Speech Signal Processing and Deep Neural Modeling for Pronunciation Evaluation in College English Teaching

Accurate and objective pronunciation evaluation in college English teaching remains challenging because accent, speaking rate, intonation, articulation, fluency, and recording conditions can substantially alter observed acoustic patterns. The main aim of this study is to develop a reproducible pronunciation-evaluation framework integrating speech signal processing, deep neural modeling, and dynamic state-space learning. The proposed Cognitive-Prosodic Neural Digital Twin Framework (CP-NDTF) jointly represents segmental, spectral, pitch, energy, rhythm, speech rate, fluency, and temporal stability information to dynamically characterize pronunciation performance. Evaluation is conducted using the publicly available L2-ARCTIC non-native English speech corpus, comprising 26,867 utterances from 24 speakers balanced across six first-language groups and gender. Manual phone-level annotations for 3,599 utterances identify substitution, deletion, and addition errors, providing an independent annotation basis for pronunciation-error targets. Original 44.1-kHz recordings are resampled to 16 kHz during preprocessing to establish a uniform analysis pipeline. CP-NDTF is compared with a conventional static deep neural pronunciation-scoring baseline using identical speaker-independent partitions, preprocessing procedures, feature inputs, and evaluation criteria. Controlled Gaussian and prosodic perturbations are applied exclusively for augmentation and robustness testing and are not presented as authentic environmental recordings. Results characterize model convergence, pronunciation-score and error behavior, prosodic alignment, representation structure, robustness, cross-speaker generalization, and feature interactions. The term "digital twin" is deliberately qualified because the framework updates a latent pronunciation state from discrete speech observations rather than implementing continuous bidirectional physical-virtual synchronization. Accordingly, CP-NDTF is positioned as a digital-twin-inspired computational learner-state model for reproducible pronunciation analysis. Hardware latency, peak memory consumption, parameter count, and floating-point operations are not claimed as measured deployment outcomes and remain important targets for subsequent device-level benchmarking and prospective educational validation. These findings provide a structured basis for future classroom-oriented pronunciation feedback and deployment studies.