Paper demo and analysis portal

Evaluating the Efficacy of Modern Speech Enhancement on Real-Time MRI Speech: A Multi-Task Benchmark and Open Resource

Authors: Huang-Cheng Chou, Sean Foley, Haley Hsu, Kevin Huang, Szu-Jui Chen, Rong Chao, Louis Goldstein, Khalil Iskarous, Dani Byrd, Yu Tsao, Sudarsana Reddy Kadiri, John H. L. Hansen, and Shrikanth S. Narayanan

3 demo pages

Demo Overview

This site accompanies a study of modern speech enhancement for real-time MRI speech. The demos are designed to show both listening quality and acoustic fidelity, because cleaner-sounding speech can still alter speaker-specific, phonetic, or time-varying cues that matter for speech science.

  • Post-DSP Audio compares Denoiser, PASE, and RE-USE on released rtMRI audio across multiple corpora.
  • Raw Audio tests whether starting from Raw scanner-contaminated audio changes the enhancement outcome.
  • TIMIT Acoustic Analysis uses clean Lab audio as a paired reference and focuses on the additive MRI-noise condition, Noisy (add), with 95% speaker-cluster bootstrap confidence intervals from 5,000 resamples.
8rtMRI demo samples in Post-DSP Audio
5TIMIT Lab/add conditions
12paired TIMIT sentence blocks analyzed

Open A Demo Page

Post-DSP Audio

Four rtMRI corpora

Compare publicly released Post-DSP audio against Denoiser, PASE, and RE-USE, with synchronized rtMRI video, transcript, and scrollable spectrograms.

Raw Audio

Long Single-Speaker raw basis

Use the Raw LSS rtMRI signal as the enhancement input to inspect whether Raw audio preserves details lost after corpus DSP cancellation.

TIMIT Acoustic Analysis

Clean Lab plus additive MRI noise

Inspect paired clean Lab, Noisy (add), and SE outputs with illustrative speaker-reference, formant, harmonic, and time-varying acoustic probes.