GB2590879A - Discrete dither - Google Patents

Abstract

A bit is buried in a pair of data samples by identifying two possible 2-tuples (pairs) of output values, where each pair conveys the same bit of data, and pseudo-randomly choosing one of the two possible pairs with equal probability (50%). The two possible 2-tuples may differ by exactly one step in both dimensions. The 2-tuples of output values can be thought of as lying on a grid, and in an embodiment, a 1 bit can be conveyed using either of the grid locations shown on the left, and a 0 bit can be conveyed using either of the grid locations shown on the right. Using these locations would remove noise modulation that would otherwise arise due to some least-significant bit pairs already conveying the desired bit and other pairs needing to be altered. The output signals would differ from the original signals by half a quantisation step, on average.

Description

DISCRETE DITHER

Background of the Invention

It is well known that quantisation of audio can result in objectionable artefacts.

These are commonly ameliorated by incorporating a suitable noise signal known as dither into the quantiser. Addition of dither with a rectangular probability density function (RPDF) spanning one quantisation step makes the resultant quantisation error uncorrelated with the signal but the dithered quantiser may still exhibit noise modulation. Changing the dither to a triangular probability density function (TPDF) spanning two quantisation steps also removes the noise modulation.

However these steps come with a noise penalty, a RPDF dithered quantiser is typically 3dB noisier than an undithered quantiser and a TPDF dithered quantiser 4.77dB noisier.

It is an object of the current invention to reduce the noise penalty in certain circumstances whilst still removing all noise modulation.

Summary of the Invention

The invention pertains to reducing the wordwidth of audio by 1 bit. For example, this might be reducing 17 bit audio to 16 bits. Or it might be to subtractively embed data into the audio as a fragile watermark.

According to an unclaimed aspect, there is provided a quantisation method for reducing the wordwidth of audio by 1 bit comprising the step of pseudo-randomly choosing one of two consecutive output values such that the probability of choosing one of them is 75% and the probability of choosing the other is 25%.

In this way, the noise introduced into the output audio is constant yet less than would be expected from the common practice of TPDF dithered quantisation.

According to a first aspect of the invention, there is provided a quantisation method for burying one bit of data in a pair of signal samples comprising the steps of: identifying two possible 2-tuples of output values both of which convey said one bit of data; and pseudo-randomly choosing one of the two possible 2-tuples with 50% probability.

According to a second aspect of the invention, there is provided a computer readable medium comprising instructions that, when executed by one or more processors, cause said one or more processors to implement the method of the first aspect.

In this way there is uncertainty and hence quantisation error in the output regardless of whether the pair of signal values would be interpreted as containing the desired bit of data or not. There is therefore the possibility of avoiding the noise modulation that would arise if the method specified no signal modification in the eventuality that the signal pair already conveys the desired bit. And yet since the selection is between two possibilities rather than the four neighbours that naturally occur in two dimensions, the noise can be lower than expected.

Brief Description of the Drawings

Figure 1 (prior art) shows how a dithered quanfiser adds dither (10) to a high precision input signal (1) prior to selecting (11) the nearest representable output value (2).

Figure 2 (prior art) shows how, when the dither is RPDF spanning 1 quantisation step size, the error power of the dithered quantiser varies with the high precision input.

Figure 3 shows how the invention incorporates a carefully chosen offset (12) which allows the dithered quantisation to have a constant variance independent of the input level (1) yet lower than the prior-art TPDF dither.

Figure 4 shows how a prior art watermarker might embed 1 data bit into two samples of audio as the XOR of their Isbs. In the left hand picture, the input pair of samples does not embed the desired value so one of the 4 neighbours is selected with equal probability. In the right hand picture, the input pair of samples does embed the correct value. However to keep a constant variance, one of the 4 more distant neighbours is sometimes selected.

Figure 5 shows a watermarker according to the invention embedding 1 data bit into two samples. By applying a suitable offset to the mean output value, the correct value can be embedded by selecting from one of two pairs of values with constant variance that is lower than in Figure 4.

Detailed Description

Figure 1 shows how a dithered quanfiser adds pseudo-random dither to a high precision input signal prior to selecting the nearest output value. We shall denote the quanfiser step size as A. It is well known that if the overall expected error of the dithered quanfiser is required to be zero, then the noise power from the dithered quantisation cannot be less than the values shown in Figure 2, and that these values are achieved by choosing the pseudorandom noise (10) to be RPDF spanning a range from -0.54 to +0.54.

It is also well known that it is further desirable for the dithered quantiser to have constant noise power. Since when the input level is exactly midway between permissible output values, the noise power cannot be less than 0.2542 this is achieved in the prior art by increasing the noise power up to 0.25442 for other input values by randomly selecting from 3 output values. One way to implement this is by making the pseudo-random noise source (10) have a triangular pdf (probability density function) spanning -A to +A.

