DSP Performance Analysis in Audio Decoding Based on Psycho-Acoustics

Usually when we talk about DSP decoding and overall system performance , we mainly use traditional methods such as SNR, instantaneous error and phase error to make judgments. This article uses the method of psycho-acoustic compression design to examine the performance and performance of DSP decoding, introduces the concept of audio compression decoding based on psycho-acoustics, and gives a DSP performance analysis based on psycho-acoustics.

Since the 1990s, digital signal processing technology has gradually occupied an important position in the consumer audio market. Digital signal processors were originally used to process digitized analog audio signals, that is, PCM data. In the current era of system design, DSP system design solutions based on flexible software design features are an ideal alternative to traditional designs.

In the design of audio systems, the psycho-acoustic model is usually used to remove redundant data in the signal when the signal source is compressed and encoded. The performance of the system can be guaranteed by selecting a DSP with an appropriate number of bits. The selection of DSP in practical applications involves many factors, including accuracy (24-bit/32-bit), main frequency, cost, and memory capacity. This article analyzes the performance of DSP based on the psycho-acoustic model in audio decoding applications.

Relationship between DSP decibel and sound pressure decibel

The data described in the rest of this article are all measured at dBFS, i.e. decibels full scale. From the perspective of audibility, these values ​​need to be related to dB SPL, i.e. converted to decibels of sound pressure intensity. The analog signal chain after the DSP includes DAC, preamplifier, power amplifier and speaker. Although the gain and performance of each component may vary significantly from system to system, it is still possible to relate dBFS to dB SPL with sufficient accuracy from the perspective of system configuration alone.

Typically, digital tracks are recorded at -20dBFS, which fully meets the amplitude required for the signal peak, while also having enough dynamic range to show the quiet part of the audio file, and will not be distorted in different formats such as CD, Dolby Digital and DTS. As we all know, the auditory configuration recommended by THX is to reproduce a -20dBFS sound signal at 85dB sound pressure intensity, which is usually very loud, and normal listening will be much lower than this intensity.

From the above facts, it can be concluded that there is a linear mapping between dBFS and dB SPL, with the following relationship: a signal at 0dBFS can be reproduced at 105dB SPL. It should be noted that the sound produced in this case is very high and is not suitable for long-term listening; 0dB SPL corresponds to -105dBFS.

Hearing and hearing threshold

Human hearing has its limits, and usually 0dB is set as the lowest audible level in the sound pressure intensity design. Most of the sound spectrum (below 300Hz and above 10KHz) can only be heard above 10dB sound pressure intensity. The highest sensitivity of sine waves is 3~4KHz, and such sounds can be perceived by people with excellent hearing at -3~-4 dB SPL.

From a physiological point of view, to reach the threshold of hearing, the energy of the sound needs to be large enough to generate a standing wave in the eardrum, causing the tiny hairs there to fluctuate. Without this fluctuation, the neurons connected to the auditory cortex cannot be triggered, and the sound cannot be perceived. From the above discussion, we get inspiration about audio system design, that is, when the noise level is lower than the hearing threshold of people, it is meaningless to blindly pursue high-precision DSP implementation solutions.

Using the previously obtained listening configuration relationship, the lowest audible sound pressure is -4dB SPL, or -109dBFS. Assuming that all other parts of the signal chain (DAC, preamplifier, etc.) are zero distortion, this means that any DSP that can produce a better than 109dB signal-to-noise ratio will not become a bottleneck in system performance, which is a very important issue in system design using DSP. In actual applications, the analog signal chain is the main source of noise in the system, and the contribution of DSP to noise is much lower than these analog devices.

DSP bit number that meets system performance

The above analysis is based on an average level of -20dBFS and a THX listening configuration. Although this is an extreme case, some performance margin should be left in the design to take into account the change in the dBFS/dB SPL conversion relationship. Therefore, a well-considered design should make the DSP bit number about two more than the theoretical bit number, that is, 121dB using a 6dB/bit configuration, corresponding to a 20-bit dynamic range of the PCM output.

The above analysis is consistent with the assumptions of Dolby's Dolby Digital design solution, which also uses 20-bit accuracy. At the same time, the actual ADC/DAC is also limited to 20-bit performance (<120dB), and even DAT recording uses 20-bit accuracy. All of this verifies the correctness of the above analysis.

