JP2569470B2 - Formant extractor - Google Patents

Description

Description: TECHNICAL FIELD The present invention relates to a formant extractor.

2. Description of the Related Art Speech formants are used as extremely effective parameters in the fields of speech analysis, synthesis, and recognition.

Such a formant is usually set in advance by n
Next LPC (Linear Prediction Coding)
The formant frequency is determined based on n / 2 conjugate complex solutions obtained by solving a higher-order equation having coefficients as constants, that is, a known method using a pole frequency.

[Problems to be solved by the invention]

However, the above-mentioned conventional formant extraction means,
There is a drawback that the formant frequency extracted for each LPC analysis frame does not easily lose continuity between frames. The cause of this is that basically the effects of fluctuations in speech from frame to frame and fluctuations in the cut-out phase at the time of clipping by the window function in speech quantization cannot be avoided in the formant extraction processing, and this will occur as a formant extraction error. It depends.

An object of the present invention is to eliminate the above-mentioned drawbacks and to significantly reduce formant extraction errors by providing a means for determining a formant through a pattern matching process between a formant standard pattern and analytical data. It is to provide an extractor.

[Means for solving the problem]

A formant extractor according to the present invention is a standard that uses two angular relationship information between formant frequencies obtained by analyzing first to third formant frequencies obtained by analyzing an input audio signal in polar coordinates as a vector element. A formant trajectory obtaining means for obtaining a formant trajectory through pattern matching using time normalization of a pattern and analysis data on the angle relationship information of the input audio signal, and a formant ratio of the first to third formants is determined by the formant Conversion means for converting the frequency into an absolute value.

〔Example〕

 Next, the present invention will be described in detail with reference to the drawings.

FIG. 1 is a block diagram showing one embodiment of the present invention.
A / D converter 1, autocorrelation coefficient calculator 2, formant judgment unit 3, polar coordinate converter 4, pattern collator (1) 5, standard pattern file 6, pattern collator (2) 7, orthogonal coordinate converter 8 It is comprised including.

The input audio signal supplied via input line 101 is A / D
The converter 1 cuts off unnecessary high frequency components through an LPF (Low Pass Filter), samples it at a predetermined sampling frequency, performs quantization of a predetermined number of bits, and converts it into a quantized audio signal. In this embodiment, the quantization process is performed at a sampling frequency of 8 KHz and a quantization bit number of 12 bits.

The quantized audio signal is, for example, 30 mSEC
240 samples are temporarily stored in the internal memory as window time, and weighted multiplication by a window function such as a Hamming function or a rectangular function is performed for each predetermined repetition cycle of 20 mSEC, and this becomes an analysis frame, which is quantized for each analysis frame. The audio signal is supplied to the autocorrelation coefficient calculator 2.

The autocorrelation coefficient calculator 2 extracts an autocorrelation coefficient within a required time delay range of the quantized speech signal input for each analysis frame to a predetermined order, up to the 12th order in the present embodiment, and extracts this to a formant determiner 3. To supply.

The formant extractor 3 has 12 input frames for each analysis frame.
A known formant extraction method for the following autocorrelation coefficient, for example, AbS (Analysis by Synthesi
s) to determine the first to third formant frequencies. In this case, the formant frequency to be determined is set to the third
It is known that if the first to third formant frequencies are secured, the expression of voiced speech for purposes such as speech recognition and synthesis has an almost dominant meaning, and the variation in appearance This is for excluding the fourth formant or higher, which is highly likely. It is also well known that this formant frequency substantially coincides with the pole frequency as a vocal tract resonance point.

The first to third formant frequencies ₁ , ₂ , and ₃ thus determined by the formant determiner 3 have individual differences in their absolute values, but the distribution of the ratio ₁ : ₂ : ₃ depends on the speaker. It is known that it is almost constant without any.

The polar coordinate converter 4 receives such first to third formant frequencies and performs the polar coordinate conversion.

FIG. 2 is a three-dimensional polar coordinate display diagram of the first to third formant frequencies.

Three coordinate axes orthogonal to each other at the origin 0 indicate formant frequencies ₁ , ₂ , and ₃ , respectively, and _two angles θ that determine the spatial direction of the composite vector V ( ₁ , ₂ , ₃ , ₃ ) of these formant frequencies are defined between the formant frequencies. The relationship angle. These two angles θ are ₁ :
If the ratio of ₂ : ₃ is the same, it is the same, and from a different viewpoint, it is an angular expression of the frequency spectrum of the audio signal for each analysis frame. The angular formant frequency thus expressed is substantially constant irrespective of the speaker, except that the synthesized vector absolute value l differs for each speaker. In the case where pattern matching is used instead of feature parameters, matching processing excluding speaker dependency can be performed.

The two mutual relationship angles θ, indicated by the composite vectors of the formant frequencies ₁ , ₂ , and ₃ expressed in polar coordinates, are stored in the internal memory in units of conversational terms and read out, and the pattern matching unit (1) 5 and the pattern matching The resultant vector absolute value 1 is supplied to an orthogonal coordinate converter 8.

The formant frequency correlation angle θ, supplied to the pattern collator (1) 5, performs pattern collation while receiving the standard pattern from the standard pattern file 6 via the output line 601.

The standard pattern file 6 contains audio data by a specific speaker,
In the case of the present embodiment, a limited number of conversation terms used for communication in the formant vocoder used for voice narrowband communication were obtained by offline analysis with the formant extractor according to the present embodiment or a separately prepared computer system. A standard pattern of analysis frame unit regarding θ, is stored and supplied to the pattern matching unit (1) 5.

