological knowledge are provided. By comparing and matching the features of the computer and obtaining phonological candidates based on the results of this comparison and matching, highly accurate phonological recognition can be performed and human speech recognition and continuous speech recognition can be performed. be.

[Conventional technology]

In conventional phoneme recognition, the spectral pattern of input speech is recognized by comparing and matching it with a standard phonological spectral pattern.
96104] was carried out for reference. However, with pattern matching, it is difficult to detect subtle differences between consonants, such as ^TA and APA.

Therefore, in order to improve this problem, attempts have been made to perform phoneme recognition using an expert system that uses phoneme features, phoneme identification rules, etc. as a phoneme knowledge base.

[Problem to be solved by the invention]

However, in phoneme recognition in expert systems, there is a problem in how to describe phoneme features, phoneme identification rules, etc. in a phoneme knowledge base.

Therefore, an object of the present invention is to provide a speech recognition device that improves the description of phoneme features, phoneme identification rules, etc. for a knowledge base.

[Means to solve the problem]

This invention divides speech into phonological segments based on constant parts, transition parts, etc. of vowels, consonants, formants, etc., has a phonological knowledge base described in units of segments, and has characteristics of phonological segments extracted from input speech signals. and the features of the phoneme segment based on the phoneme knowledge base are compared and verified, and phoneme candidates are obtained based on the results of this comparison and verification.

[Effect]

Divide speech into phonological segments based on predetermined conditions,
Next, features are obtained for each phoneme segment and input into the inference means.

On the other hand, in the phoneme knowledge base, the characteristics of each phoneme are described in units of phoneme segments using, for example, if---then--type rules. The inference means compares and collates the features of each phoneme segment with the features of the phoneme segment based on the phoneme knowledge base, obtains phoneme candidates based on this, and specifies the phoneme.

As a result, highly accurate phoneme recognition can be performed, making it possible to recognize large words and continuous speech.

〔Example〕

An embodiment of the present invention will be described below with reference to FIGS. 1 to 8.

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

Input audio is converted into an audio signal by a microphone 1, and is supplied to an A/D conversion circuit 4 via an amplifier 2 and a low-pass filter 3. The audio signal is sent to the A/D conversion circuit 4.
For example, it is converted into a 12-bit digital audio signal at a sampling frequency of 12.5 Hz. This digital audio signal is supplied to the acoustic analysis circuit 5.

The acoustic analysis circuit 5 includes a transient detection parameter generation circuit 51 having a bandpass filter bank, a logarithmic power detection circuit 52 for detecting audio power, a zero cross rate calculation circuit 53, and a primary , a calculation circuit 54 for calculating the Percoll coefficient, a calculation circuit 55 for calculating the slope of the power spectrum, a formant detection circuit 56 for determining changes in the formant in the time direction, and a detection circuit 57 for the fundamental period of speech.

The transient detection parameter is used to detect the transient nature and stationarity of the input audio, and the amount of change in the audio spectrum is defined as the sum of variances within a block in the time direction of each channel (frequency). That is, the gain of the audio spectrum 5i(n) is normalized using the average value S avg (n) shown below in the frequency direction.

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

The speech spectrum 5i(n) after gain normalization is S i (n) = S i (n) S avg
(n) ・----(2).

The transient detection parameter T (n) is the sum of H frames before and after that frame (2M+1) (n-M, n+
[gate] is defined as the sum of the time-direction variance of each channel within the block.

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

In practice, changes near the center of the (n - M + n + M) block are likely to pick up sound fluctuations or noise, so they are removed from the calculation of the transient detection parameter T (n), and Equation (3) is transformed as follows.

In equation (5), - as an example, a.1. M
- The transient detection parameter T (n) in the case of 2B, m=3.4=32 is found. For example, ``Akiyo (akiyo)
o) For an input voice J, a transient detection parameter T (n) as shown in FIG. 2A is obtained.

Other parameters, such as the logarithmic power shown in Figure 2B, the zero crossing rate shown in Figure 2C, the first-order Percoll coefficient shown in Figure 2, and the slope of the power spectrum shown in Figure 2 Detection and calculation of parameters such as the fundamental period shown in FIG. 2H are also performed using the transient detection parameter T.
Similarly to (n), consider a window that has a time width of h frames before and after the point in time (frame) at the center, and sequentially move this window one sample point at a time in the time direction. Obtained by performing calculations. In addition, Fig. 2 F and J show the waveform of the input voice ``Akyo J'', Fig. 2 ■ shows the formant transition, and Fig. 2 G and K show the phoneme obtained based on the above-mentioned parameters. An example of a boundary candidate is shown. In Fig. 2, F
, and J, G, and K have the same content and are shown redundantly for convenience of comparison with other parameters.

Each parameter obtained by the acoustic analysis circuit 5 is supplied to the phoneme recognition circuit 8 as a parameter for recognition processing, and the circuit 51
Each parameter output from 56 is supplied to the feature point extraction circuit 61 of the first segmentation circuit 6 as a segmentation parameter.