When the input (1) is already quantised to one more bit than the output (2) there are only two cases to consider: i) The input is x + 0.54 for some permissible output value x. In this case both the above recipes lead to choosing an output of x with 50% probability or x + 1 with 50% probability. The output value has mean x + 0.54 which matches the input, and variance 0.2542.

ii) The input is some permissible output value x. In this case the RPDF recipe outputs x with 100% probability. This has zero variance, which differs from the prior case and hence the RPDF recipe exhibits noise modulation. The TPDF recipe outputs one of fx -1, x, x + 1} with probability [25%,50%,25%}. This still preserves the mean output value but increases the variance to 0.2542 which removes the noise modulation.

According to the invention, the dithered quantiser incorporates a constant offset of 0.254. A small DC offset is immaterial to audio but this small but crucial relaxation allows the variance to be constant but at a lower level than the

triangular dithered prior art.

If we now reconsider the two cases i) The input is x + 0.54 for some permissible output value x. We output x with 25% probability or x + 1 with 75% probability. The mean output is x + 0.754 but the variance is 0.187542 The input is some permissible output value x. We output x with 75% probability or x + 1 with 25% probability. The mean output is x + 0.254 and the variance is 0.187542 We thus have a constant expected error from the dithered quantiser and a constant variance of the output. However, this variance is 1.25dB lower than that produced by a triangular dithered quantiser.

One possible implementation of this is shown in Figure 3. The high precision input (1) has one more bit of precision than the output (2) and the pseudorandom noise (10) is RPDF spanning -0.54 to +0.54. The offset (12) of 0.254 ensures that quantisation decisions are made as described above.

The pseudo-random noise can take discrete values instead of continuous with exactly the same outcome. Possible ways of achieving the outcome described above are 4 equally possible values f -0.375a, -0.125A, 0.125A, 0.375A} or 3 unequally probable values 0,0.541 with probabilities [25%, 50%,25%}.

S

The offset (12) can also be incorporated into the pseudorandom noise before adding to the signal.

The above procedure generalises to the case where the input precision exceeds the output by k bits where k > 1. In this case the offset (12) is chosen as 2-("WA and the pseudorandom noise might take values 2-1'na with probability (2" -ml)2-2k for integer -2Ic < n < 2". As k increases, this approaches a continuous TPDF distribution and the variance advantage over a normal TPDF dithered quantiser decreases One scenario where quantisation reducing precision by a single bit can occur is embedding data into the audio lsb (least significant bit), for example into the 241h bit as a fragile watermark. Reducing the variance of this embedding increases the transparency of the watermark.

In this scenario, the desired data bit can be subtracted from an audio sample. This is then quanfised to 23 bits as described above and the desired data bit added back. This subtractive method ensures that the Isb of the audio holds the desired data and yet the whole procedure has constant expected error (independent of both the data bit and the original audio Isb) and a small variance.

The invention is also applicable for embedding data at a lower data rate, for example one data bit in a single stereo sample (or two consecutive samples on a single channel).

This might be done by defining the embedded data bit to be the XOR of the Isbs of the two samples.

A prior art approach is illustrated in Figure 4. We first consider if the pair of lsbs already convey the desired value.

If they do not, then one Isb needs changing to embed the value. There are 4 ways to do this with minimum error, by adding or subtracting A from either channel. The expected error is held zero by making adding and subtracting equally likely, the error variance is A2 and can be distributed evenly across both channels by randomly choosing which channel to alter. The net result is that the 4 neighbours are chosen each with 25% probability.

If they do contain the correct value, then the sample pair can be left unchanged. But if we are to have constant variance then with 50% probability we must alter both samples by +A. There are 4 ways to do this, which we do with 25% probability each.

According to the invention however, we introduce an offset of OSA on both samples as illustrated in Figure 5. Now there are 4 neighbours, two of which embed one value for the embedded bit and two of which have the other. We can make a 50% choice between the two neighbours with the desired embedded value. This approach has a constant mean error and variance 0.5A2 -half that

generated by the prior art approach.

Any of the methods described herein may be implemented by one or more processors executing instructions stored on a non-transitory data storage device or computer readable medium, the instructions causing the one or more processors to implement the respective methods.

Numerous modifications, adaptations and variations to the embodiments described herein will become apparent to a person skilled in the art having the benefit of the present disclosure, and such modifications, adaptations and variations that result in additional embodiments of the present invention are also within the scope of the accompanying claims.

Claims (3)

1. CLAIMS1. A quantisation method for burying one bit of data in a pair of signal samples comprising the steps of: identifying two possible 2-tuples of output values both of which convey said one bit of data; and, pseudo-randomly choosing one of the two possible 2-tuples with 50% probability.
2. 2. A quantisafion method according to claim 1, wherein the two possible 2-tuples differ by exactly 1 quanfisafion step in both dimensions.
3. 3. A computer readable medium comprising instructions that, when executed by one or more processors, cause said one or more processors to perform the method of claim 1 or claim 2.