RTFS-Net: Recurrent Time-Frequency Modelling for Efficient Audio-Visual Speech Separation

TL;DR

RTFS-Net addresses the cocktail party problem in audio-visual speech separation by performing independent time and frequency modeling on STFT bins within a lightweight, TF-domain framework. It introduces a Cross-dimensional Attention Fusion for multimodal integration, a dual-path SRU-based RTFS Block for efficient recurrent processing, and a Spectral Source Separation (S^3) module that preserves complex-valued spectral information. The approach yields state-of-the-art separation quality with dramatically fewer parameters and MACs across multiple AVSS datasets, validated by extensive ablations. Its TF-domain emphasis, efficient fusion, and spectral-aware reconstruction offer practical benefits for real-time AVSS applications and broader audio-visual signal processing tasks.

Abstract

Audio-visual speech separation methods aim to integrate different modalities to generate high-quality separated speech, thereby enhancing the performance of downstream tasks such as speech recognition. Most existing state-of-the-art (SOTA) models operate in the time domain. However, their overly simplistic approach to modeling acoustic features often necessitates larger and more computationally intensive models in order to achieve SOTA performance. In this paper, we present a novel time-frequency domain audio-visual speech separation method: Recurrent Time-Frequency Separation Network (RTFS-Net), which applies its algorithms on the complex time-frequency bins yielded by the Short-Time Fourier Transform. We model and capture the time and frequency dimensions of the audio independently using a multi-layered RNN along each dimension. Furthermore, we introduce a unique attention-based fusion technique for the efficient integration of audio and visual information, and a new mask separation approach that takes advantage of the intrinsic spectral nature of the acoustic features for a clearer separation. RTFS-Net outperforms the prior SOTA method in both inference speed and separation quality while reducing the number of parameters by 90% and MACs by 83%. This is the first time-frequency domain audio-visual speech separation method to outperform all contemporary time-domain counterparts.

Figures (4)

• Figure 1: The overall pipeline of RTFS-Net. The red and blue solid lines signify the flow directions of auditory and visual features respectively. The snowflake indicates the weights are frozen and the component is not involved in training.
• Figure 2: Structure of our CAF Block. As in Figure \ref{['fig:av_pipeline']}, the red and blue solid lines signify the flow directions of auditory and visual features respectively.
• Figure 3: RTFS Block design. After compressing the data to a more efficient size, we process first the frequency dimension, then the time dimension, then both dimensions in tandem using TF-domain self-attention to capture inter-dependencies. We then carefully restore the data to its original dimensions using our Temporal-Frequency Attention Reconstruction units.
• Figure 4: Comparison of target speaker extraction methods. In \ref{['fig:suba']}, \ref{['fig:subb']} and \ref{['fig:subc']}, the top spectrogram is the target speakers original audio, the middle spectrogram is obtained using RTFS-Net-4 with $S^3$, and the bottom spectrogram is obtained using RTFS-Net-4 using the traditional masking approach.