• Model-based probabilistic frameworks unify speech enhancement with recognition by explicitly modeling speech and noise statistics, enabling Bayesian MMSE/MAP estimators, HMM decompositions, and state-dependent dynamics to separate speech from interference [2], [10], [3], [11], [19].

• Iterative, constrained spectral-domain optimization employs EM-like iterations and explicit spectral constraints, often guided by auditory constraints, to iteratively refine speech spectra under noisy conditions [1], [17], [13].

• Auditory-inspired and perceptual strategies use auditory evidence, Lombard-effect considerations, and morphological spectral constraints to preserve intelligibility and robustness under varying acoustic environments [5], [4], [9].

• Feature-level robustness for recognition in noise emphasizes robust spectral estimation and transform-based features, using energy-conditioned estimates, subband/envelope processing, and reliable filterbank representations to improve recognition [14], [12], [19], [18].

• Quality assessment and perceptual evaluation align objective measures with subjective judgments, guiding development and benchmarking of enhanced and coded speech [8], [20].