The above data is based on the worst case, because in reality the amount of noise generated by the power amplifier, preamplifier and DAC has a greater impact on the overall system performance than the performance of the DSP. The best power amplifier can only achieve a signal-to-noise ratio of 109dB, because noise energy can be accumulated in the linear region, which means that a 20-bit DSP with an output of 121dB will only generate 6.66% of the noise of the amplifier. If the performance of the speaker is also taken into consideration, the noise generated by the DSP is 1/6 - speaker distortion, which can be completely ignored.

The above analysis does not even take into account coding distortion, ADC or microphone noise, all of which are critical. If the entire signal chain is considered, it is clear that a 20-bit DSP is more than enough. Only by testing with abnormal sine waves, artificially synthesizing signals with a precision greater than 20 bits can a measurable difference be obtained, and this difference is actually imperceptible to human hearing.

Lossy compression with "transparent" audio quality

Psycho-acoustic compression is designed to compress a given signal in a lossy way, and then understand to what extent different frequency/time domain signals are audible or inaudible, so as to adjust the encoding process accordingly and reduce the introduced noise below the hearing threshold. The basic phenomenon is that the strong sound part of the signal will mask the adjacent weak sound part. Ideally, such data reduction will not lead to a perceived loss of sound quality, which leads to the concept of "transparent" audio coding or compression.

This is fundamentally different from a simple SNR measurement, and more complex, because it requires accurate reproduction of the relevant audible portion of a particular signal. In other words, while SNR is a good criterion for determining encoding/decoding quality, it is not an appropriate criterion for judging that a DSP that can produce -140dB THD+N is necessarily better than -130dB THD+N. Since psycho-acoustic compression designs are based on the human hearing threshold curve, the above conclusion becomes very obvious, and signals below this threshold cannot be heard.

Question about getting "transparent" audio compression

The actual encoding/decoding output quality is determined by the following factors:

1. Algorithm used

2. Compression bit rate

3. Psycho-acoustic model used in analyzing input signals

4. Transient Analysis Architecture and Transformation Filter Bank

5. Bit allocation strategy

In the encoding/decoding process, the above factors are independent of the algorithm accuracy. Even if infinite precision is used, the above factors still have a decisive influence on the audio quality.

Based on the lossy compression systems discussed above, the following conclusions can be drawn: Traditional measurements such as SNR, THD+N, transient error, and phase error are no longer the ultimate metrics for comparing the performance of different implementations. They can only be used as a reference when identifying and verifying system performance, and cannot be used to rate systems that vary slightly from the psycho-acoustically validated threshold of approximately 120dB performance.

DSP value-added suggestions

Usually 20-bit DSP is enough for the system, and 16-bit is enough from a psycho-acoustic point of view. For a given compression scheme, once the decoder reaches a certain performance, simply increasing the accuracy of the DSP will not further improve the system performance. The actual DSP value-added solution becomes the post-processing of the decoded audio program and the system-level features it provides. In fact, the final consumer also needs the product to have more additional features, such as automatic detection, error shielding and a post-processor that provides virtual sound effects.

Judging from the market response alone, the use of 32-bit DSP is still very successful in concept, because the performance improvement can be intuitively felt from the numbers. In fact, it is an obvious misunderstanding to consider the overall performance and more functions. This misunderstanding is equivalent to comparing only the CPU frequency of PCs without considering the overall performance of the system.

32-bit DSP does not really help improve the final performance of the system, it requires more memory (about 33% more than 24-bit DSP). At the same time, because 32×32MAC is slower than 24×24MAC, a 32-bit DSP core will always be slower than the decoder of the corresponding 24-bit core. In terms of signal-to-noise ratio, actual application results have proved that through optimization, 24-bit DSP can get the same (or even higher) performance as 32-bit DSP.

It should be noted that the above discussion is only based on the basic decoder, and the 32-bit DSP still has certain advantages in post-processing. Therefore, if the front-end uses a highly optimized 24-bit DSP decoding engine and the back-end processor is a 32-bit DSP, such an optimized design combines the advantages of both and is a good choice. Such a system-level chip solution is already available.

Conclusion

In the real world, when using psycho-acoustic compression coding, the SNR test method is not a suitable standard for comparing system performance. The ITU PEAQ test platform method is more suitable for measuring the quality of perceptible audio. In addition, when we evaluate a DSP solution, we also need to consider the system's error shielding, automatic detection and post-processing characteristics, as well as the system's startup response time and batch delay.

The factors that affect the quality of AV receiver system decoders are not all about DSP accuracy. Through strong DSP algorithm skills and optimization, a 24-bit precision DSP can outperform a 32-bit DSP. However, with the continuous improvement of manufacturing technology, 32-bit DSP will eventually solve the above problems, making 32-bit decoders more cost-effective and technically competitive.