The first segmentation circuit 6 extracts negative feature points in order to find phoneme boundary candidates from the segmentation parameters. In this example, the following 15 types of feature points are used.

■ Rising point - the point where the flat part changes to an increasing direction ■ Falling point - the point where it changes to a decreasing direction and then becomes flat ■ Increased change point - the point where the rate of increase changes ■ Decrease change point - the decreasing rate ■Peak point - the position of the peak ■Positive zero crossing point - the point that crosses the zero level in the increasing direction ■Negative zero crossing point - the point that crosses the zero level in the decreasing direction 0 Word beginning, ending (rising from silence) , falling to silence), the point at which the unstable part at the beginning and end of the 0th word becomes stable [phase]
Rising and falling to silent intervals due to mid-word pauses ■Consonant interval ← → Change point in vowel interval @ From the steady formant interval within the vowel interval to the start point of the transition interval, or the end point of the transition interval ■Formant within the consonant interval From the steady interval to the start point of the transition interval, or the end point of the transition interval [Phase] Point where the formant occurs, point where it disappears■Start point and end point of the voice bar interval.In this specification, the voice bar refers to the point before a voiced consonant. A voice signal consisting only of low frequency components that occurs when the lips are closed and the vocal cords vibrate.

The feature point extraction circuit 61 refers to the feature point information from the feature point information storage circuit 62 and extracts feature points for each parameter. Among the parameters in FIGS. 2A to 2E, the positions indicated by the curtain lines in the time axis direction are the positions of each feature point.

Each parameter obtained from the first segmentation circuit 6 and to which feature points are attached is sent to the second segmentation circuit 7.
is supplied to

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

Since the feature points obtained by the first segmentation circuit 6 may have positional deviations or undetected points for each parameter, the feature point integration processing circuit 71 refers to the feature point integration information from the feature point integration information storage circuit 73 and The feature points of the parameters are summarized and phoneme boundary candidates are determined.

Note that the feature point integration information is information about which parameter's feature point should be prioritized.

The phoneme boundary feature detection circuit 72 determines the phoneme boundary feature of each phoneme boundary candidate. In this example, the following phonetic boundary features are used.

■ Rise from silence (SIL-R) ■ Consonantity → Vowelness (C-V) ■ Vowelness → Vowelness (V-V) ■ Vowelness → Vowel transition (V-V, T) ■ Vowel Transitional part → Consonantity (V, T-C) ■ Consonantity → Vowel transitional part (C-
V, T) ■Vowel transition → vowel character (V, T-V) ■Falling to silence (F-S IL) ■Sound → silence (SND
-3IL) [Phase] Consonance → Consonance (C-C) ■Consonance → Consonant transition (C-C, T) ■ Consonant transition → Consonance (C, T-C) Phonological boundary feature information Memory circuit 74
These 12 types of phoneme boundary feature information are stored in the phoneme boundary feature detection circuit 72, and the phoneme boundary feature detection circuit 72 detects the phoneme boundary feature of each phoneme boundary candidate by referring to information from the phoneme boundary feature information storage circuit 74. .

The second segmentation circuit 7 obtains phoneme boundary candidate information and its phoneme boundary feature information as phoneme interval information. This phoneme segment information is then supplied to the phoneme recognition circuit 8.

The phoneme recognition circuit 8 executes phoneme recognition based on the recognition processing parameters from the acoustic analysis circuit 5 and the phoneme interval information from the second segmentation circuit 7.

The phonological recognition circuit 8 uses the phonological knowledge base [
The information is compared and collated with the phoneme features of the phoneme segment stored in the knowledge base (hereinafter simply referred to as knowledge base). Then, based on this result, a phoneme candidate sequence is output. This process will be explained based on the flowchart of FIG.

As described above, the first and second segmentation circuits 6
．． 7, a phonological segment is formed (step 101
).

Next, the phoneme recognition circuit 8 extracts phoneme features of each phoneme segment. Specifically, according to the statistics of recognition processing parameters from the acoustic analysis circuit 5, in a phoneme segment where the formant is a stationary interval, the articulation method is vowel/consonant, voiced/unvoiced, fricative/plosive. /nasal, etc. (step 102).

Furthermore, the phonological feature detector provided in the phonological recognition circuit 8 obtains information such as rupture points, voice bars, cutoff frequencies of frequency bands where frictional energy is concentrated, and formant transition methods (step 103).

Through the above processing, phoneme features for each phoneme segment are determined. Based on this, if---then type inference is performed in the following steps [Step 104]
. Incidentally, in the knowledge base used for this inference, an if--then type rule is described in step 105. Further, in the following inference, if necessary, the process returns to feature extraction of phoneme segments [step 102] and reprocessing is executed.

(2) Determine the articulatory position (eg, labial, stem, palate) of each preceding and subsequent consonant segment using the first speculative formant transition method.

Rule (11) If the formant transition for the following vowel /a/ is as shown in Figure 4, it is a bilabial sound. In the figure, Fl and F2 represent the first formant and the second formant, respectively.

