PhASE-Flow speeds speech enhancement with 4-step flow matching in SSL space
New speech enhancer runs 10x faster using SSL representation domain, not spectral.
Researchers from the audio processing community have introduced PhASE-Flow, a novel speech enhancement framework that operates entirely in the self-supervised learning (SSL) representation space rather than the traditional spectral domain. Existing flow matching methods for speech enhancement treat SSL features as external conditions, missing the structural richness of hierarchical SSL representations. PhASE-Flow instead models the conditional distribution of clean acoustic representations given phonetic representations, then reconstructs the waveform using a neural vocoder. This approach fully exploits both acoustic and phonetic levels of SSL models.
In experiments, PhASE-Flow outperformed state-of-the-art baselines in perceptual quality and intelligibility metrics. Crucially, it achieves competitive performance with only 4 sampling steps, enabling highly efficient inference suitable for real-time applications. The paper has been accepted at Interspeech 2026, and audio demos are publicly available. This breakthrough could significantly reduce computational costs for speech enhancement in hearing aids, VoIP, and noisy environment communication.
- Operates entirely in the SSL latent space, not the spectral domain, leveraging both acoustic and phonetic features.
- Achieves competitive performance with only 4 sampling steps, dramatically reducing inference time.
- Outperforms state-of-the-art baselines in perceptual quality and intelligibility metrics.
Why It Matters
Real-time, high-quality speech enhancement for hearing aids, VoIP, and noisy environments with minimal computational cost.