JP2002515608A - Method and apparatus for determining spectral speech characteristics of uttered expressions - Google Patents

Abstract

According to the invention, spectral voice characteristics are determined in a natural language expression, whereby the expression is digitized and subjected to a wavelet transformation. The speaker-specific characteristics arise from the different transformation steps of the wavelet transformation. Within the scope of a voice synthesis, these characteristics can be compared with characteristics of other expressions in order to generate a continuously sounding synthetic voice signal for the human ear. Alternatively, the characteristics can also be modified in a targeted manner in order to counteract a perceptive dissonance.

Description

DETAILED DESCRIPTION OF THE INVENTION

The present invention relates to a method and an apparatus for determining a spectral speech feature in an uttered expression.

In concatenated speech synthesis, individual sounds are synthesized from a speech data bank. In this case, in order to obtain a sound course that sounds natural to the human ear, discontinuities at the points where the sounds are assembled (connection points) must be avoided. Here, the sound is, for example, a collection of phonemes of a language or a plurality of phonemes.

The wavelet transform is known from [1]. In the wavelet transform, the wavelet filter ensures that the high-pass and low-pass components of each successive transform step completely restore the signal of the current transform step. Here, the resolution of the high-pass component or the low-pass component is reduced in one conversion step to the next conversion step (in English terminology, “subsampling”).
Is). In particular, due to this subsampling, the number of conversion steps is finite.

[0004] It is an object of the present invention to provide a method and a device for determining spectral audio features, for example to provide a natural sounding synthetic audio output.

[0005] This problem is solved by a configuration described in the characterizing part of claim 1.

[0006] Within the framework of the invention, a method is described for determining the spectral speech characteristics of an uttered expression. For this purpose, the uttered expression is digitized and subjected to a wavelet transform. The unique features of the speaker are determined based on different transform steps of the wavelet transform.

It is particularly advantageous here that in the wavelet transform, the expression is divided by a high-pass filter and a low-pass filter, and that different high-pass or low-pass components of different transformation steps have speaker-specific features. It is.

[0008] The individual high-pass or low-pass components of the different transform steps represent certain features specific to the speaker, and both the high-pass and low-pass components of each transform step can be changed, ie each feature is distinct from another feature. Can be changed to Reassembling the original signal from the high-pass and low-pass components of the individual transform steps during the inverse wavelet transform ensures that only the desired features are changed.
Thus, a predetermined, predetermined characteristic of the expression can be changed, so that the rest of the expression is not affected.

A feature of one embodiment is that the expression is windowed before the wavelet transform, ie, a predetermined amount of the sampling value is cut out and transformed into the frequency domain. For this purpose, for example, a fast Fourier transform (FFT) is applied.

[0010] A feature of another embodiment is to separate the high-pass component of the transform step into a real part and an imaginary part. The high-pass component of the wavelet transform corresponds to the difference signal between the current low-pass component and the low-pass component of the preceding transform step.

A feature of a development is in particular that the number of transform steps to be performed of the wavelet transform is determined by the fact that the last transform step consisting of a series of connected low-pass filters contains an expressed DC component. Is to be done. In this case, the signal can be represented as complete by its wavelet coefficients. This corresponds to the fact that the information of the signal part is completely transformed into the wavelet space.

In particular, if, for example, only each low-pass component is further transformed (by high-pass and low-pass filters), the difference signal remains as the high-pass component of the transformation step, as explained above. When the difference signal (high-pass component) is accumulated over the conversion step, information of an uttered expression without a DC component is obtained as the high-pass component accumulated in the last conversion step.

In an additional development, the speaker-specific features can be identified as:

A) Fundamental frequency: The expressed fundamental frequency is identified by the oscillation of the high-pass component in the first or second transform step of the wavelet transform. The fundamental frequency indicates whether the speaker is male or female.

B) Shape of the spectral envelope: The spectral envelope contains information about the transfer function of the vocal tract during articulation. In voiced regions, the formants predominate in the spectral envelope. The high-pass component in the higher-order transform step of the wavelet transform includes this spectral envelope.

C) Spectral tilt (Rauchigkeit) The degree of voice wrinkling is identified as a negative slope in the course of the previous previous low-pass component.

The features a) to c) unique to the speaker are extremely important in speech synthesis. As mentioned at the beginning, when using a large number of actual spoken expressions in concatenated speech synthesis, exemplary sounds are cut out of these expressions and then assembled into new words (synthesized speech). ). In this case, the discontinuity between the assembled sounds is disadvantageous. This is because it is unnaturally perceived by the human ear. To counter these discontinuities, it is advantageous to directly detect, possibly compare, and / or match multiple perceptual parameters.