The pattern matching unit (1) 5 performs pattern matching for each analysis frame via the standard pattern and the formant frequency correlation angle θ, and matches the standard pattern having the closest distance θ to the analysis data of the input voice signal. Then, the label number is selected as a standard pattern, and its label number is sent to the standard pattern file 6 via the output line 501, and the data of θ relating to the standard pattern from the standard pattern file 6 is output to the pattern collator (2). In this pattern matching, the matching between the standard pattern and the analysis pattern via θ is performed in the form of measuring the city distance or the Euclidean distance between the two patterns θ.
In the case of this embodiment, the city area distance di expressed by the following equation (1) is used.
It is implemented in the form of selecting a standard pattern that minimizes j.

In the equation (1), K indicates an input pattern, S indicates a standard pattern θ, and i = 1, 2,..., N
Where n is the total number of standard patterns, j = 1, 2,..., M and m is the total number of input patterns.

By the way, the number of speakers that can be used for creating a standard pattern used for pattern matching is limited for many economic and other reasons. Therefore, training by such a limited number of speakers (training, registration data) It is basically impossible to create a standard pattern that can be adapted to any speaker by using. This is because the spectral distribution of speech differs for each speaker, that is, the vocal tract characteristics and the vocal cord sound source characteristics differ for each speaker. The difference in vocal tract characteristics is due to the difference in vocal tract length for each speaker, which means that the formant frequency as a resonance point of the vocal tract differs for each speaker. On the other hand, differences in vocal fold sound source characteristics affect the slope of the general shape of the spectral envelope. Therefore, in order to perform pattern matching that is easy to adapt to unspecified speakers, it is necessary to normalize the vocal tract characteristics and vocal cord sound source characteristics that differ for each speaker by some means, or to remove the effects. Required.

Usually, since the two characteristics are convoluted, it is necessary to separate them and normalize or eliminate speaker dependence. The present invention also focuses on this point, and implements pattern matching that eliminates speaker dependence by representing a formant frequency with θ.

In this embodiment, the vocal tract length is normalized by θ, and the spectral envelope is represented by the ratio of the first to third formants.

Now, the pattern matching unit (2) 7 is supplied with θ relating to the standard pattern matched to the input pattern from the standard pattern file 6 and the polar coordinate converter 4 with θ of the input audio signal.

The standard pattern registered and stored in the standard pattern file 3 is a high-precision pattern created by itself so as to make the formant extraction error almost zero, but only a specific generation rate by a specific speaker. If the input pattern of the specific speaker as well as the specific speaker is compressed or expanded on the time axis and time normalization is not performed, the reproduction of the formant frequency of the specific speaker will not be possible. Can not.

The pattern collator (2) 7 selects the standard pattern θ,
Is subjected to time normalization for temporally compressing or expanding θ of the input pattern. This normalization is performed by finding a mapping function that makes the standard pattern and the input pattern expressed on the mutually orthogonal time axes i and j correspond to each other in a mapping relationship. This is a well-known technique that implements dynamic programming using the urban distance dij shown in equation (1) as an evaluation scale, and grasps the mapping function as a monotonically increasing dynamic programming path that minimizes the sum of the urban distances dij. , Can be easily implemented by other processing techniques.

The time-normalized standard pattern θ is then supplied to the orthogonal coordinate converter 8.

The orthogonal coordinate converter 8 converts polar coordinate data based on the time-normalized θ and the composite vector absolute value l input from the polar coordinate converter 4 into rectangular coordinates, and expresses the time expressed in a time axis versus frequency axis rectangular coordinate system. A formant trajectory is provided while maintaining continuity as sequence data, and is transmitted as output formant data. By treating the formant trajectory as a pattern on the time axis, the formant extraction error can be greatly reduced. In addition, in the case of using for voice narrow-band communication as in the present embodiment, since the conversational term to be prepared as a standard pattern is greatly limited, the system configuration is easy, and the unspecified speech which can make operation reliability high can be achieved. Formant extraction for the elderly.

In the embodiment shown in FIG. 1, the pattern matching is performed by the pattern matching unit (1) 5 and the pattern matching unit (2) 7. However, even if these functions are integrated, they can be easily realized. No problem.

〔The invention's effect〕

As described above, according to the present invention, the first to third formants obtained by analyzing the input audio signal are expressed in polar coordinates, and the standard pattern file having the vector information as the formant frequency inter-relationship angle information is obtained. Formant trajectory acquisition means for extracting a formant whose formant trajectory is secured through pattern matching using time normalization with data, and conversion means for converting the ratio of the first to third formants to the absolute value of the formant frequency With this arrangement, the formant extraction error can be significantly reduced, and a formant extractor with an unspecified speaker can be realized.

[Brief description of the drawings]

FIG. 1 is a block diagram showing an embodiment of the present invention. FIG. 2 is a diagram showing three-dimensional polar coordinates of first to third formant frequencies. 1 ... A / D converter, 2 ... Autocorrelation coefficient calculator, 3
... Formant detector, 4 ... Polar coordinate converter, 5 ... Pattern collator (1), 6 ... Standard pattern file, 7 ...
Pattern collator (2), 8... Cartesian coordinate converter.

Claims (1)

(57) [Claims] 1. A standard pattern in which two angular relationship information between formant frequencies obtained by expressing first to third formant frequencies obtained by analyzing an input audio signal in polar coordinates is used as a vector element. A formant trajectory obtaining means for obtaining a formant trajectory through pattern matching using time normalization with the analysis data relating to the angular relation information of the audio signal, and a ratio of the first to third formants is calculated by calculating an absolute value of the formant frequency. A formant extractor, comprising: conversion means for converting into a value.