EP1078354B1 - Method and device for determining spectral voice characteristics in a spoken expression - 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

The invention relates to a method and an arrangement for Determination of spectral speech characteristics in one spoken utterance.

In a concatenative speech synthesis, individual sounds composed of language databases. To get one for that human ear naturally sounding speech history too are discontinuities at the points where the sounds are to be put together (concatenation points) to avoid. The sounds are in particular phonemes of a language or a composition of several phonemes.

A wavelet transformation is known from [1]. In the Wavelet transformation is through a wavelet filter ensures that a high pass share and a Low-pass portion of a subsequent transformation stage Signal of a current transformation stage complete restore. It is done by a Transformation stage to the next a reduction of Dissolution of the high-pass component or low-pass component Technical term: "Subsampling"). In particular, by that Subsampling the number of transformation levels finally.

US-A-5528725 discloses a method for speech recognition using wavelet transformations.

EP-A-0519802 discloses a method of speech synthesis which speaker-specific characteristics with regard to a a natural sounding sequence of speech sounds adapts.

The object of the invention is a method and an arrangement for determining spectral Specify language characteristics, with the help in particular a natural-looking synthetic speech output can be determined is.

This task is carried out according to the characteristics of the independent Claims resolved.

Within the scope of the invention, a method is specified for Determination of spectral speech characteristics in one spoken utterance. This becomes the spoken utterance digitized and subjected to a wavelet transformation. Using different transformation levels of the wavelet transformation become the speaker-specific Characteristics determined.

It is a particular advantage that in the wavelet transformation by means of a high-pass filter and one Low pass filter the utterance is split and different high-pass components or low-pass components different transformation levels speaker-specific Characteristics included.

The individual high-pass components or low-pass components Different transformation levels stand for predefined ones speaker-specific characteristics, both High pass component as well as low pass component of a respective one Transformation level, i.e. the respective characteristic, can be modified separately from other characteristics. If you use the inverse wavelet transformation from the respective high-pass and low-pass proportions of the individual Transformation levels back to the original signal together, this ensures that exactly what you want Characteristic has been changed. It is therefore possible to change certain predetermined characteristics of the utterance, without affecting the rest of the utterance.

One embodiment is that before the wavelet transformation the utterance windowed, that is, a given one Set of samples cut out, and in the Frequency range is transformed. This will in particular a Fast Fourier Transform (FFT) is applied.

Another embodiment is that a High-pass component of a transformation stage in a real part and split an imaginary part. The high pass portion of the Wavelet transformation corresponds to the difference signal between the current low-pass component and the low-pass component the previous transformation level.

In particular, further training consists in the number of transformation stages of the wavelet transformation to be carried out by determining that in the last Transformation level, which consists of cascaded Low passports exist, contain a steady portion of the utterance is. Then the signal as a whole can be represented by its Wavelet coefficients. This corresponds to the complete one Transformation of the information of the signal section in the Wavelet space.

In particular, only the respective low-pass portion continues transformed (using a high pass and a Low-pass filter), one remains as a high-pass component Transformation stage the difference signal, as explained above. Cumulative difference signals (high-pass components) over the Transformation levels, you get in the last Transformation level as a cumulative high-pass component Information of the spoken utterance without a constant component.

a) Grundfrequenz:
Die Schwingung des Hochpaßanteils der ersten oder der zweiten Transformationsstufe der Wavelet-Transformation läßt die Grundfrequenz der Äußerung erkennen. Die Grundfrequenz zeigt an, ob der Sprecher ein Mann oder einen Frau ist.

b) Form der spektralen Hüllkurve:
Die spektrale Hüllkurve enthält Information über eine Transferfunktion des Vokaltrakts bei der Artikulation. In einem stimmhaften Bereich wird die spektrale Hüllkurve von den Formanten dominiert. Der Hochpaßanteil einer höheren Transformationsstufe der Wavelet-Transformation enthält diese spektrale Hüllkurve.

c) Spectral Tilt (Rauchigkeit):
Die Rauchigkeit in einer Stimme wird als negative Steigung im Verlauf des vorletzten Tiefpaßanteils sichtbar.

As part of additional training, the speaker-specific characteristics can be identified as:

a) fundamental frequency:
The oscillation of the high-pass component of the first or the second transformation stage of the wavelet transformation reveals the fundamental frequency of the utterance. The basic frequency indicates whether the speaker is a man or a woman.

b) Shape of the spectral envelope:
The spectral envelope contains information about a transfer function of the vocal tract during articulation. In a voiced area, the spectral envelope is dominated by the formants. The high-pass component of a higher transformation level of the wavelet transformation contains this spectral envelope.

c) Spectral Tilt:
The smokiness in a voice becomes visible as a negative slope in the course of the penultimate low-pass portion.

