New AI Makes Fuzzy Audio Sound Crystal Clear, Live and on a Phone
Say goodbye to muffled voice calls and grainy voice memos — instantly clearer audio.
Researchers propose StreamWSR, a streamable neural model for waveform-domain speech super-resolution. It uses a fully causal architecture with compact frame-level waveform representations, allowing zero-look-ahead streaming inference while avoiding vocoder-based reconstruction and explicit phase prediction. The model downsamples the input with strided causal convolutions, captures local waveform structures and long-range historical dependencies with a lightweight causal backbone, then generates high-resolution speech through causal transposed convolutions and a residual connection. On 16 kHz speech super-resolution, StreamWSR achieves competitive or superior quality and intelligibility compared with representative waveform- and spectrum-based baselines, while keeping a zero-look-ahead advantage at just 9M parameters and 2G FLOPs.
- StreamWSR improves speech quality, turning muffled audio into clear, high-definition sound.
- It's a real-time system that works with zero delay, needing just a split second of audio, not the full clip.
- Because it's small and lightweight (9 million parameters), it could run on phones, tablets, and other everyday devices.
Why It Matters
Better-sounding calls, hearing aids, and voice assistants without extra delay or bulging costs.