Case study · AI + speech + DSP
Attention-based speech denoising.
A four-person team project combining signal processing and learned temporal-spectral modeling, evaluated against MetricGAN+ across three test datasets.
Problem
Recover clearer speech without erasing it.
Speech enhancement must reduce background noise while preserving the structure listeners and downstream models depend on. The challenge is not simply generating a smoother waveform; it is balancing speech quality, signal distortion, and residual noise.
Architecture
Local patterns, long-range context, and frequency focus.
The model uses audio-to-STFT preprocessing, residual convolutional encoder-decoder blocks, Gated DeltaNet temporal modeling, frequency-axis attention, and a learned noise gate.
Engineering decisions
Give each block one job.
- STFT representation: exposes local spectral structure that is difficult to see directly in raw samples.
- Residual convolutions: learn local time-frequency features while preserving trainability through a deeper encoder-decoder.
- Gated DeltaNet: captures temporal relationships beyond a local convolutional window.
- Frequency attention: lets the model emphasize informative spectral regions.
- Learned noise gate: provides a trainable final suppression mechanism.
Dataset + method
Paired training and three held-out evaluations.
The team trained on 23,075 paired clean/noisy samples at 16 kHz. Evaluation used three 824-example test sets. The same metric family was applied to the proposed system and the MetricGAN+ comparison.
Results
Higher than MetricGAN+ across every reported dataset and metric.
The model outperformed MetricGAN+ on PESQ, CSIG, CBAK, and COVL across all three evaluation datasets. Exact University of Edinburgh results are shown below; additional per-dataset numbers will be added with the public report.
| Dataset | PESQ | CSIG | CBAK | COVL | vs. MetricGAN+ |
|---|---|---|---|---|---|
| University of Edinburgh | 3.1383 | 4.5913 | 3.9756 | 3.9236 | Higher on all four |
Ablations
Attention was critical; Gated DeltaNet was consistently useful.
Controlled ablations showed that removing attention caused the clearest degradation. Gated DeltaNet also improved performance consistently, supporting the value of explicit temporal modeling rather than relying on convolutional blocks alone.
My contribution
Led the team and drove implementation.
I led the four-person team and drove much of the implementation. The benchmark results are team outcomes; this case study separates that shared result from my individual ownership.
Demonstration
Audio and spectrogram release in preparation.
Public audio, aligned spectrograms, source code, and the final report will be linked here after the team confirms a suitable release package.
Limitations
Objective scores are not the whole listening experience.
PESQ, CSIG, CBAK, and COVL are useful, but they do not replace controlled listening tests or evaluation on deployment-specific noise. Public release should also state the exact data splits, preprocessing configuration, hardware, and inference boundary.
Lesson
Ablations turn architecture into an argument.
The strongest result was not only the benchmark win. The ablation study showed which parts of the model actually mattered, making the design more interpretable and the claims more credible.