The speaker-specific characteristics a) to c) are in speech synthesis of great importance. As at the beginning mentioned, one uses the concatenative Speech synthesis of large amounts of uttered utterances cut out the example sounds and later to a new one Word to be put together (synthesized language). there are discontinuities between compound sounds of Disadvantage as these are considered unnatural by the human ear be perceived. To counteract the discontinuities it is advantageous to directly add the perceptually relevant sizes record and if necessary to compare and / or adapt to each other.

This can be done through direct manipulation by a Speech in at least one of his speaker-specific Characteristics is adjusted so that it is in the acoustic Context of concatenated sounds not as disturbing is perceived. It is also possible to choose one to align the appropriate sound with the speaker-specific Characteristics of sounds to be linked as good as possible match each other, e.g. that the sounds are the same or similar Smokiness is inherent.

An advantage of the invention is that the spectral Envelope reflects the speaker's articulation tract and not, e.g. a pole position model, on formants is supported. Go further with the wavelet transformation as a nonparametric representation, no data is lost that The utterance can always be completely reconstructed. From the individual transformation levels of the wavelet transformation resulting data are linear from each other independently, can thus be influenced separately and later on to the influenced utterance - lossless - be put together.

Furthermore, an arrangement for determining spectral Speech characteristics indicated a processor unit has, which is set up such that an utterance can be digitized. Then the utterance becomes a Wavelet transform and subjected to different levels of transformation speaker-specific characteristics determined.

This arrangement is particularly suitable for implementation of the method according to the invention or one of its above explained further training.

Further developments of the invention also result from the dependent claims.

Exemplary embodiments of the invention are described below shown and explained in the drawing.

Fig.1

eine Wavelet-Funktion;

Fig.2

eine Wavelet-Funktion, unterteilt nach Realteil und Imaginärteil;

Fig.3

eine kaskadierte Filterstruktur, die die Transformationsschritte der Wavelet-Transformation darstellt;

Fig.4

Tiefpaßanteile und Hochpaßanteile unterschiedlicher Transformationsstufen;

Fig.5

Schritte der konkatenativen Sprachsynthese.

Show it

Fig.1

a wavelet function;

Fig.2

a wavelet function, divided into real part and imaginary part;

Figure 3

a cascaded filter structure that represents the transformation steps of the wavelet transformation;

Figure 4

Low-pass components and high-pass components of different transformation levels;

Figure 5

Steps of concatenative speech synthesis.

wobei

f

die Frequenz,

σ

eine Standardabweichung und

c

eine vorgegebene Normierungskonstante

bezeichnen. 1 shows a wavelet function which is determined by

in which

f

the frequency,

σ

a standard deviation and

c

a given standardization constant

describe.

In particular, the standard deviation σ is determined by the Predeterminable position of the sideband minimum 101 in FIG. 1.

2 shows a wavelet function with a real part according to equation (1) and a Hilbert transform H of the real part as an imaginary part. The complex wavelet function thus arises Ψ (f) = ψ (f) + jH {ψ (f)}

wobei Ψ die konjugiert komplexe Wavelet-Funktion bezeichnet. The constant c from equation (1) is used to normalize the complex wavelet function:

in which Ψ called the conjugate complex wavelet function.

3 shows the cascaded application of the wavelet transformation. A signal 301 is filtered both by a high pass HP1 302 and by a low pass TP1 305. In particular, subsampling takes place, ie the number of values to be stored is reduced per filter. An inverse wavelet transformation ensures that the original signal 301 can be reconstructed from the low-pass component TP1 305 and the high-pass component HP1 304.

In the high pass HP1 302 is separated according to real part Re1 303 and Imaginary part Im1 304 filtered.

The signal 310 after the low pass filter TP1 305 is again both by a high pass HP2 306 and by a Filtered low pass TP2 309. The HP2 306 high pass includes again a real part Re2 307 and an imaginary part Im2 308. Das Signal after the second transformation stage 311 is again filtered, etc.

If one goes with a (FFT-transformed) short-term spectrum 256 values, so there are eight transformation steps (subsampling rate: 1/2) until the signal from the last low-pass filter TP8 corresponds to the DC component.

4 shows various transformation stages of the wavelet transformation, divided into low-pass components (FIGS. 4A, 4C and 4E) and high-pass components (FIGS. 4B, 4D and 4F).

From the high-pass component according to FIG. 4B, the fundamental frequency is spoken utterance evident. In addition to the fluctuations in the amplitude is clearly a predominant periodicity in Wavelet-filtered spectrum to recognize the fundamental frequency of the speaker. On the basis of the fundamental frequency it is possible given utterances in speech synthesis each other adapt or use appropriate statements from a database to determine given utterances.

In the low-pass portion of Fig.4C are as pronounced minima and Maxima the formants of the speech signal section (the length of the speech signal section corresponds approximately to twice Fundamental frequency). The formants represent Resonance frequencies in the speaker's vocal tract. The clear one Representability of the formants enables an adjustment and / or selection of suitable sound modules at concatenative speech synthesis.

