A Predictive Dual-Stage Neural Framework for Phase-Coherent Auditory Synthesis on Edge Devices

Area of Science:

• Auditory Signal Processing
• Artificial Intelligence in Acoustics
• Digital Signal Processing

Background:

• Real-time binaural beat synthesis faces challenges like carrier non-stationarity and processing delays in dynamic environments.
• Conventional digital signal processing (DSP) pipelines struggle with phase coherence and introduce artifacts.

Purpose of the Study:

• To propose a predictive dual-stage neural framework for phase-coherent auditory synthesis in non-stationary acoustic conditions.
• To decouple carrier estimation from signal generation for improved real-time performance.

Main Methods:

• A dual-stage neural framework with an intelligent acoustic sensing module (AI-1) for adaptive carrier estimation.
• A predictive phase-coherent generator (AI-2) using short-horizon carrier trajectory forecasting for continuous phase evolution.
• Objective evaluation using Phase Coherence Factor (PCF) and Signal-to-Artifact Ratio (SAR).

Main Results:

• The framework achieved a PCF > 0.91 and SAR > 39.8 dB across various acoustic profiles and noise conditions.
• Performance improvement is attributed to the combined adaptive estimation and predictive generation, outperforming baselines.
• Hardware profiling demonstrated efficient INT8 inference time (2.4 ms/frame) on edge devices.

Conclusions:

• The proposed framework offers a lightweight solution for real-time, phase-continuous auditory synthesis in dynamic environments.
• It effectively reduces phase-boundary misalignment through predictive compensation, enhancing auditory signal processing.
• The results highlight the framework's potential for robust binaural beat generation in challenging acoustic settings.