This paper introduces automatic contextual audio denoising (ACAD), where what counts as “noise” depends on the situation rather than being fixed in advance. For example, traffic sounds may be unwanted during a phone call but useful in an urban monitoring application. The proposed system first infers the acoustic context of the recording, such as whether it comes from a street, restaurant, kitchen, or subway, and then uses that context to decide which sounds are relevant and which are out of context and should be removed. To study this problem, the paper also constructs a dataset in which the same type of sound can be considered noise in one context but useful in another, forcing the model to make genuinely context-dependent decisions. The method combines a context-recognition network with a neural denoiser whose processing is guided by the learned context. Experiments show that using inferred context improves denoising compared with a conventional context-independent system, and the best version improves SI-SDR by almost 2 dB over the standard baseline. Interestingly, the learned context performs even better than simply providing the correct scene label, suggesting that the model captures richer information about the acoustic situation than a predefined class alone. The results support the broader idea that future audio-processing systems can benefit from understanding where and under what circumstances a sound occurs before deciding what should be preserved or removed.

This paper tackles a practical problem in spatial audio: how to convert recordings from the small, irregular microphone arrays found in real devices into Ambisonics, a device-independent format used for immersive and 3D audio. Existing neural methods can work well for a specific microphone layout, but often need to be retrained when the array changes, and methods that only use microphone positions cannot describe more realistic effects such as microphone directionality or sound scattering caused by the device body. The proposed method addresses this by giving the neural network not only the recorded audio but also a detailed description of how the microphone array responds to sound coming from different directions and frequencies. Separate neural encoders process the audio and these directional characteristics, and cross-attention combines the two to create an array-independent spatial representation from which the Ambisonics signal is generated. The method is evaluated on simulated microphone arrays in both free-field conditions and a more realistic mobile-phone scenario with complex body scattering. It achieves the best overall signal reconstruction quality in both settings, outperforming conventional Ambisonics encoding and, in realistic phone-like conditions, existing neural approaches. Importantly, the model can generalize to previously unseen microphone configurations without retraining and does so with fewer than one million parameters, showing that detailed knowledge of a device’s directional acoustic behaviour can enable compact and flexible neural spatial-audio capture systems.

This paper asks a simple but important question about audio super-resolution: when a generative model creates the missing high-frequency content of a low-bandwidth recording, does the result merely sound realistic, or is it actually statistically similar to real wideband audio? To investigate this, the study evaluates several neural super-resolution systems for speech and music, including GAN-based and diffusion-based models, and compares real and generated audio not only with conventional signal-quality metrics and listening tests, but also by testing whether simple classifiers can tell them apart from learned audio representations. The results reveal a striking mismatch: some generated samples achieve very high perceptual quality and can sound close to the original wideband recordings, yet classifiers operating on audio embeddings can still distinguish real from synthetic samples with near-perfect accuracy. This behaviour is observed across different datasets, sampling-rate expansion tasks, and modern generative models. The findings therefore show that good perceptual quality and strong objective scores do not necessarily mean that a generative model has learned the true distribution of real audio, highlighting an important limitation in how audio super-resolution systems are currently evaluated and suggesting that representation-based tests can expose differences that conventional metrics and listening tests may miss.

This paper addresses a practical problem in long speech separation: when a long recording is split into shorter segments and each segment is separated independently, the same speaker may appear in different output channels from one segment to the next. This makes it difficult to reconstruct continuous, speaker-consistent audio, especially in meetings, interviews, or other recordings where speakers may remain silent for long periods. The proposed method solves this by maintaining a small, continuously updated reference pool of speaker embeddings for each output stream. For every new segment, the system compares its speaker embeddings with these reference pools and assigns the segment to the most likely speakers, while retaining only the most representative and reliable embeddings over time. The approach requires no additional training and can be added as a post-processing stage to existing speech-separation systems. Experiments on long recordings with both dense and sparse speech show that the method substantially improves cross-segment speaker consistency, reaching 98.5% permutation accuracy in dense scenarios and 96.8% in sparse ones, with separation quality close to the oracle upper bound. It is particularly effective when there are long gaps between utterances, where conventional overlap- or clustering-based methods degrade, and it remains robust even when the number of speakers is estimated incorrectly.

This paper addresses a major limitation of neural sound-source localization: most deep-learning methods work well only with the microphone array they were trained on, making deployment to new devices difficult without retraining. The proposed method instead uses detailed acoustic information about the microphone array—its array transfer functions, which describe how sounds arriving from different directions are captured by the microphones—rather than relying only on microphone positions. A neural network separately processes the recorded audio and these array characteristics, then combines them through cross-attention to estimate the directions of one or more sound sources. This allows the same model to operate on previously unseen microphone layouts, including realistic mobile-phone-like arrays where the device body alters the incoming sound field. Experiments in reverberant and noisy environments show that the method generalizes to new arrays with only moderate loss in performance and remains effective as the number of simultaneous sources increases. Classical MUSIC localization is still more accurate for simple single-source cases, but its performance degrades much faster as scenes become more complex, whereas the proposed neural approach remains more stable and can jointly detect and localize multiple sources. The results demonstrate that conditioning neural localization systems on the actual acoustic behaviour of a device is a promising route toward microphone-array-independent spatial audio processing.