Robust Speech Recognition with Schrödinger Bridge-Based Speech Enhancement

TL;DR

The paper tackles ASR robustness in noisy, reverberant environments by deploying a Schrödinger-bridge-based generative speech enhancement front-end. It formulates SE as a data-to-data problem with Gaussian boundary conditions and trains a score-based backbone (NCSN++) to predict clean TF representations, using a time-domain auxiliary loss. Empirical results show the SB front-end yields large WER improvements (up to roughly 40% relative vs unprocessed and ~8% relative vs a similarly sized predictive model) and maintains advantages across multiple ASR architectures and noise conditions. The findings establish Schrödinger-bridge SE as a competitive alternative to diffusion-based and predictive methods for ASR enhancement in realistic, adverse acoustic settings.

Abstract

In this work, we investigate application of generative speech enhancement to improve the robustness of ASR models in noisy and reverberant conditions. We employ a recently-proposed speech enhancement model based on Schrödinger bridge, which has been shown to perform well compared to diffusion-based approaches. We analyze the impact of model scaling and different sampling methods on the ASR performance. Furthermore, we compare the considered model with predictive and diffusion-based baselines and analyze the speech recognition performance when using different pre-trained ASR models. The proposed approach significantly reduces the word error rate, reducing it by approximately 40% relative to the unprocessed speech signals and by approximately 8% relative to a similarly sized predictive approach.

Figures (2)

• Figure 1: WER and SI-SDR for the SB model with ODE sampler (configuration 4 in Table \ref{['table:models']}) and different number of sampling steps.
• Figure 2: ASR performance in terms of WER vs. RSNR for the best SB model from Table \ref{['table:models']} using ten sampling steps.