In the low-pass portion of the penultimate transformation stage (at 256 Frequency values in the original signal: TP7), the smokiness one vote. The descent of the curve between maximum Mx and minimum Mi denotes the degree of Smokiness.

The three speaker-specific characteristics mentioned are thus identified and targeted for speech synthesis to be influenced. It is particularly important that manipulation in the inverse wavelet transform of a single speaker-specific characteristic only this influences the other perceptually relevant variables stay untouched. Thus, the fundamental frequency can be targeted can be adjusted without changing the smokiness of the voice being affected.

Another possible application is the selection of one suitable sound section for concatenative linking with another sound section, both sound sections originally from different speakers in different Contexts were included. With determination spectral Speech characteristics can be a more suitable one to be linked Phonetic section can be found as with the characteristics Criteria are known that allow a comparison of Sound sections among themselves and thus a selection of the matching sound section automatically according to certain specifications enable.

5 shows steps of a concatenative speech synthesis. A database is created with a predetermined amount of naturally-spoken language from different speakers, sound sections in the naturally-spoken language being identified and stored. There are numerous representatives for the different sound sections of a language that the database can access. The sound sections are in particular phonemes of a language or a series of such phonemes. The smaller the section of the sound, the greater the possibilities when composing new words. For example, the German language contains a predetermined amount of approximately 40 phonemes, which are sufficient for the synthesis of almost all words in the language. Different acoustic contexts must be taken into account, depending on the word in which the respective phoneme appears. Now it is important to embed the individual phonemes in the acoustic context in such a way that discontinuities which are perceived by the human ear as unnatural and "synthetic" are avoided. As mentioned, the sound sections come from different speakers and thus have different speaker-specific characteristics. In order to synthesize a statement that looks as natural as possible, it is important to minimize the discontinuities. This can be done by adapting the identifiable and modifiable speaker-specific characteristics or by selecting suitable sound sections from the database, the speaker-specific characteristics also being a decisive aid in the selection.

5 shows two sounds A 507 and B 508 by way of example shown, each of the individual sound sections 505 and 506 exhibit. Lute A 507 and B 508 are from a spoken utterance, whereby the sound A 507 clearly is different from the sound B 508. A dividing line 509 indicates where the A 507 sound should be linked to the B 508 sound. In the present case, the first three sound sections should of the sound A 507 with the last three sound sections of the According to B 508 can be linked concatenatively.

There is a temporal stretching or along the dividing line 509 Upsetting (compare arrow 503) the successive Phonetic sections performed to the discontinuous Reduce impression at transition 509.

A variant consists in an abrupt transition along the the dividing line 509 divided sounds. However, this happens the discontinuities mentioned, which the human ear as distracting. If you put together a sound C, that the sound sections within a transition area 501 or 502 are taken into account, where a spectral Distance between two assignable Sound sections in the respective transition area 501 or 502 is adjusted (gradual transition between the According sections). As the distance measure is used especially in the wavelet space the Euclidean distance between the relevant coefficients in this area.

Bibliography:

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

Claims (10)

1. Method for determining spectral speech characteristics in a spoken utterance,

   a) in which the utterance is digitised

   b) in which the digitised utterance is subjected to a wavelet transformation, and

   c) in which the speaker-specific characteristics are determined with the aid of different transformation stages of the wavelet transformation.
2. Method according to Claim 1, in which a windowed transformation of the digitised utterance is carried out in a frequency band prior to the wavelet transformation.
3. Method according to Claim 2, in which the transformation is carried out in the frequency band by means of Fast Fourier Transformation.
4. Method according to one of the preceding claims, in which a low-pass component and a high-pass component of a signal to be transformed are determined in each stage of the wavelet transformation.
5. Method according to one of the preceding claims, in which a high-pass component is subdivided according to a real part and an imaginary part.
6. Method according to one of the preceding claims, in which the wavelet transformation comprises a plurality of transformation stages, the last transformation stage supplying a direct component of the utterance in repeated low-pass filtering corresponding to the number of transformation stages.
7. Method according to one of the preceding claims, in which the speaker-specific characteristics are determined by means of:

   a) a fundamental frequency of the spoken utterance;

   b) a spectral envelope; and

   c) a spectral tilt of the spoken utterance.
8. Use of the method according to one of Claims 1 to 7 for speech synthesis, in which individual speaker-specific characteristics are adapted with regard to a naturally sounding juxtaposition of speech sounds.
9. Use of the method according to one of Claims 1 to 7 for speech synthesis, in which those speech sounds which ensure a naturally sounding juxtaposition of speech sounds are selected from a prescribed data volume with the aid of individual spectral speech characteristics.
10. Arrangement for determining spectral speech characteristics in a spoken utterance, having a processor unit which is set up in such a way that the following steps can be carried out:

    a) the utterance is digitised;

    b) the digitised utterance is subjected to a wavelet transformation; and

    c) the speaker-specific characteristics are determined with the aid of different transformation stages of the wavelet transformation.

Images