EP1390946B1 - Method for estimating a codec parameter - Google Patents

Description

The invention relates to a method for estimating a Parameters occurring in the course of speech coding, in particular a filter coefficient, a gain factor, or a basic speech frequency.

• Die 10 Koeffizienten eines Filters, welches die spektrale Einhüllende des Sprachsignals im Bereich des aktuellen Rahmens repräsentiert, werden mit 26 Bit pro Rahmen quantisiert. Diese Koeffizienten werden auch Spektralkoeffizienten oder Spektralparameter genannt.
• Mittels 4x7 bit werden vier Subrahmen eines Anregungssignals für dieses Filter quantisiert.
• Mittels 2x8 bit und 2x5 bit werden vier Werte einer Sprachgrundfrequenz repräsentiert.
• Mittels 4x7 bit werden vier Verstärkungsfaktorpaare pro Rahmen vektorquantisiert.

In digital communication systems such as the Internet or mobile radio systems such as GSM or UMTS, source coding methods, for example voice, audio, image or video coding methods, are used in order to reduce the bit rate to be transmitted. The source coding methods usually provide a bit stream that is divided into frames. In the case of voice transmission in the GSM system, a frame of voice-coded bits represents 20 ms of the voice signal. The bits within a frame represent, among other things, a certain set of parameters. These parameters describe, for example, the spectral envelope of the speech signal, the basic speech frequency, or a signal energy or amplification.
A frame is again often divided into subframes, so that some parameters are transmitted once per frame, others once per subframe.
In the case of the US TDMA Enhanced Full Rate (EFR) speech codec with 7.4 kbps, a 20 ms frame contains 148 bits. A frame here consists of four subframes. The parameters here are:

• The 10 coefficients of a filter, which represents the spectral envelope of the speech signal in the area of the current frame, are quantized with 26 bits per frame. These coefficients are also called spectral coefficients or spectral parameters.
• Four subframes of an excitation signal for this filter are quantized using 4x7 bits.
• Using 2x8 bit and 2x5 bit, four values of a basic speech frequency are represented.
• Four pairs of amplification factors per frame are vector-quantized using 4x7 bits.

In summary, it can be said that the bits are within of a frame generally a certain set of parameters represent which one depends on the particular one used Source coding method is.

The digitized signal within a frame was up the transmission side through the so-called source coding redundancy withdrawn. On the receiving side, this is due to the source decoding, such as speech decoding, largely reversed made.

It can now happen that one or more consecutive Frame is lost or from a network component be marked as unusable. These frames, So-called "bad frames" cannot or should not be used then become. The source decoder, for example the speech decoder, must take measures on the receiving side, that such a loss of frame is not audible if possible or in the case of image or video transmissions is not visible.

Generally there is an indicator on the receiving side indicating whether a frame was received without errors, the so-called bad frame indicator (BFI). BFI = 0 means in following that one assumes that the received frame is correct, while BFI = 1 indicates an error, for example that no frame was received in time or a disturbed frame was received. Of course, bit errors, that is, the reversal of individual bits, within one Framework depending on the system conditions. These are supposed to but in the further either no differentiated treatment experienced on the reception side, or the appropriate framework is marked with BFI = 1.

So far, in the case of BFI = 1, the past has already been decoded speech signal, for example by correlation formation the current speech signal frame is estimated. alternative are known from the past Speech codec parameters are the parameters of the current frame estimate, and then operate the decoder in a similar manner leave as if these estimated parameter values were correct. These are usually extrapolative processes, only on the bits or parameter values already received To fall back on.

