EP1033702A2 - Method to separate speech and noises - Google Patents

Abstract

The process is used to separate out the speech signal from background noise and involves a transformation process from intensity v. time to amplitude v. frequency and frequency v. time. An analytical process is used to determine the speech interval and classification. Separation is based upon a comparison process.

Description

The invention relates to a method for separating speech and noises.

The object of the invention is to provide a method with which reliably separates speech and noise, if possible in real time.

This object is achieved with the method according to claim 1.

The report "Separation of language of non - language based on group theory ", as of Tuesday, February 29th, 2000, with the corresponding test diagrams is part of the description of the invention and is used Completion of the disclosure content.

The method according to the invention comprises the following steps:

Preparatory arithmetic steps Step 1:

The digitized speech signal to be freed from interference is in its individual sounds using an envelope method disassembled.

2nd step:

The speech signal from the Plot intensity over time in a plot Amplitude versus frequency and frequency versus time transformed.

3rd step:

In the frequency domain, a band is placed around the intensities. This will only compare all other intensities approved within this band. The size of the band depends among other things from the intensity or from the Volume down.

Step 4:

With the help of number theoretical functions, intervals of Language is calculated and a space of the equivalence classes of the Speech sounds are generated.

Separation of language from non-language 5th step:

The frequencies within a sound (envelope) are indicated by suitable loop methods compared in pairs. The division then results for each frequency pair Intervals. These intervals are calculated with the Intervals of the equivalence classes compared. Ales result This check shows whether the frequencies considered have harmonious relationships with each other or not. This harmonious relationships have speech frequencies with everyone other speech frequencies within an envelope or one Loud. Only if in the room of the equivalence classes Reference interval is found, the corresponding frequency transferred to the language file. If there is no reference interval the frequency is transferred to the non-language file.

Step 6 Separate language and non-language with the help of a Amplitude band in the frequency domain.

If the band around the amplitudes is made much smaller than 1 then the intervals of the equivalence classes interact which in turn are less than 1 with those to be compared Frequencies of the speech signal to be examined. This effect separates part of non-language from language.

7th step:

The ratio of the pitch difference is replaced by a logarithmic relationship between the amplitudes a new Relationship formed and for the separation of noise / speech used. For the result of this relationship, the procedure repeatedly, in the room of the equivalence classes Search reference interval.

Step 8: Separate language and non-language with the help of Limit of symmetry

If the frequency ratio becomes greater than 8, then one approaches the symmetry limit for harmony or consonance. If Removed frequencies whose intervals are larger than a certain one Become a factor, then it is non-language.

The individual separation processes (steps 5 to 8) can be carried out by Appropriate group-theoretical procedures linked together and be optimized. They can also be used individually become.

Advantageous embodiments of the method according to the invention are described below.

The voiced speech sounds of our speech organ are those Harmonious sounds. The harmonious family relationships between the partials make the unchangeable (invariant) properties of the sounds. In the learning process, which from the earliest age to speech recognition enables people to learn the invariants with their hearing Sound characteristics. Because it enables information-bearing from non-information-bearing To distinguish sound characteristics.

Calculate the envelope curve of a sound (envelope curve)

In order to be able to process the speech signals in real time, it is required that the signal be sounded sound by sound in turn after can be processed. For an arbitrary cut to prevent the signal from being processed is a individual flexible process needed. For this, the Envelope calculated because it provides the start, end and the duration of the sound.

The digitized speech signal can be converted using the Differential calculation first the base points of the envelopes won, which include all individual sounds, indifferent how badly the signal is disturbed. All constrictions in the signal are defined as the start and end points of the envelope. The signal can then be processed sound by sound.

Differentiating the envelope leads to the minima of Envelope. The minima of the envelope form the start and End points of the envelopes. The envelopes of the sounds included different frequencies and are therefore temporal different lengths.

Transformation of the speech signal with integral transformations

Fourier transform:
Representation of the speech signal in the frequency domain: amplitude as ordinate and frequency as abscissa

Hilbert transformation:
Representation of the speech signal in the frequency domain: frequency as ordinate and time as abscissa

Each wave packet (each sound) contains n summands. Every nth Summand is the nth partial oscillation from the wave packet.

The original signal is like this with integral transformations transforms it as an amplitude over frequency, and as Frequency over time for each individual's data set Envelope is available for further processing.

