Noisy Disentanglement with Tri-stage Training for Noise-Robust Speech Recognition

Abstract: To enhance the performance of end-to-end (E2E) speech recognition systems in noisy or low signal-to-noise ratio (SNR) conditions, this paper introduces NoisyD-CT, a novel tri-stage training framework built on the Conformer-Transducer architecture. The core of NoisyD-CT is a especially designed compact noisy disentanglement (NoisyD) module (adding only 1.71M parameters), integrated between the Conformer blocks and Transducer Decoder to perform deep noise suppression and improve ASR robustness in challenging acoustic noise environments. To fully exploit the noise suppression capability of the NoisyD-CT, we further propose a clean representation consistency loss to align high-level representations derived from noisy speech with those obtained from corresponding clean speech. Together with a noisy reconstruction loss, this consistency alignment enables the NoisyD module to effectively suppress noise while preserving essential acoustic and linguistic features consistent across both clean and noisy conditions, thereby producing cleaner internal representations that enhance ASR performance. Moreover, our tri-stage training strategy is designed to fully leverage the functionalities of both the noisy disentanglement and speech recognition modules throughout the model training process, ultimately maximizing performance gains under noisy conditions. Our experiments are performed on the LibriSpeech and CHiME-4 datasets, extensive results demonstrate that our proposed NoisyD-CT significantly outperforms the competitive Conformer-Transducer baseline, achieving up to 25.7% and 10.6% relative word error rate reductions on simulated and real-world noisy test sets, respectively, while maintaining or even improving performance on clean speech test sets. The source code, model checkpoint and data simulation scripts will be available at https://github.com/litchimo/NoisyD-CT.