For voice transmission over the Internet, for example Voice over IP (VoIP), or for voice transmission over the Internet in connection with a mobile communication system (such as GSM or UMTS) is on at the receiving end Buffer memory required because packets received do not arrive in a fixed time grid, but with different ones Delay times arrive (delay jitter). Such a buffer memory can possibly have several frames to include length, causing too frequent frame losses at cost an increased transmission delay can be prevented can. However, it often happens that several successive frames are lost, but the subsequent one Frame is received correctly. In such cases an interpolation of the use of a buffer memory Speech codec parameters of the lost frames instead a conventional extrapolation advantageous because it Is generally more accurate. A simple solution would be a linear one Interpolation based on the parameter values of the last decoded frame (time t = n-1) and the parameter values of the correctly received frame (time t = m> n) over all m-n intervening lost frames (Times t = n, n + 1, ..., m-1). A buffer storage and therefore Parameter interpolation can also be used in streaming applications can be realized since they are usually not sensitive regarding the delay time, see e.g. EP-A-0,459,358th

However, this has the disadvantage that there are parameters that cannot simply be interpolated. These often include the amplification factors, the basic speech frequency values or the spectral parameters V_i (t) of a speech frame at time t because they are differentially coded. In the case of speech coding, a spectral parameter V_i (t) of a speech frame is, for example, the filter coefficient of the time-dependent, digital filter, with the aid of which the vocal tract is modeled:
Speech is encoded, for example, using the LPC principle (Linear Predictive Coding). Voiced sounds are generated in this case via a periodic sequence of pulses, unvoiced sounds, for example, by means of a random noise generator (random noise generator). Plosive sounds are simulated by changing the amplification, and the vocal tract is simulated using a time-varying digital filter. The coefficients of this varying digital filter are obtained with the help of linear prediction, that is to say a prediction of the following value on the basis of previous values.

Differential or predictive coding is understood an encoding of a parameter at a time n at which also values of the parameter before n Times to be involved.

A parameter in the sense of the following explanations can for example a gain factor, a basic speech frequency or be a spectral parameter. Usual forms of presentation spectral parameters are, for example, the filter coefficients itself (in so-called direct form), autocorrelation coefficients, Reflection coefficients or so-called Log-area ratios. A state-of-the-art presentation are for example the ISF (imittance spectral frequencies), LSF (line spectral frequencies) or LSP (line spectral pairs). For the sake of simplicity, a parameter in the following without restricting generality as a spectral coefficient accepted.

Differential coding and decoding of the parameter V_i (t) can be carried out, for example, in the following way: A differential signal X_i (t = n) is determined on the transmission side in accordance with: X_i (n) = V_i (n) -a_i * Q [X_i (n-1)], i = 1,2, ..., 10, where V_i (n) is the parameter to be encoded, a_i is a prediction coefficient, and Q [X_i (n-1)] is the quantized difference signal determined for the encoding of V_i (n-1) in the previous frame. So-called vector quantization is often used for quantization. This is the joint quantization of several X_i (n) for certain values of i. Vector quantization can also mean the joint quantization of two or more different parameter types that occur in a speech coding method. In the case described, vector quantization could look like this: i = 1,2,3, and i = 4,5,6 and i = 7,8,9,10. The quantized difference signal Q [X_i (n)], i = 1,2, ..., 10 is thus represented by a number of bits, for example 26 bits per frame, and transmitted.
It can be seen from equation (1) that such coding leads to data compression: the storage effort for the difference values X_i, which represent the difference of almost equally large numbers, is less than for the values of V_i.

At the receiving end, a quantized value W_i (n) of the spectral parameter V_i (n) is reconstructed from the currently received difference signal value Q [X_i (n)] and the previously received Q [X_i (n-1)]: W_i (n) = a_i * Q [X_i (n-1)] + Q [X_i (n)], i = 1,2, .., 10

The form of parameter decoding described here is common in many coding methods currently used, under other, for example, in the AMR and EFR speech coder (adaptive multi-rate or enhanced full-rate). in principle are of course also higher orders of prediction imaginable. Usually, the equations (1), (2) mentioned regulations for the reduced by the mean Parameter value carried out. The mean becomes Finally added as an addition of a constant.

A predictive coding as exemplified above indicates an interpolative determination of the spectral coefficients missing frame disadvantages on:

With first-order predictive quantization (see Equations (1) and (2)) is there for an interpolative determination the quantized parameter value W_i (n) required, that two successive values of the quantized Difference signals {Q [X_i (m)], Q [X_i (m + 1)]} are received, which is often not the case with packet-switched transmission methods the case is. This will be explained in more detail in the following illuminated; this is the quantized difference signal Q [X_i (n)] hereinafter referred to as quantity Y_i (n):

So the following applies: W_i (n) = a_i * Y_i (n-1) + Y_i (n), i = 1,2, .., 10.