In the Fourier space, for example, are the amplitudes or the intensities for an envelope plotted against frequency. Here with one Amplitude band within the frequency space around the intensities a comparison of all other intensities only within this Band approved. The size of the amplitude band depends on the Intensity or volume.

All frequencies belonging to the intensities are in the Functional space of the Hilbert transformation as frequencies over the time is available.

Unfolding of the signal with integral transformations

In the digitized speech signal as the output signal is the Information as intensity over time in the form of frequencies encoded in a two-dimensional field.

The integral transformations are called linear operators of the Group theory defined and for the unfolding of the signal used. For example, Fourier and Hilbert transformation used. They transform the signal into one required for subsequent processing State in functional spaces, for example in the frequency space.

Tones and sounds

A sound system is an ordered pair of elements. The frequency interval is a group and the intervals are an injective mapping of the sound system into the set of all positive real numbers.
A finite subset of the frequencies is called sound.
An ordered n-tuple of tones is called an ordered n-sound

Order the speech signal by congruence groups

Number theoretical functions from prime numbers and powers of two form fields for equivalence classes

The equivalence classes become the speech signal in the form of Equivalence relations related.

Algebraic congruence groups arrange the speech signal in Language and non-language in different files.

Recovery of the voice signal in the time domain Integral inverse transformations

Equivalence relations

Element swaps are specified by a two or more digit relation on the set of all sounds of a sound system.
The assignment to speech or non-speech applies to an element from the set of the speech signal if and only if there is a reference tone for the intervals of the signal to be checked from the equivalence class of the number theoretical functions of the prime number field.

Calculation of the intervals from the frequencies of the speech signal

All frequencies of the frequency space of the transformed Voice signals must be checked to see if they Symmetry relationships with other frequencies of the same Have envelope. In the symmetry relationships of the elements there are encrypted invariant properties of the language.

Of the signal to be processed are therefore within the Envelope all frequencies in pairs with all divided other frequencies to form intervals. This happens until each frequency in the continuous Sequence through all other frequencies in the same envelope has been divided to temporarily become one Ratio number (interval) of this frequency pair to come.

The first frequency is through the second, the third and so divided until the end of the envelope. Then the second one divided by the third, the fourth and so on. In a Envelope can, for example, thousands of frequencies his. The order of the frequencies must never be interchanged become.

The separation of language and non-language becomes a consequence performed by several symmetry operations.

1. Equivalence classes with reference interval as Decision criterion

After each interval formation within the frequency range of the checking speech signal must have a reference interval from the Equivalence classes of the calculated intervals can be found.

The equivalence classes are in turn calculated using various number-theoretical functions. The separating effect becomes significantly better if one demands the fulfillment of a reference interval for every number-theoretical function. If these are n functions (where n = natural integers), then these number theoretical functions span an n-dimensional metric space.
If the interval obtained from the signal to be tested finds exactly one reference coordinate in all number-theoretical functions, then all number-theoretical functions are fulfilled at least once. The coordinates lead to exactly one reference interval from the space of the equivalence classes.

You can then use this reference tone to call other speech parameters such as the sound's decay time and decay time be queried if this n-dimensional space is a metric is imprinted. This can be done with the help of group theory on the Basis of matrix operations are performed.

This reference interval can be interpreted as a reference tone. Then many more parameters for speech recognition are used, such as the attack time and the Cooldown of a tone.

If there is a reference interval, then the calculation procedure must memorize these elements until all the elements are together were compared. Only after completing this Sorting process, the process knows which elements to use have other interval relationships.

All the frequencies over the intervals relationships after all, stand with everyone else Frequencies within the same envelope and thus the same sound of language in relation. The process notices yourself all these elements. At the end of the calculation process all such items are stored in the memory for speech. All other frequencies that are not in this relationship system Get involved are unrelated elements and come in the memory for non-language. Becomes part of the non-language hereby already separated from the language.

Elements that are related to each other through symmetrical relationships Intervals connected are elements of language. elements those in no symmetrical context with everyone else Elements within the envelope are not a language. They do not transmit any information.

2. Improved separation of language and non-language by one Linking the pitch difference with the intensities

A two-digit relation of the intervals is defined that a reference tone within several coordinates of the Equivalence class space the element from the one to be examined Signal corresponds exactly.

