Exploring Disentangled Neural Speech Codecs from Self-Supervised Representations

Neural audio codecs (NACs), which use neural networks to generate compact audio representations, have garnered interest for their applicability to many downstream tasks -- especially quantized codecs due to their compatibility with large language models. However, unlike text, speech conveys not only linguistic content but also rich paralinguistic features. Encoding these elements in an entangled fashion may be suboptimal, as it limits flexibility. For instance, voice conversion (VC) aims to convert speaker characteristics while preserving the original linguistic content, which requires a disentangled representation. Inspired by VC methods utilizing $k$-means quantization with self-supervised features to disentangle phonetic information, we develop a discrete NAC capable of structured disentanglement. Experimental evaluations show that our approach achieves reconstruction performance on par with conventional NACs that do not explicitly perform disentanglement, while also matching the effectiveness of conventional VC techniques.

FasTUSS: Faster Task-Aware Unified Source Separation

Time-Frequency (TF) dual-path models are currently among the best performing audio source separation network architectures, achieving state-of-the-art performance in speech enhancement, music source separation, and cinematic audio source separation. While they are characterized by a relatively low parameter count, they still require a considerable number of operations, implying a higher execution time. This problem is exacerbated by the trend towards bigger models trained on large amounts of data to solve more general tasks, such as the recently introduced task-aware unified source separation (TUSS) model. TUSS, which aims to solve audio source separation tasks using a single, conditional model, is built upon TF-Locoformer, a TF dual-path model combining convolution and attention layers. The task definition comes in the form of a sequence of prompts that specify the number and type of sources to be extracted. In this paper, we analyze the design choices of TUSS with the goal of optimizing its performance-complexity trade-off. We derive two more efficient models, FasTUSS-8.3G and FasTUSS-11.7G that reduce the original model's operations by 81\% and 73\% with minor performance drops of 1.2~dB and 0.4~dB averaged over all benchmarks, respectively. Additionally, we investigate the impact of prompt conditioning to derive a causal TUSS model.

Physics-Informed Direction-Aware Neural Acoustic Fields

This paper presents a physics-informed neural network (PINN) for modeling first-order Ambisonic (FOA) room impulse responses (RIRs). PINNs have demonstrated promising performance in sound field interpolation by combining the powerful modeling capability of neural networks and the physical principles of sound propagation. In room acoustics, PINNs have typically been trained to represent the sound pressure measured by omnidirectional microphones where the wave equation or its frequency-domain counterpart, i.e., the Helmholtz equation, is leveraged. Meanwhile, FOA RIRs additionally provide spatial characteristics and are useful for immersive audio generation with a wide range of applications. In this paper, we extend the PINN framework to model FOA RIRs. We derive two physics-informed priors for FOA RIRs based on the correspondence between the particle velocity and the (X, Y, Z)-channels of FOA. These priors associate the predicted W-channel and other channels through their partial derivatives and impose the physically feasible relationship on the four channels. Our experiments confirm the effectiveness of the proposed method compared with a neural network without the physics-informed prior.

Data Augmentation Using Neural Acoustic Fields With Retrieval-Augmented Pre-training

This report details MERL's system for room impulse response (RIR) estimation submitted to the Generative Data Augmentation Workshop at ICASSP 2025 for Augmenting RIR Data (Task 1) and Improving Speaker Distance Estimation (Task 2). We first pre-train a neural acoustic field conditioned by room geometry on an external large-scale dataset in which pairs of RIRs and the geometries are provided. The neural acoustic field is then adapted to each target room by using the enrollment data, where we leverage either the provided room geometries or geometries retrieved from the external dataset, depending on availability. Lastly, we predict the RIRs for each pair of source and receiver locations specified by Task 1, and use these RIRs to train the speaker distance estimation model in Task 2.

Task-Aware Unified Source Separation

Several attempts have been made to handle multiple source separation tasks such as speech enhancement, speech separation, sound event separation, music source separation (MSS), or cinematic audio source separation (CASS) with a single model. These models are trained on large-scale data including speech, instruments, or sound events and can often successfully separate a wide range of sources. However, it is still challenging for such models to cover all separation tasks because some of them are contradictory (e.g., musical instruments are separated in MSS while they have to be grouped in CASS). To overcome this issue and support all the major separation tasks, we propose a task-aware unified source separation (TUSS) model. The model uses a variable number of learnable prompts to specify which source to separate, and changes its behavior depending on the given prompts, enabling it to handle all the major separation tasks including contradictory ones. Experimental results demonstrate that the proposed TUSS model successfully handles the five major separation tasks mentioned earlier. We also provide some audio examples, including both synthetic mixtures and real recordings, to demonstrate how flexibly the TUSS model changes its behavior at inference depending on the prompts.

