Can You Hear the Difference? A Quick Blind Audio Test for a New Digital Audio System

[Tom Jerry]

Hi All, I have also participated the test. My feeling is that, this survey is more for evaluating the quality of the audio, for example, audible distortion like digital hiss due to poor quantization, background noise, etc, rather than checking the subjective preference on audio genres/rhythms. So, I guess we should not let the results reflect our personal preferences on this listening test.

In my opinions, in the last 5x tests, clearly, 2/3 of the audio clips in each test sound flawless (they may sound shrill due to the electronic musical instruments, but there is no factual distortion), and should be rated with high scores.

 

[RandomEar]

@USF-Audio-2026 Interesting work! I checked the article you've linked and it certainly sounds neat from a mathematical / signal processing point of view. From my understanding, you cannot guarantee that the reconstruction gets all the folding events correct, though. Is that right? That seems like an important trade-off. If so, would it not be an acceptable compromise to sacrifice 1 or 2 bits of ADC resolution to log each folding event? If you use just 1 bit, you could log "did fold" for each sample point without the direction. This would probably improve the accuracy and reliabilty of your reconstruction by an order of magnitude. If you sacrifice 2 bits, you could log "did fold up/down", which would guarantee a perfect reconstruction. This sounds like a desirable outcome to me - at least as an optional mode of operation. The user could then select "give me 24 bits with some reconstruction errors", "give me 23 bits with far fewer reconstruction errors" or "give me 22 bits and a perfect reconstruction".

That being said, the word "folding" sounds a bit weird to me in this context. There are a couple of uses in programming or science which are quite far off from what happens in the context of this DSP technique. It also reminds me of the convolution, which used to be called Faltung and is definitely not what this is. Technically, it's kind of a controlled integer overflow. Maybe "wrapping" would be more fitting? I guess you've used the term folding extensively already, so it may be too late to switch now.

 

[Tom Jerry]

@USF-Audio-2026 Interesting work! I checked the article you've linked and it certainly sounds neat from a mathematical / signal processing point of view. From my understanding, you cannot guarantee that the reconstruction gets all the folding events correct, though. Is that right? That seems like an important trade-off. If so, would it not be an acceptable compromise to sacrifice 1 or 2 bits of ADC resolution to log each folding event? If you use just 1 bit, you could log "did fold" for each sample point without the direction. This would probably improve the accuracy and reliabilty of your reconstruction by an order of magnitude. If you sacrifice 2 bits, you could log "did fold up/down", which would guarantee a perfect reconstruction. This sounds like a desirable outcome to me - at least as an optional mode of operation. The user could then select "give me 24 bits with some reconstruction errors", "give me 23 bits with far fewer reconstruction errors" or "give me 22 bits and a perfect reconstruction".

That being said, the word "folding" sounds a bit weird to me in this context. There are a couple of uses in programming or science which are quite far off from what happens in the context of this DSP technique. It also reminds me of the convolution, which used to be called Faltung and is definitely not what this is. Technically, it's kind of a controlled integer overflow. Maybe "wrapping" would be more fitting? I guess you've used the term folding extensively already, so it may be too late to switch now.

Interesting observation. I also checked the linked paper and went through the math details. I think the authors managed to prove an exact signal reconstruction for bandlimited signals, without depending on any prior knowledge. There is some mathematical insight behind it, showing the perfect encoding/decoding of a smooth signal via its modulo (fractional) part. In some sense, this is equivalent to say, the audio can be recovered from its quantization noise, which is very counter-intuitive but exciting.

As for the term "folding", my thinking is that "folding" has more implications than "wrapping" or "overflow", since they usually refer to the arithmetic/digital operation after the quantization, i.e., the analog signal has been sampled and quantized. Yet, the subtlety in this work/survey is that, the folding is purposefully implemented in the continuous-time world, which is unprecedented and by careful design. This requires innovations across theory, hardware, and algorithms.

 

[RandomEar]

Interesting observation. I also checked the linked paper and went through the math details. I think the authors managed to prove an exact signal reconstruction for bandlimited signals, without depending on any prior knowledge. There is some mathematical insight behind it, showing the perfect encoding/decoding of a smooth signal via its modulo (fractional) part. In some sense, this is equivalent to say, the audio can be recovered from its quantization noise, which is very counter-intuitive but exciting.

Yeah, this topic is still a bit fuzzy to me. In practice the signal is likely not going to be perfectly bandlimited, so I would expect that some reconstruction errors are unavoidable.

As for the term "folding", my thinking is that "folding" has more implications than "wrapping" or "overflow", since they usually refer to the arithmetic/digital operation after the quantization, i.e., the analog signal has been sampled and quantized. Yet, the subtlety in this work/survey is that, the folding is purposefully implemented in the continuous-time world, which is unprecedented and by careful design. This requires innovations across theory, hardware, and algorithms.

True.

 

[Tom Jerry]

Yeah, this topic is still a bit fuzzy to me. In practice the signal is likely not going to be perfectly bandlimited, so I would expect that some reconstruction errors are unavoidable.

This is a cool insight and intuition. My thinking is that, the results from this folding work is akin to Shannon's sampling theorem. So, to make an analogy, the digital acquisition works flawlessly with analog input, and this underpins our modern technology. Following the same way, such imperfection of bandlimitedness should not be an issue. Moreover, any electronics that can be implemented will enforce the bandlimitedness due to the underlying physics.

Interestingly, this line of folding research has led to a series of works with hardware data validation, which is rare in signal processing/applied math papers.

 

[Sebby]

I just ran the test and only the last 3 samples had an acceptable audio file among the 3 options.
The other 7 samples had issues in all 3 options.
One of the 3 options was consistently terrible across all 10 samples.
In the first 7 samples I could assume that the problem isn’t the format but the track itself, with dynamic compression and extremely high background noise.