The elements of the existing set of factors Frequency intervals from the equivalence classes are now as Harmonies defined.

V = Intervall aus den Äquivalenzklassen aller zahlentheoretischen Funktionen;

f1/f2 = Verhältnis der Frequenzen = Tonhöhenabstand, (A1/A2) = Verhältnis der Intensitäten

In the following formula, the pitch difference is related to the logarithmic ratio of the amplitudes. V = exp (abs ((factor x (ln (f1 / f2)) / (factor x ln (A1 / A2)))); f1 / f2 = integer division it mean:

V = interval from the equivalence classes of all number theoretical functions;

f1 / f2 = ratio of frequencies = pitch spacing, (A1 / A2) = ratio of intensities

A reference tone is defined using this relationship.

Then the previous procedure is repeated.

This group operation improves the quality of the separation Language of non-language.

3. Separate language and non-language with the help of a Amplitude band relationship in the frequency domain

If the band around the amplitudes is made much smaller than 1, then the intervals of the equivalence classes, which in turn are smaller than 1, interact with the frequencies of the speech signal to be examined to be compared.
For example, the maximum of the first wave packet is 500 units. In order to compare all the amplitudes, the amplitude band with 50 units of bandwidth across the entire frequency space is set as the input band for the processing program.

There is a new effect when this processing program a pre-program is used, in which only the bandwidth the amplitude band is set much smaller than 1, for example 0.05 instead of 50.

Then this procedure separates language from non-language with the help of the discussed method of equivalence classes, exactly the opposite. The memory allocation must then be reversed since this procedure is reversed.
This effect separates part of the non-language from language. Because the procedure is very fast, it is a good idea to start with it. Since any process that removes non-language speeds up the work of subsequent processes, you have to start with the fastest process.

4. Separate language and non-language using the Limit of symmetry

If the frequency ratio becomes too large, for example larger than 8, then one approaches the symmetry limit for harmonic sound or consonance. If you remove frequencies, their intervals become larger than a certain factor, then it is about non-language.

Quality of separation

The more correctly calculated intervals in a group are around so the separation process gets better.

If intervals are double or multiple in a group, then the separation process deteriorates.

The more different number theoretical functions like this be used so that each is in its own group the better the separation process.

If different number theoretical functions in a group are summarized, then the deteriorates Separation process.

If the intervals are not suitable, for example number theoretical functions are then calculated the separation process deteriorates.

Existence of intervals

The existence of music theory has been known for centuries of intervals. Kepler was already looking for general ones Laws of harmony. He was looking for world harmony laws. Next Many, however, should be mentioned in particular Leonhard Euler, who Relationship between the decomposition of natural numbers into Products of prime powers, the Fermat theorems and the Number theory with number theoretical functions recognized.

The mathematician Leonhard Euler presented a complicated one number theoretical function (Euler grading function) for Calculation of such intervals and calculated them for the Numbers 1 to 10. That is 92 intervals for the Music sounds.

Leibnitz and Herder said that our soul counts unconsciously.

For the method according to the invention for speech recognition, the Euler's degree function is particularly advantageous as number theory function can be applied.

There are also number theoretical functions. Stand behind it complicated structures and facts of number theory and algebra, such as the gamma function.

The intervals are with number theoretical functions calculated. It arises from n number theoretical functions n-dimensional metric space that comes from the fields of Equivalence classes of symmetry relationships exist. There are n number theoretical functions, for example 10, used to calculate the intervals. This n number-theoretical functions generate intervals that are too an n-dimensional metric space of the equivalence classes let it stretch. These functions may produce all intervals for all human languages.

The relationships between the elements becomes more important than that Elements themselves.

The interval groups of the language can e.g. by following number theoretical functions are generated: 1. Function is the Euler Gradus function

a₁, a₂, a_m müssen ganze Zahlen sein.