Rule (12) If the formant transition for the following vowel /a/ is as shown in FIG. 5, it is an alveolar stage.

Rule (13) If the formant transition for the following vowel /a/ is as shown in Figure 6, it is a palatal sound. Note that other rules are omitted.

(2) Determine the phonology of the consonant segment using the second inferential articulation method and articulation position.

Rule (21) If the articulation method is a voiceless fricative and the articulatory position determined from the subsequent formant transition is a bilabial, then /f/.

Rule (22) If the articulation method is a voiceless fricative and the articulation position is an alveolar sound, it is /s/.

Rule (23) If the method of articulation is a voiceless fricative and the position of articulation is the palatal base, it is /sh/.

Rule (24) If the method of articulation is a voiceless plosive and the position of articulation is the palatal stand, then it is /ni/.

Rule (25) If the method of articulation is a voiced plosive and the position of articulation is the labial stand, it is /b/. Note that other rules are omitted.

(2) Third inferential articulation method When the phoneme cannot be specified by the articulatory position, the phoneme of the consonant segment is determined using phoneme features such as burst, voice bar, and cut-off frequency.

Rule (31) If there are two or more bursts in a voiceless plosive, it is /ni/.

Rule (32) The length of the rupture section is /p/ </l/
The order is <8/. Note that other rules are omitted.

(2) A semi-vowel is determined for a long interval of formant transition in the fourth inferred vowel interval.

Rule (41) If the formant transition in the following vowel 10/ is as shown in FIG. 7, it is a consonant /y/.

Note that other rules are omitted.

Checking the consistency of the phoneme candidate string found by the above-mentioned if-the-n type inference [step 106]
is performed based on the phoneme connection knowledge base described in step 107. In step 106, it is checked whether the phoneme candidate string is correctly connected as a Japanese musical piece. Incidentally, if the phoneme candidate string cannot be specified even by this check, the process returns to the previous steps 102 and 104 and reprocessing is performed.

The above reasoning will be explained using the input voice "Akyo J" as an example.

(2) After phoneme segmentation, articulation methods are classified based on the various parameters mentioned above.

The phoneme boundary features of the phoneme boundary candidates shown in FIGS. 2G and 2K are defined as follows.

■(SIL-R) ~(C-V) Consonant Voiced■
(C-V) ~ (V-V, T) Vowelness Voiced ■ (V-
V, T) ~ (V, T-3IL) Vowel voiced ■ (V, T-3IL) ~ (SND-SIL) Consonant voiceless ■ (SND-3IL) ~ (SIL-R) Silent ■ (S
IL-R) ~ (C-V, T) Consonant Voiceless (plosive, 2 bursts) ■ (C-V, T) ~ (V, T-V) Vowel Voiced ■ (V
, T-V) ~ (F-3IL) Vowelic Voiced ■ (F-3I
L) ~ (SND-3IL) Consonant voiceless 11, (V
Since the preceding vowel of the formant transition between -V, T) and (V, T-3IL) is /a/, the following consonant is a palatal base from the formant transition shown in FIG. 8 and rule (13).

The phonological segment of IIl, (SrL-R-C-V, T) is 7kl based on the plosive and palatine features of rule (24).

IV, (C-V, T ~ V, T-V) (7)
Since the segment has a long formant transition period, a semi-vowel check is performed, and the result is the consonant /y/ since it is the same as the formant transition in rule (41).

Based on 70 or more inferences, the phoneme recognition circuit 8 outputs /a/
A phoneme candidate string of /+/y/+10/ is output to +/.

In this way, the phonological features of the phonological segment formed based on the input speech signal are compared and matched with the phonological features of the phonological segment that have been improved and described in the knowledge base, and the sound [
Since it recognizes y2, it can perform highly accurate phoneme recognition, and it becomes possible to recognize large words and continuous speech.

〔Effect of the invention〕

According to this invention, the description of phoneme features, phoneme identification rules, etc. for the phoneme knowledge base is improved, and the characteristics of the phoneme segment extracted from the input speech signal and the characteristics of the phoneme segment described in the phoneme knowledge base are improved. Since the sound n is recognized by comparing and collating the sound n, it is possible to perform the sound fl#X1lk with high accuracy, and there is an effect that it is possible to perform large word comprehension and continuous speech recognition.

[Brief explanation of drawings]

Fig. 1 is a block diagram showing an embodiment of the present invention, Fig. 2 is a waveform diagram of each parameter, Fig. 3 is a flowchart showing the inference procedure, and Figs. 4 to 8 each show a pattern of formant transition. FIG. Explanation of main symbols in the drawings 5: Acoustic analysis circuit, 6: First segmentation circuit,
7: Second segmentation circuit, 8: Phonological recognition circuit.

Claims (1)

[Claims] Speech is divided into phonological segments by constant parts, transition parts, etc. of vowels, consonants, formants, etc., and a phonological knowledge base described in units of segments is provided, and the features of the phonological segments extracted from the input speech signal and the above A speech recognition device that compares and matches features of phoneme segments based on phoneme knowledge and obtains phoneme candidates based on the results of this comparison and check.