It is assumed below that the last one, already according to Equation (3) heard decoded frames at time t = n-1, and that the frame t = n is currently being decoded should, but BFI (n) = 1 applies, ie a "bad" framework is present. Let the frame t = m> n be the first frame after t = n-1, for which BFI = 0 applies. The spectral coefficients all other m-n intermediate frames with BFI = 1 are now to be interpolated. The spectral coefficient W_i (n-1) now forms the lower one (that is, in the past support point of the interpolation. The spectral coefficient W_i (m) should normally be the top one (that is form the interpolation base in the future. However, it cannot be calculated with predictive coding are received, since the quantity Y_i (m) is received for equation (3) Y_i (m-1) but is missing on condition. Only after two consecutive correctly received frames m and m + 1 could be a spectral coefficient W_i (m + 1) = a_i * Y_i (m) + Y_i (m + 1) can be calculated and on the receiving side as a support point serve for an interpolation. In principle, however, this requires an additional delay of one frame, at least what a significant one for bidirectional voice transmission Represents problem, or two consecutive Frame with BFI = 0, which is particularly true for packet-switched Transmission modes is not always given.

With L-order prediction, the problem is aggravated according to the aspects mentioned above: The differential Decoding according to equation (2) requires L + 1 consecutive Variables or difference signals Y_i (t), that is, for interpolation the spectral coefficients of previous frames with BFI = 1 a number of L + 1 consecutive correct ones Frames are received to last in this frame again a completely error-free set of spectral coefficients and thus to obtain an upper interpolation point.

Even if for the reason in common speech coding methods error propagation often uses a linear prediction L = 1 is selected, it can be summarized that yes two consecutive correct frames are received need before getting a correct spectral coefficient again W_i (m + 1) receives. Statistically speaking, this is natural less likely than receiving a correct frame. This fact usually results in longer delay times, what real-time sensitive applications is intolerable.

The present invention is therefore based on the object specify a method with which codec parameters are received let determine, even if the underlying Data in single or multiple successive time periods absence.

This object is achieved by independent claim 1 solved. Further developments result from the dependent claims.

The invention relates to a method for receiving-side estimation of a time-variable parameter at an nth point in time. The parameter was coded predictively on the transmission side and is determined interpolatively on the reception side depending on at least two variables. One interpolation interpolation point, the first variable, forms an earlier value of the parameter that has already been decoded, and a second interpolation interpolation point, the second variable, is determined by extrapolative measures.
The interpolative determination of the parameter can be carried out by means of known interpolation measures, for example by means of linear interpolation between the first and second variable. In an embodiment variant, a weighted summation is also used for the interpolation.

The advantage of this method is that it is interpolated to determine the parameter can, as soon as the second size is known.

The invention further relates to a method for receiving Estimation of an assigned to an nth frame Codec parameter. The codec parameter is predictive on the transmission side encoded and is received at the receiving end as a function of at least two signals determined by interpolation. A The interpolation is supported by the previously decoded Parameters of the (n-1) th frame formed, another Support point is determined by the parameter of the mth frame with m > n formed, which is determined by extrapolative measures has been.

A further development is that an interpolation then takes place as soon as the data of a correct framework are available. This has the advantage of a short delay time simultaneous use of an interpolative measure for parameter estimation.

Another further training provides that the quality of the Receive is indicated by an indicator size. This indicator size can e.g. B. the "bad frame" indicator BFI.

Figur 1

die Simulationsergebnisse einer GSM-Vollratenkanal-Übertragung, wobei die Ergebnisse verschiedener Extrapolationsmethoden dargestellt sind.

The invention is explained in more detail below on the basis of a few exemplary embodiments.
In the further shows

Figure 1

the simulation results of a GSM full rate channel transmission, the results of various extrapolation methods being shown.

In one possible embodiment, they are differential encoded parameters a procedure which consists of There are two steps: First, the parameters of the Frames with poor reception, BFI = 1, extrapolatively estimated. On this basis, the first can now frames received correctly are decoded again. He then forms the basis for an interpolative reassessment of the Parameters of the previous frame with BFI = 1.