Enhanced Reverberation as Supervision for Unsupervised Speech Separation

Reverberation as supervision (RAS) is a framework that allows for training monaural speech separation models from multi-channel mixtures in an unsupervised manner. In RAS, models are trained so that sources predicted from a mixture at an input channel can be mapped to reconstruct a mixture at a target channel. However, stable unsupervised training has so far only been achieved in over-determined source-channel conditions, leaving the key determined case unsolved. This work proposes enhanced RAS (ERAS) for solving this problem. Through qualitative analysis, we found that stable training can be achieved by leveraging the loss term to alleviate the frequency-permutation problem. Separation performance is also boosted by adding a novel loss term where separated signals mapped back to their own input mixture are used as pseudo-targets for the signals separated from other channels and mapped to the same channel. Experimental results demonstrate high stability and performance of ERAS.

TF-Locoformer: Transformer with Local Modeling by Convolution for Speech Separation and Enhancement

Time-frequency (TF) domain dual-path models achieve high-fidelity speech separation. While some previous state-of-the-art (SoTA) models rely on RNNs, this reliance means they lack the parallelizability, scalability, and versatility of Transformer blocks. Given the wide-ranging success of pure Transformer-based architectures in other fields, in this work we focus on removing the RNN from TF-domain dual-path models, while maintaining SoTA performance. This work presents TF-Locoformer, a Transformer-based model with LOcal-modeling by COnvolution. The model uses feed-forward networks (FFNs) with convolution layers, instead of linear layers, to capture local information, letting the self-attention focus on capturing global patterns. We place two such FFNs before and after self-attention to enhance the local-modeling capability. We also introduce a novel normalization for TF-domain dual-path models. Experiments on separation and enhancement datasets show that the proposed model meets or exceeds SoTA in multiple benchmarks with an RNN-free architecture.

Why does music source separation benefit from cacophony?

In music source separation, a standard training data augmentation procedure is to create new training samples by randomly combining instrument stems from different songs. These random mixes have mismatched characteristics compared to real music, e.g., the different stems do not have consistent beat or tonality, resulting in a cacophony. In this work, we investigate why random mixing is effective when training a state-of-the-art music source separation model in spite of the apparent distribution shift it creates. Additionally, we examine why performance levels off despite potentially limitless combinations, and examine the sensitivity of music source separation performance to differences in beat and tonality of the instrumental sources in a mixture.

NIIRF: Neural IIR Filter Field for HRTF Upsampling and Personalization

Head-related transfer functions (HRTFs) are important for immersive audio, and their spatial interpolation has been studied to upsample finite measurements. Recently, neural fields (NFs) which map from sound source direction to HRTF have gained attention. Existing NF-based methods focused on estimating the magnitude of the HRTF from a given sound source direction, and the magnitude is converted to a finite impulse response (FIR) filter. We propose the neural infinite impulse response filter field (NIIRF) method that instead estimates the coefficients of cascaded IIR filters. IIR filters mimic the modal nature of HRTFs, thus needing fewer coefficients to approximate them well compared to FIR filters. We find that our method can match the performance of existing NF-based methods on multiple datasets, even outperforming them when measurements are sparse. We also explore approaches to personalize the NF to a subject and experimentally find low-rank adaptation to be effective.

Generation or Replication: Auscultating Audio Latent Diffusion Models

The introduction of audio latent diffusion models possessing the ability to generate realistic sound clips on demand from a text description has the potential to revolutionize how we work with audio. In this work, we make an initial attempt at understanding the inner workings of audio latent diffusion models by investigating how their audio outputs compare with the training data, similar to how a doctor auscultates a patient by listening to the sounds of their organs. Using text-to-audio latent diffusion models trained on the AudioCaps dataset, we systematically analyze memorization behavior as a function of training set size. We also evaluate different retrieval metrics for evidence of training data memorization, finding the similarity between mel spectrograms to be more robust in detecting matches than learned embedding vectors. In the process of analyzing memorization in audio latent diffusion models, we also discover a large amount of duplicated audio clips within the AudioCaps database.