The present invention relates generally to apparatuses and methods for improving speech quality in e.g. IP-telephony systems. More particularly, the present invention relates to a method and apparatus for reducing audio artifacts due to overrun or underrun in a playout buffer.
The invention also relates to an arrangement for carrying out the method.
When sampling frequencies, in e.g. a speech coding system, are not controlled, underrun or overrun might occur in the playout buffer, which is a buffer storing speech samples for later playout. Underrun means that the playout buffer will run into starvation, i.e. it will no longer have any samples to play on the output. Overrun means that the playout buffer will be filled with samples and that following samples cannot be buffered and consequently will be lost. Underrun is probably more common than overrun since the size of the playout buffer can increase until there is no memory left, while it only can decrease until there are no samples left.
Currently, most systems do not deal with the problem that the sampling frequency might differ considerably between the sending and the receiving side. One possible solution proposed in, EP-0680033 A2, works on pitch periods. Adding or removing pitch periods in the speech signal achieves a different duration of a speech segment without affecting other speech characteristics other than speed. This proposed solution might be used as an indirect sample rate conversion method.
Another solution uses the beginning of talkspurts as an indication to reset the playout buffer to a specified level. The distance, in number of samples, between two consecutive talkspurts is increased if the receiving side is playing faster than the sending side and decreased if the receiving side is playing slower than the sending side. In IP-telephony solutions using the IP/UDP/RTP-protocols (Internet Protocol/User Datagram Protocol/Real Time Protocol), a marker flag in the RTP header is used to identify the beginning of a talkspurt. At the beginning of a talkspurt, the playout buffer is set to a suitable size.
The solution according to EP-0680033 A2, where pitch periods are removed or inserted, assumes a fixed conversion factor between the receiving and transmitting side. Therefore, it cannot be used in dynamic systems, i.e. where the sampling frequencies varies. Further, it does not solve the problem with underrun or overrun situations, but is instead focused on changing the playback rate of a speech signal stored in compressed form for playback later and at a different speed to that at which it was stored.
Using the method of resetting the playout buffer to a certain size causes problems if there are very long talkspurts, e.g. broadcast from one speaker to several listeners. Since the length of a talkspurt is not defined in the beginning of the talkspurt, the size to reset to might be either too small or too large. If it is too small, underrun will occur and if it is too large, unnecessary delay is introduced. Thus, the problem persists.
The general problem with the currently known approaches is that they are static and inflexible. Therefore, dynamic solutions are required.
The present invention deals with the problem of improving speech quality in systems where the sampling rate at a transmitting terminal differs from the playout rate of a receiving buffer at a receiving terminal. This is often the case in e.g. IP-telephony.
When sampling frequencies are not controlled, underrun or overrun might occur in the playout buffer at the receiving side, which causes audible artifacts in the speech signal. To avoid said overrun or underrun there is an need for dynamically keeping the playout buffer to an average size, i.e. controlling the fullness of the playout buffer.
One object of the present invention is thus to provide a method for reducing audio artifacts in a speech signal due to overrun or underrun in the playout buffer.
Another object of the invention is to dynamically control the fullness of the playout buffer so as not to introduce extra delay.
The above mentioned and other objects are achieved by means of dynamic sample rate and conversion of speech frames, i.e. converting speech frames comprising N samples to instead comprise either N+1 or N−1 samples. More specifically, the invention works on an LPC-residual of the speech frame. By adding or removing a sample in the LPC-residual, a sample rate conversion will be achieved. The LPC-residual is the output from an LPC-filter, which removes the short-term correlation from the speech signal. The LPC-filter is a linear predictive coding filter where each sample is predicted as a linear combination of previous samples.
By using the proposed sample rate conversion method, the playout buffer, of e.g. an IP-telephony terminal, can be continuously controlled with only small audio artifacts. Since the method works on a sample-by-sample basis, the playout buffer can be kept to a minimum and hence no extra delay is introduced. The solution also has very low complexity, especially when the LPC-residual already is available, as in the case in e.g. a speech decoder.
The term “comprises/comprising” when used in this specification is taken to specify the presence of stated features, integers, steps or components but does not preclude the presence or addition of one or more other features, integers, steps, components or groups thereof.
Although aspects of the invention have been summarised above, the method and apparatus according to the appended claims define the scope of the invention.
Referring to
The above and other problems with underrun and overrun are solved by using dynamic sample rate conversion based on modifying the LPC-residual of the speech signal and will be further described with reference to
The LPC-filter is usually denoted:
By feeding a speech frame through the LPC-filter, H(z), the LPC-residual is found. The LPC-residual, shown in
F(z)=b·z−L
When the LPC-residual is fed through the inverse of the LTP-filter F(z), and LTP-residual is created. In the LTP-residual, the long-term correlation in the LPC-residual is removed, giving the LTP-residual a noise-like appearance.
The solution according to the invention modifies the LPC-residual, shown in
A sample rate conversion consists of four modules, shown in FIG. 4.
1) A Sample Rate Controller (SRC) module 400 that calculates whether a sample should be added or removed;
2) LPC-Residual Extraction (LRE) modules 410 that are used to obtain the LPC-residual rLPC;
3) Sample Rate Conversion Methods (RCM) modules 420 that find the position at which to add or remove samples and determine how to perform the insertion and deletion, i.e. converting the LPC residual block rLPC comprising N samples to a modified LPC-residual block rLPC comprising N+1 or N−1 samples; and
4) A Speech Synthesiser Module (SSM) 430 to reproduce the speech.