For each received frame with BFI = 1, that is, a frame that is not error-free, it is provided that the parameters are first extrapolated conventionally. This includes (at least in the last frame with BFI = 1 before a frame with BFI = 0) with differentially coded parameters a calculation of the quantized difference signal or the quantity Y "afterwards". In the example given at the beginning, this conventional procedure provides that after the extrapolative determination of W_i (n) in the context t = n according to equation (3), the quantity Y_i (n) is determined by converting equation (3): Y_i (n) = W_i (n) - a_i * Y_i (n-1), i = 1,2, ..., 10,

A difference signal is thus again present at the time t = n + 1 of the previous frame, i.e. Y_i (n), so that at any time can be decoded again using equation (3). Due to the (provisional) extrapolative approach, one can upper support point W_i (m) can be determined, if only for the frame m BFI (m) = 0 applies. No further correct framework is required. The interpolation of the m-n past The frame can take place directly at time t = m.

Because of the memory of differential coding the support point W_i (m) has errors. This mistake disappears completely only when receiving L consecutive Frame with BFI = 0. Own, to test this procedure however, simulations show that W_i (m) can be used as an upper support point, opposite one approximation substantially improved in the prior art to enable the parameters. The main advantage this method is that an error burst, that is, a Sequence of m-n bad frames, waiting for a single correct one Frame can be interpolated, even if differentially coded parameters are present. No additional Delay is necessary; moreover, that becomes statistical rarer case of L using consecutive frames BFI = 0 not required.

In a first exemplary embodiment, parameters with a first order prediction, that is L = 1 considered:

• Der Spektralkoeffizient W_i(n-1) sei bereits decodiert,
• Y_i(n-1) liege entweder empfangen [BFI(n-1)=0] oder nach Gleichung (4) rekonstruiert vor [BFI(n-1)=1].
• Als Resultat des nachfolgend genannten rekursiven Algorithmus' liegen auch Y_i(n), ..., Y_i(n+K-1) vor.
• Der aktuelle Zeitpunkt sei t = n+K, zu diesem Zeitpunkt solle der Spektralkoeffizient W_i(n) bestimmt werden.

Das heißt also, daß eine Zeitverzögerung von K Rahmen zur Interpolation erlaubt ist.
Das Vorgehen erfolgt nun in zwei Schritten:

a) Operationen am Rahmen n+K:

Falls BFI(n+K) = 0: Berechne W_i(n+K) nach Gleichung (3).

Falls BFI(n+K) = 1: Berechne eine vorläufige extrapolierte Version W_i(n+K) mit einem beliebigen extrapolativen Verfahren.

b) Decodiere den Rahmen n:

Falls BFI(n) = 0: Berechne W_i(n) nach Gleichung (3).

Falls BFI(n) = 1: Berechne m > n, wobei m der erste Rahmen mit BFI(m) = 0 nach dem Rahmen n ist.

Falls m > n+K: Berechne mit einem beliebigen Extrapolationsverfahren W_i(n).

Falls m <= n+K: Dann liegt für Rahmen m als korrekt empfangener Rahmen ja bereits ein vorläufig extrapolativ bestimmter Spektralkoeffizientwert W_i(m) vor. Er bildet die obere (oder zukünftige) Stützstelle für eine Interpolation des Parameters W_i(n). Die untere (oder zurückliegende) Stützstelle sei der Spektralkoeffizient W_i(n-1).

Man kann nun zum Beispiel eine lineare Interpolation durchführen. Dies geschieht unter Berücksichtigung der zeitlichen Abstände des Rahmens n zu den Stützstellen durch: W_i(n) = [W_i(n-1) - W_i(m)] * (m-n) / (m-n+1) + W_i(m). Die obere Stützstelle W_i(m) ist bereits vorläufig extrapolativ bestimmt, die untere Stützstelle W_i(n-1) bereits endgültig decodiert. The following assumptions are made:

• The spectral coefficient W_i (n-1) is already decoded,
• Y_i (n-1) is either received [BFI (n-1) = 0] or reconstructed according to equation (4) before [BFI (n-1) = 1].
• As a result of the recursive algorithm mentioned below, there are also Y_i (n), ..., Y_i (n + K-1).
• The current point in time is t = n + K, at this point the spectral coefficient W_i (n) should be determined.

