A research team at Imperial College London, made up of Ruiming Guo, Yuliang Zhu and Ayush Bhandari, has developed USF-Audio, a new technique that folds an audio waveform before digitizing it and then mathematically reconstructs the original waveform afterward. The work was published in the journal IEEE Transactions on Audio, Speech and Language Processing (TASLP) and became available on IEEE Xplore on September 30, 2026. The university presented the results on November 11, saying a 6-bit prototype achieved at least 18 dB more dynamic range than a conventional 6-bit ADC. According to the paper's abstract, improvements were also confirmed on 14 real-world datasets and in a listening test with 52 participants. The idea is to change how information is represented within a limited number of bits, rather than adding more bits. However, the detailed measurement conditions behind the 18 dB improvement cannot be determined from the university's release or the paper's abstract.
Digital recording limits are set by bit depth and input range
PCM (pulse-code modulation), the foundation of digital recording, was conceived by Alec Reeves at IT&T's Paris laboratory in 1937. It measures the analog signal from a microphone at regular intervals and converts each value into an integer with a fixed number of bits. This conversion is called quantization, and the number of bits determines how many levels the signal's amplitude can be divided into.
For example, 6 bits represent amplitude with 64 levels (2 to the 6th power), and 7 bits with 128 levels. Because a fixed input range is divided into equal steps, each additional bit doubles the number of levels and halves the width of each step. This reduces the error between the original analog signal and the digitized value, allowing finer amplitude changes to be recorded.
According to Analog Devices' technical note MT-001, when a full-scale sine wave is input to an ideal N-bit ADC, the signal-to-noise ratio (SNR) is given by "6.02N + 1.76 dB." This value applies to the bandwidth from DC to half the sampling frequency, and it means SNR improves by about 6.02 dB with each additional bit. Applying the formula to 6 bits gives an ideal SNR of about 37.9 dB (editorial calculation).
However, increasing the number of bits to improve quantization precision does not widen the input range the ADC can handle. When an input signal exceeds the upper limit, the excess is pinned at the maximum value, and the peaks of the waveform are sliced flat in what is known as clipping. The paper's abstract points out that current digital audio technology relies on uniform sampling and uniform quantization based on the Shannon-Nyquist sampling theorem, which imposes fundamental limits on dynamic range and digital precision. Signals beyond the input range are lost to clipping, and even for signals within range, quantization noise limits the effective resolution.
Traditional responses to these problems include increasing the bit depth, setting headroom to leave margin below the maximum input, and oversampling, which raises the sampling frequency. Each involves trade-offs, however. For instance, larger headroom increases the relative impact of quantization noise on quiet sounds. All of them essentially retain the basic mechanism of measuring a signal at regular intervals and quantizing it within a fixed input range. USF-Audio aims to expand dynamic range in a different way by changing how the signal is acquired.
How USF-Audio folds the waveform instead of clipping it
The defining feature of USF-Audio is that it folds the analog signal through modulo arithmetic (modulo wrapping) before the ADC digitizes it. In a conventional ADC, once the input reaches the upper limit, anything beyond it cannot be recorded and the waveform is cut off. In USF-Audio, when the signal exceeds the upper limit, the excess is recorded as a value continuing from the bottom of the input range. According to the paper's abstract, this nonlinear acquisition method prevents measurement saturation and allows the original waveform to be mathematically recovered from the folded data.
A clock offers a useful analogy. After passing 12 o'clock, a clock starts counting again from 1, but looking only at the position of the hands, you cannot tell how many times they have gone around. A signal folded by USF-Audio is similar: a single measurement alone cannot reveal how many times the original amplitude exceeded the input range. What matters, then, is a mechanism that estimates the original waveform by exploiting the smoothness of audio signals and the relationships between consecutive measurements. However, the specific reconstruction procedure used in USF-Audio is not described in the university's release or the paper's abstract.
The foundation of this technology is a signal acquisition framework called Unlimited Sensing, which Associate Professor Bhandari and colleagues have been developing. According to the research group, for smooth signals, even the fractional part that corresponds to quantization noise contains clues for recovering the integer part, and this property can be used to reconstruct the original signal. As a result, they say, a signal with high dynamic range can be recovered with a fixed number of bits and a constant oversampling factor that does not depend on the maximum recorded voltage. Even so, this approach presupposes denser sampling than usual and does not overturn the sampling theorem itself. The oversampling factor used in this USF-Audio work cannot be confirmed from the release or the abstract.
Associate Professor Ayush Bhandari of Imperial College London's Department of Electrical and Electronic Engineering described the study this way: "what if we changed what the bits represent, rather than simply adding more of them?" (November 11, 2026). Techniques that repair audio lost to clipping, known as declipping, already exist, but they are fundamentally different from USF-Audio. Declipping estimates a waveform that has already been cut off from surrounding information, without directly recording the lost portion. USF-Audio, by contrast, applies folding at the acquisition stage, aiming to record amplitude information that would otherwise be lost in a form that can be recovered.
