One Model, Many Latencies: Universal Speech Enhancement for Diverse Real-Time Applications
This work addresses the practical problem of deploying speech enhancement models in real-time applications with varying latency requirements, offering a flexible and efficient solution for developers.
The paper proposes a single universal speech enhancement model that can be deployed across diverse real-time applications with different latency budgets, eliminating the need for separate models. The model achieves competitive performance across various latency settings, with a performance gap of less than 0.1 dB PESQ compared to specialized models.
Different real-time speech applications impose distinct latency budgets, often requiring separately trained enhancement models for each scenario. In this paper, we propose a one-for-all, real-time universal speech enhancement model that provides explicit control over both algorithmic and computational latency. Algorithmic latency is flexibly adjusted via configurable look-ahead frames. To avoid learning inefficiency caused by varying padding configurations, we introduce parallel convolutional layers corresponding to different look-ahead settings. Computational latency is controlled through an early-exit mechanism, enabling inference at different network depths. To narrow the performance gap between specialized and flexible models, we propose a two-stage training strategy with a shared-to-multiple decoder transition. Overall, the proposed framework enables a single model to be deployed across diverse latency budgets without retraining separate models.