Table 5.
Comparative analysis of noise mitigation techniques for TinyML-based voice assistants.
| Technique | Accuracy improvement (%) |
Computational cost |
Latency | Power consumption |
Hardware suitability |
Deployment challenges |
|---|---|---|---|---|---|---|
| ASR-based models | 10–12% | Medium | Low | Moderate | MCUs & DSPs | Requires well-trained noise suppression models |
| Spectral subtraction | 6–9% | Low | Very Low | Very Low | Ideal for low-power MCUs | Less effective in non-stationary noise environments |
| Deep learning-based methods (SEGAN, CNNs) | 15–18% | High | Medium-High | High | Requires advanced MCUs or Edge TPUs | High computational demand & energy usage |