From PCM in 1937 to USF-Audio in 2026
The main technical milestones leading to this study run chronologically from the conception of PCM in 1937, to a prototype ADC based on Unlimited Sensing in 2021, an integrator-based modulo ADC in 2024, and USF-Audio for audio in 2026. According to an article Imperial College London published in 2021, Unlimited Sensing developed from Associate Professor Bhandari's doctoral thesis and a 2017 research paper, and a related US patent was granted in 2020. On September 16, 2021, Bhandari and colleagues published the paper "Unlimited Sampling from Theory to Practice: Fourier-Prony Recovery and Prototype ADC" in IEEE Transactions on Signal Processing, presenting a prototype ADC using the modulo approach.
In 2024, Yuliang Zhu and Associate Professor Bhandari presented research on a low-cost modulo ADC using an integrator at IEEE Sensors 2024. That work was not specific to audio, however, and the signals and experimental conditions it covered differ from those of USF-Audio. The USF-Audio paper then became available on IEEE Xplore on September 30, 2026, and the university presented the results on November 11.
The paper reportedly presents theoretical guarantees for recovering the original signal from folded quantized measurements, a robust reconstruction algorithm that handles audio-scale data, and a hardware-based signal acquisition system. It also includes evaluations on real audio and a listening test with 52 people. Whereas the 2021 work centered on a modulo ADC prototype and the 2024 work on a low-cost conversion circuit, the present study is distinctive in combining theory, hardware, and perceptual quality evaluation for the specific application of audio. The team says this series of validations shows that Unlimited Sensing could be applied to real audio acquisition.
18 dB more dynamic range with 6 bits, though some measurement conditions remain unclear
According to Imperial College London, the prototype operates at 6 bits per sample and achieved at least 18 dB more dynamic range than a conventional 6-bit ADC. The university explains that this corresponds to roughly an eightfold expansion of the amplitude range that can be recorded. The prototype ADC was assembled from commercially available electronic components and reportedly does not require a special semiconductor manufacturing process. However, the 18 dB figure is based on the university's announcement, and we have not verified the details in the body of the paper, including measurement conditions.
It is worth noting that the 18 dB improvement refers to an expansion of dynamic range and does not necessarily mean an improvement in signal-to-noise ratio (SNR). The university's release describes a wider range of recordable amplitudes and does not say SNR improved by 18 dB. The 18 dB therefore cannot simply be added to the ideal 6-bit ADC SNR of about 37.9 dB shown above. Converting an amplitude ratio to decibels with 20×log10(8) gives about 18.06 dB, which is consistent with the university's explanation of roughly eightfold (editorial calculation).
The paper's abstract says performance was evaluated on 14 real-world datasets covering speech, music, biological sounds, animal calls, historical recordings, and more. A subjective listening test with 52 participants was also conducted, comparing recordings from conventional ADCs and state-of-the-art declipping methods. The results reportedly showed consistent improvements in dynamic range, ENOB (effective number of bits), and perceived audio quality. The university's release says that in the listening test, audio recovered with USF-Audio was rated higher than conventional 6-bit recordings and recordings with clipping.
On the other hand, the university's release and the paper's abstract alone do not provide enough of the conditions needed to examine the results in detail. For example, the type and amplitude of the input signal, the oversampling factor, and the sampling frequency used when measuring the 18 dB improvement are not disclosed. For the listening test, the rating scale, number of trials, and statistical test details are also not given. Nor can we confirm the names of the declipping methods used for comparison, their individual improvement margins, or the specific assumptions under which the theoretical recovery guarantees hold.
These conditions are important for judging how far the results can be generalized. Without knowing the type and amplitude of the input signals, for instance, it is impossible to tell whether the 18 dB improvement applies to a wide range of audio or was measured under specific conditions. Likewise, the claim of consistent superiority over existing declipping methods requires checking the specific margin of improvement against each comparison method. These are points to be examined in the body of the paper.
Digital stethoscopes and environmental sensors envisioned; practical use needs more validation
The research team cites digital stethoscopes, environmental sensors, and sensing systems for AI as potential applications for USF-Audio. In such devices, small signals must be measured with high precision while sudden large inputs may also occur, so handling a wide range of amplitudes is important. Because USF-Audio may be able to retain information from large input signals without adding bits, the team expects it could eventually lead to compact, low-power sensing devices.
However, lower power consumption and smaller size are currently presented as future possibilities rather than quantitatively demonstrated results in actual devices. The university's release gives no specific figures on how much performance would improve in each application, such as digital stethoscopes or environmental sensors. The prototype is a research-stage device built from commercial components, and the team itself acknowledges that further engineering work is needed to establish these benefits in practical equipment.
What USF-Audio shows is that recording performance may be improved not only by increasing an ADC's bit depth but also by changing how signals are acquired. Reporting at least 18 dB of dynamic range expansion with a 6-bit prototype, along with improvements confirmed in a 52-person listening test, can be seen as an early result supporting its effectiveness. Next steps include a detailed examination of the measurement conditions, as well as evaluating performance, power consumption, and circuit complexity when it is built into practical audio equipment.