Clean Lab
Lab referencen/aNearest clean speaker
n/aF1-F3 drift (Hz)
n/aF0 drift (%)
n/aAmplitude-envelope r
Paired TIMIT acoustic probes
TIMIT Acoustic Analysis Demo
240
analyzed audio files
Research Question
This page uses clean Lab recordings from the four paired TIMIT speakers as references. MRI scanner noise is synthetically added to create Noisy (add), and the noisy signals are then enhanced with Denoiser, PASE, and RE-USE. The goal is to illustrate possible changes in speaker-related and acoustic cues, not only whether the output sounds cleaner.
- The speaker-reference ABX proxy compares each processed sample with clean same-speaker and other-speaker references for the same sentence block.
- The nearest-speaker mismatch proxy reports how often the nearest clean reference is not from the source speaker; lower is better.
- Formant, F0, harmonicity, envelope, and spectral-flux probes summarize phonetic, harmonic, and time-varying acoustic preservation.
- Uncertainty is shown as 95% speaker-cluster bootstrap confidence intervals from 5,000 resamples, resampling speaker IDs with replacement.
- Scope: These are illustrative, demo-only summaries and are not the manuscript's reported vowel-category ABX, PLV, FGER/FEE, SpkSim, or LPS analyses.
4paired speakers
12common sentence blocks
5clean/additive conditions
100%clean-reference ABX ceiling
Summary Across The Paired Additive Subset
| Condition | Speaker-reference ABX proxy (%) | Nearest-speaker mismatch proxy (%) | F1-F3 drift (Hz) | F0 drift (%) | Periodicity retained (%) | Amplitude-envelope r | Spectral-flux r |
|---|---|---|---|---|---|---|---|
| Clean Lab | 100.0 ± 0.00 | 0.0 ± 0.00 | 0 ± 0.00 | 0.0 ± 0.00 | 100 ± 0.00 | 1.00 ± 0.00 | 1.00 ± 0.00 |
| Noisy (add) | 53.5 ± 29.86 | 75.0 ± 25.00 | 532 ± 88.37 | 63.0 ± 22.11 | 81 ± 1.84 | 0.57 ± 0.20 | 0.57 ± 0.02 |
| Denoiser (add) | 66.7 ± 12.50 | 72.9 ± 27.08 | 203 ± 38.76 | 5.1 ± 5.16 | 108 ± 2.06 | 0.98 ± 0.02 | 0.60 ± 0.02 |
| PASE (add) | 80.6 ± 13.19 | 43.8 ± 31.25 | 107 ± 43.88 | 6.0 ± 4.28 | 107 ± 4.71 | 0.99 ± 0.00 | 0.71 ± 0.04 |
| RE-USE (add) | 81.2 ± 17.36 | 39.6 ± 37.50 | 167 ± 176.68 | 5.2 ± 3.74 | 102 ± 3.38 | 0.99 ± 0.01 | 0.75 ± 0.02 |
Values after ± are 95% speaker-cluster bootstrap CI half-widths computed with 5,000 resamples over speaker IDs.
Interactive Audio Examples
Select a speaker and sentence block, then compare Clean Lab, Noisy (add), and each SE output. The spectrogram scrolls with playback for timing inspection.
Noisy (add)
Noisy (add) proben/aNearest clean speaker
n/aF1-F3 drift (Hz)
n/aF0 drift (%)
n/aAmplitude-envelope r
Denoiser (add)
SE outputn/aNearest clean speaker
n/aF1-F3 drift (Hz)
n/aF0 drift (%)
n/aAmplitude-envelope r
PASE (add)
SE outputn/aNearest clean speaker
n/aF1-F3 drift (Hz)
n/aF0 drift (%)
n/aAmplitude-envelope r
RE-USE (add)
SE outputn/aNearest clean speaker
n/aF1-F3 drift (Hz)
n/aF0 drift (%)
n/aAmplitude-envelope r