This means that a time delay of K frames for interpolation is allowed.
The procedure now takes place in two steps:

a) Operations on frame n + K:

If BFI (n + K) = 0: Calculate W_i (n + K) according to equation (3).

If BFI (n + K) = 1: Calculate a preliminary extrapolated version W_i (n + K) using any extrapolative method.

b) Decode the frame n:

If BFI (n) = 0: Calculate W_i (n) according to equation (3).

If BFI (n) = 1: Calculate m> n, where m is the first frame with BFI (m) = 0 after frame n.

If m> n + K: Calculate W_i (n) with any extrapolation method.

If m <= n + K: Then, for frame m as the correctly received frame, there is already a provisionally extrapolative spectral coefficient value W_i (m). It forms the upper (or future) support point for an interpolation of the parameter W_i (n). The lower (or back) support point is the spectral coefficient W_i (n-1).

For example, you can now perform a linear interpolation. This takes into account the time intervals between the frame n and the support points by: W_i (n) = [W_i (n-1) - W_i (m)] * (mn) / (m-n + 1) + W_i (m). The upper interpolation point W_i (m) has already been provisionally determined extrapolatively, the lower interpolation point W_i (n-1) has already been finally decoded.

In Figure 1 is a simulation of a GSM full rate channel transmission with various C / I ratios (carrier-tointerferer ratio), which describe the channel quality, too see. For the curves, the spectral distortion (spectral distortion, SD), a common measure of quality for coding or transmission of spectral coefficients against the C / I ratio is plotted. The higher the SD, the more the reception-side speech quality is worse.

Curve 1 shows an extrapolation as used in previous decoding methods is used. Curves 2 to 5 show the results for the above embodiment depend on of size K, which is the maximum allowed time delay indicates in frame. Curve 2 shows a delay by one frame (K = 1), curve 3 a delay by two frames (K = 2), curve 4 a delay of three frames (K = 3) and Curve 5 shows a delay of four frames (K = 4).

You can see that with a delay of just one Frame (K = 1) is a huge profit, more than K = 2 future frames will not bring much additional profit. These simulation results are extremely beneficial for the Transfer of real-time sensitive applications, since here only a slight delay is allowed. At very low C / I ratios however, you can see slight differences for the different deceleration values (K = 1,2,3,4). The cause the reason for this is that with such a poor C / I ratio often several successive frames are "bad frames".

In addition to the examples explained above, there are many further embodiments within the scope of the invention defined by the claims, the not described further here. But you can based on the preceding statements by the specialist without to put a lot of effort into practice.

Claims (4)

1. Method for estimating, on the receiving side, the value of a temporally variable parameter at an nth time point,

   wherein the parameter is predictively encoded on the sending side and, as a result of this encoding, a quantised difference signal belonging to the time point n is formed between the value of the parameter at the nth time point and the quantised difference signal at the (n-1)th time point, said formed difference signal being transmitted,

   wherein an indicator is present on the receiving side and shows whether or not a correctly received quantised difference signal is present for the nth time point,

   wherein the parameter is determined on the receiving side as a function of at least two variables by means of interpolation if the associated difference signal has not been correctly received, and

   the first variable in this case represents a first interpolation node, which is formed by a value of the decoded parameter which is assigned to a time point prior to the nth time point, and

   wherein the second variable represents a further interpolation node, which is formed by a further value of the parameter at an mth time point after the nth time point, and this further value of the parameter is determined by means of extrapolation.
2. Method according to claim 1, wherein data belonging to the nth time point is transmitted in an nth frame and the temporally variable parameter is a codec parameter.
3. Method according to claim 1 or 2, wherein the quantised signal for the (n-1)th time point on the sending side is weighted with a prediction coefficient for forming the quantised difference signal for the nth time point.
4. Method according to one of the preceding claims, wherein an interpolation is carried out as soon as the data for only one correct frame is received.

Images