This can be done by direct manipulation. This direct manipulation is performed by adapting the audio sound in at least one of the speaker's unique characteristics so that the audio sound is not perceived as an obstacle in the acoustic context of the sound combined by the concatenation. Will be Also, the selection of matching sounds can be adjusted so that the speaker-specific characteristics and the sounds to be combined are matched as well as possible to each other, for example, so that these sounds have the same or similar wrinkling characteristics. It is also possible to

An advantage of the present invention is that the spectral envelope can be controlled by the speaker's Artikulationstrakt.
And does not rely on formants, such as the polar position model. Furthermore, in the wavelet transform as a non-parametric representation, no data is lost, and the expression can always be completely restored. The data resulting from the individual transform steps of the wavelet transform are linearly independent of one another and can therefore be varied separately and later reassembled (without loss) into the altered representation.

The invention further provides an apparatus for determining a spectral feature, comprising a processor unit. The processor unit is configured to digitize the representation. Based on this, the expression is subjected to wavelet transformation, and the unique characteristics of the speaker are determined by different transformation steps.

This device is particularly advantageous for implementing the method of the invention or the development described above.

[0022] Developments of the invention are set out in the dependent claims.

Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.

Here, FIG. 1 shows a wavelet function, FIG. 2 shows the wavelet function divided into a real part and an imaginary part, and FIG. 3 shows a cascade connection showing a transform step of the wavelet transform. FIG. 4 shows a low-pass component and a high-pass component of different conversion steps, and FIG. 5 shows a concatenated speech synthesis step.

FIG. 1 shows a wavelet function determined by the following equation.

[0026]

(Equation 1)

Here, f represents a frequency, σ represents a standard deviation, and c represents a predetermined normalization constant.

The standard deviation σ is determined, for example, by the position of the sideband minimum value 101 in FIG. 1, which can be set in advance.

FIG. 2 shows a real part of Expression (1) and a wavelet function having the real part as an imaginary part obtained by performing the Hilbert transform H. Therefore, this complex wavelet function is obtained by the following equation.

[0030]

(Equation 2)

[0031]

[Outside 1]

FIG. 3 shows that the wavelet transform is cascaded and applied. The signal 301 includes a high-pass filter HP1 302 and a low-pass filter TP13.
05. Here, for example, subsampling is performed, that is, the number of values to be stored is reduced for each filter. The inverse wavelet transform ensures that the original signal 301 is restored again from the low-pass component TP1 305 and the high-pass component HP1 304.

In the high-pass HP 1 302, filtering is performed separately according to the real part Re 1 303 and the imaginary part Im 1 304.

The signal 310 after the low-pass filter TP 1 305 is newly filtered by the high-pass filter HP 2 306 and the low-pass filter TP 2 309. The high-pass filter HP2 306 also has a real part Re2 307 and an imaginary part Im2.
308. After the second conversion step 11, filtering is performed again, and this is repeated.

If one starts with a short-time (FFT-transformed) spectrum having 256 values, the eight transform steps (subsampling rate: 1/2) are based on the fact that the signal from the last low-pass filter TP8 is a DC component. Is executed until it is equal to

FIG. 4 shows different transform steps of the wavelet transform, which are represented by low-pass components (FIG. 4).
A, 4C and 4E) and high-pass components (FIGS. 4B, 4D and 4F).

The fundamental frequency of the uttered expression can be seen from the high-pass component of FIG. 4B. In addition to amplitude variations, periodicities that clearly dominate the wavelet filtered spectrum can be identified. This is the speaker's fundamental frequency.
With this fundamental frequency, a preset expression can be adapted to each other at the time of speech synthesis, or a suitable expression can be obtained from a data bank provided with the preset expression.

In the low-pass component of FIG. 4C, the formant of the audio signal portion (the length of the audio signal portion is approximately twice the fundamental frequency) is shown as the remarkable minimum and maximum values. These formants represent the resonant frequency of the speaker's vocal tract. The ability to clearly indicate the formants makes it possible to adapt and / or select speech units that are suitable in concatenated speech synthesis.

In the low-pass component of the previous conversion step (when the original signal has 256 frequency values: TP 7), the degree of wrinkling of the voice can be obtained. The drop in the curve course between the maximum value Mx and the minimum value Mi characterizes the degree of wrinkling.

Thus, the above three unique features of the speaker can be identified and modified appropriately for speech synthesis. Of particular importance here is that, during the inverse wavelet transform, the manipulation of individual features specific to the speaker changes only those features and leaves other perception-related parameters intact. Therefore, the fundamental frequency can be adjusted as desired, without changing the degree of wrinkling of the voice.

A feature of another use case is that it is possible to select an advantageous sound part which is linked to and combined with another sound part, where both of these sound parts are originally separated from different speakers by different contexts. It was recorded in. Once the spectral audio features have been determined, it is possible to find advantageous sound parts to be combined. This is because these features make the evaluation criterion well known, and this evaluation criterion enables the comparison of the sound parts with one another and thus the selection of the matching sound parts automatically according to certain criteria.