Beispiel: 1440 = 2⁵ · 3² · 5¹

Dabei ist P₁ = 2, P₂ = 3, P₃ = 5

A₁ = 5; a₂ = 2; a₃ = 1

Euler definierte den Gradus als G(n) = a1 P1 + a2 P2 + am Pm - (a1 + a2 + ... am - 1)

Beispiel mit Zahlen: G(1440) = 5 · 2 + 2 · 3 + 1 · 5 - (5 + 2 + 1 - 1) = 10 + 6 + 5 - 7 = 14 als Konsonanzgrad für 1440

Für die Intervallmaße muß einem beliebigen Intervall a : b eine Zahl n so zugeordnet werden, daß n sinngemäß als Grad von a : b betrachtet werden kann. Entsprechend muß verfahren werden, wenn eine Verhältniskette a : b : c oder a₁ : a₂ : ... a_n vorliegt, also ein Mehrklang.Each natural number n is composed of prime numbers P ₁ , P ₂ , ... P _n . Then a number n becomes: n = P 1 a1 · P 2nd a2 ·. . . P m at the

a ₁ , a ₂ , a _m must be integers.

Example: 1440 = 2 ⁵ · 3 ² · 5 ¹

P ₁ = 2, P ₂ = 3, P ₃ = 5

A ₁ = 5; a ₂ = 2; a ₃ = 1

Euler defined the degree as G (n) = a 1 P 1 + a 2nd P 2nd + a m P m - (a 1 + a 2nd + ... a m - 1)

Example with numbers: G (1440) = 5 · 2 + 2 · 3 + 1 · 5 - (5 + 2 + 1 - 1) = 10 + 6 + 5 - 7 = 14 as a degree of consonance for 1440

For the interval measurements, a number n must be assigned to any interval a: b in such a way that n can be regarded as the degree of a: b. The same procedure must be followed if there is a ratio chain a: b: c or a ₁ : a ₂ : ... a _n , i.e. a multiple sound.

n must itself be a function of the system (a ₁ , a ₂ , ... a _n ). This value of a consonance is the quotient of the smallest common multiple and the largest common divisor of the natural numbers describing a sound.

The degree function for shortened fractions a / b results from G (a / b) = G (abb)

2. Function

2nd n  / (3rd x  · 5 y  · 7th e.g. ) (3rd x  · 5 y  · 7th e.g. ) / 2 n

3. Function

(n + 2) / n n / (n + 2)

4. Function

(2n + 2) / n n / (2n + 2)

5. Function

3rd x  5th y 5 y  / 3rd x

6. Function

2nd n  / 3rd x 3rd x  / 2nd n

7. Function

2nd n  5th y 5 y  / 2nd n

8. Function

3rd x  / 7th e.g. 7 e.g.  / 3rd x

9. Function

2nd n  / 7th e.g. 7 e.g.  / 2nd n

10. Function

7 e.g. / 5 y 5 y / 7th e.g.

Claims (1)

Process for the separation of speech and noise, with the following process steps: The digitized speech signal to be freed from interference is broken down into its individual sounds using an envelope method. The speech signal is transformed from the intensity versus time representation to an amplitude versus frequency and frequency versus time representation by means of integral transformations. In the frequency domain, a band is placed around the intensities. This allows a comparison of all other intensities only within this band. The size of the band depends, among other things, on the intensity or volume. Using number-theoretical functions, intervals of the language are calculated and a space of the equivalence classes of the speech tones is generated.
The actual separation of speech and noise is then carried out by one or more of the following process steps: The frequencies within a sound (envelope curve) are compared with one another in pairs using suitable loop methods. Intervals then arise for each frequency pair from the division. These intervals are compared with the calculated intervals of the equivalence classes. The result of this check is whether the considered frequencies have harmonious relationships with each other or not. These harmonic relationships have speech frequencies with all other speech frequencies within an envelope or a sound. The corresponding frequency is only transferred to the language file if a reference interval is found in the room of the equivalence classes. If there is no reference interval, the frequency is transferred to the non-language file. Separating speech and non-speech using an amplitude band in the frequency domain:
If the band around the amplitudes is made much smaller than 1, then the intervals of the equivalence classes, which in turn are smaller than 1, interact with the frequencies of the speech signal to be examined to be compared. This effect separates part of the non-language from language. A logarithmic relationship between the amplitudes creates a new relationship from the ratio of the pitch difference and uses it for the separation of noise and speech. For the result of this relationship, the procedure is repeated to look for a reference interval in the space of the equivalence classes. Separating language and non-language using the symmetry boundary:
If the frequency ratio becomes greater than 8, then the symmetry limit for harmony or consonance is approached. Those frequencies are removed whose intervals are larger than a certain factor.

Images