An idea behind embodiments of the invention is that it is possible to change the playout rate of the playout buffer 440 by removing or adding samples in the LPC-residual rLPC.
The SRC module 400 decides whether samples should be added or removed in the LPC residual rLPC. This is done on the basis of at least one of the four following parameters: the sampling frequencies of the sending TRX1 and receiving terminal units TRX2, information about the speech signal e.g. a voice activity detector signal, status of the playout buffer, an indicator of the beginning of a talkspurt. The four parameters are designated SRC Inputs in FIG. 4. On the basis of a function of one or several of these parameters the SRC 400 decides when to insert or remove a sample in the LPC residual rLPC and optionally which RCM 420 to use. Since digital processing of speech signals usually is made on a frame-by-frame basis, the decision of when to remove or add samples basically is to decide within which LPC-residual rLPC frame the ROM 420 is to insert or remove a sample.
There are basically three methods of obtaining the LPC-residual rLPC that is needed as input to the RCM's 420. The methods depend on the implementation of the speech encoder and will be described with reference to
In
As can be seen in
If the speech encoder, on the other hand, is an analysis-by-synthesis speech encoder where the LTP-filter 540 is exchanged to an adaptive codebook 590 as shown in
When the speech encoder has some sort of backward adaptation, it is not feasible to make alterations in the LPC-residual since this would affect the adaptation process in a detrimental way. In
In
Referring again to
The first and most primitive method arbitrarily removes or adds a sample whenever this becomes necessary. If the sample rate difference between the terminals is small this will only lead to mirror artifacts since the adding or removing is performed very seldom.
By inserting or removing samples at positions where the energy in the LPC-residual is low the synthesis will be less affected. This is due to the fact that segments close to pitch pulses will then be avoided. To find these segments of low energy either a sliding window method or a simplier block energy analysis can be used.
The second method, called the sliding window energy method, calculates a weighted energy value for each sample in the LPC-residual. This is done by multiplying k samples surrounding a sample with a window function of size k (k<<N), where N equals the number of samples in the LPC-residual. Each sample is then squared and the sum of the resulting k values is calculated. The window is shifted one position and the procedure is repeated. The position where to insert or remove samples is given by the sample with the lowest weighted energy value.
The third method, block energy analysis, is a simpler solution for finding the insertion or deletion point. The LPC-residual is simply divided into blocks of equal length and an arbitrary sample is removed or added in the block with the lowest energy.
The fourth method, illustrated in
When the RCM-module 420 has calculated the position at which to add or remove a sample it must be determined how to perform the insertion or deletion. There are three methods of performing such insertions or deletion depending on the type of LRE-module used.
In the first method, either zeros are added or samples with small amplitudes are removed. This method can be used for all LRE solutions described above. (See
In the second method, insertion is carried out by adding zeros and interpolating surrounding samples. Deletion is performed by removing samples and preferably smoothing surrounding samples. This method can also be used for all of the LRE solutions described above. (See FIGS. 5C-5F). Notice that in
In the third method, the SCR/RCM-modules 545 are placed within the feedback loop of the speech decoder instead of after the feedback loop as in the previous methods. (See
For the speech decoder with LTP filter (see
The alterations on the LPC residual consists of removing or adding samples just before, but since the SRC/RCM-modules 545 are placed within the LTP feedback loop, some modifications must be done. The extending or shortening of a segment can be done in three ways either at the respective ends of the segment or somewhere in the middle of the segment.
For all implementations except when the SRC/RCM-modules 545 are placed between the fixed codebook 530 and the LTP filter 540 the history of the LPC residual also has to be modified. The lag L will be increased or decreased for the specific part of the history where a sample is inserted or deleted. Thus the starting position of the segment that will be copied from the history of the LPC residual, Pointer 1 or Pointer 2 in
When the SRC/RCM-modules are placed before the summation of the outputs from the adaptive and the fixed codebook as shown in
Embodiments of the invention being thus described, it will be obvious that the same may be varied in many ways. Such variations are not to be regarded as a departure from the scope of the invention, and all such modifications as would be obvious to a person skilled in the art are intended to be included within the scope of the following claims.
Number | Date | Country | Kind |
---|---|---|---|
9903223 | Sep 1999 | SE | national |
Number | Name | Date | Kind |
---|---|---|---|
4435832 | Asada | Mar 1984 | A |
4864620 | Bialick | Sep 1989 | A |
5067158 | Arjmand | Nov 1991 | A |
5073938 | Galand | Dec 1991 | A |
5426718 | Funaki | Jun 1995 | A |
5517521 | Strawn | May 1996 | A |
5598466 | Graumann | Jan 1997 | A |
5617507 | Lee | Apr 1997 | A |
5694521 | Shlomot | Dec 1997 | A |
5717823 | Kleijn | Feb 1998 | A |
5825771 | Cohen et al. | Oct 1998 | A |
6029134 | Nishiguchi | Feb 2000 | A |
6073092 | Kwon | Jun 2000 | A |
6311154 | Gersho | Oct 2001 | B1 |
Number | Date | Country |
---|---|---|
0680033 | Apr 1995 | EP |
0 680 033 | Nov 1995 | EP |
0743773 | Nov 1996 | EP |