FIG. 5 shows the steps of concatenated speech synthesis. The databank is composed of a preset set of sounds naturally uttered by various speakers, wherein the sound parts of the naturally uttered sounds are identified and stored. A large number of samples occur for different sound parts of one voice, and the data bank can access them. A sound part is, for example, a voice or a phoneme or a sequence of such phonemes. The smaller the sound part, the greater the potential for constructing new words. German contains about 40 predetermined amounts of phonemes, which are sufficient for the synthesis of almost all words of this language. At this time, various acoustic contexts,
Consideration must be given to which words each phoneme occurs. What is important here is that the individual phonemes be properly inserted into the acoustic context so that discontinuities which are unnatural and very "synthetic" to the human hearing are avoided. As mentioned above, the sound parts are obtained from different speakers and therefore have different speaker-specific characteristics. It is important to minimize these discontinuities in order to synthesize expressions that act as naturally as possible. This can be done by adapting identifiable and modifiable speaker-specific features or by selecting a matching sound segment from a data bank. Again, the speaker-specific features are an important aid in the selection.

FIG. 5 shows an example of two sounds A 507 and B 508, each of which has a separate sound part 505 or 506. Sounds A 507 and B 5
08 are obtained from the uttered expressions, and the sound A 507 and the sound B
508 is clearly different. Separation line 509 consists of sound A 507 and sound B
508 indicates the location where it must be joined. In this case, the first three sound parts of the sound A 507 and the last three sound parts of the sound B 508 must be connected and connected.

A continuous sound portion along the dividing line 509 is temporally extended or compressed (arrow 50).
3 (see FIG. 3) to avoid the discontinuity in the transition 509.

A feature of the modification is that the sound divided along the dividing line 509 elapses sharply. In this case, the discontinuity described above, which is perceived as a disturbance in human hearing, occurs. On the other hand, the sound C in the transition region 501 or 502 is taken into account, and the transition region 501 or 502 and the spectral interval measure between the sound portions that can be associated with each other are matched. (Elapses gradually between sound parts). The Euclidean distance between the coefficients associated with this region, for example in wavelet space, is used as this interval measure.

References [1] I. Daubechies: "Ten Lectures on Wavelets", Siam Verlag 1992, ISBN 0
-89871-274-2, Chapter 5.1, pp. 129-137

[Brief description of the drawings]

FIG. 1 is a diagram showing a wavelet function.

FIG. 2 is a diagram showing a wavelet function divided into a real part and an imaginary part.

FIG. 3 is a diagram showing a cascade-connected filter structure representing a transform step of a wavelet transform.

FIG. 4 is a diagram showing a low-pass component and a high-pass component of different conversion steps.

FIG. 5 is a diagram showing steps of concatenated speech synthesis.

[Procedural Amendment] Submission of translation of Article 34 Amendment of the Patent Cooperation Treaty

[Submission date] July 7, 2000 (200.7.7)

[Procedure amendment 1]

[Document name to be amended] Statement

[Correction target item name] 0003

[Correction method] Change

[Correction contents]

The wavelet transform is known from [1]. In the wavelet transform, the wavelet filter ensures that the high-pass and low-pass components of each successive transform step completely restore the signal of the current transform step. Here, the resolution of the high-pass component or the low-pass component is reduced in one conversion step to the next conversion step (in English terminology, “subsampling”).
Is). In particular, due to this subsampling, the number of conversion steps is finite. US-A-5528725 describes a speech recognition method using wavelet transform. EP-A-0519802 describes a speech synthesis method that adapts the unique characteristics of the speaker in consideration of the arrangement of naturally sounding speech sounds.

Claims (10)

[Claims] 1. A method for determining spectral speech characteristics of an uttered expression, comprising: a) digitizing the expression; b) performing a wavelet transform on the digitized expression; c) a phase of the wavelet transform. A method for determining spectral features of a voice, wherein the unique features of the speaker are determined using different transformation steps. 2. The method according to claim 1, wherein a windowing transformation of the digitized representation to the frequency domain is performed before the wavelet transformation. 3. The method according to claim 2, wherein the conversion to the frequency domain is performed by a fast Fourier transform. 4. The method according to claim 1, wherein in each step of the wavelet transform, a low-pass component and a high-pass component of a signal to be transformed are obtained. 5. The method according to claim 1, wherein the high-pass component is divided into a real part and an imaginary part. 6. The method according to claim 1, wherein the wavelet transform comprises a plurality of transform steps, wherein the last transform step supplies the expressed DC component by repeated low-pass filtering corresponding to the number of transform steps. The method according to any one of the preceding claims. 7. The speaker-specific features are determined by: a) the fundamental frequency of the uttered expression, b) the spectral envelope, and c) the degree of wrinkling of the uttered expression. A method according to any one of the preceding claims. 8. Use of the method according to claim 1, wherein individual features specific to the speaker are adapted taking into account the naturally audible sequence of the audio sound. . 9. A speech sound that guarantees a natural arrangement of speech sounds that are heard from a preset data set is selected based on individual spectral speech characteristics. How to use the described speech synthesis method. 10. An apparatus for determining spectral audio characteristics of an uttered expression comprising a processor unit, the apparatus comprising: a) digitizing said expression; and b) wavelet transforming said digitized expression. C) performing a step of determining a speaker-specific feature using different transform steps of a wavelet transform, wherein the apparatus determines spectral utterance features of the uttered expression.