Neural Network Speech Quality Assessment Without Reference Signals

The patent transforms the measurement approach by changing from classical physical parameters (signal-to-noise ratio) to perceptual parameters that model human speech quality perception. This includes using spectral distortion measures, temporal distortion measures, and other parameters that better reflect how humans perceive speech quality in modern communication systems with coding distortions, packet loss, and variable delays.

The patent replaces classical mechanical/electrical measurement systems with a perceptual model that simulates human auditory processing. This includes implementing models of human speech perception, auditory filtering, and quality assessment that better match subjective human evaluations rather than relying on traditional engineering metrics.

The patent extracts and removes the requirement for reference signals from the measurement system. By developing single-ended objective models that assess speech quality using only the degraded signal, the system eliminates the need for complex reference signal management, synchronization, and comparison infrastructure while maintaining assessment accuracy through alternative perceptual modeling approaches.

The patent enables the degraded signal to assess its own quality without external reference. The single-ended objective models analyze the degraded signal's inherent characteristics, spectral properties, and temporal features to determine speech quality, allowing the system to be self-sufficient and avoid the complexity of requiring pristine reference signals from the original source.

The patent creates objective models that copy and simulate human subjective quality assessment behavior. By training objective models on large datasets of paired subjective evaluations and objective measurements, the system learns to replicate human perception patterns, producing scores that closely match subjective LQS evaluations but can be generated automatically without human test participants.

The patent replaces the human testing system with an automated computational model. Instead of requiring human listeners to evaluate speech quality in controlled experiments, the system uses trained objective models that automatically predict subjective quality scores from signal characteristics, eliminating the need for human involvement while maintaining the perceptual accuracy of subjective tests.

The patent develops universal objective models that can assess speech quality across multiple bandwidths (narrow-band, wide-band, ultra-wide-band) and various coding types (codec standards, packet loss conditions, noise environments). The models are designed to be bandwidth-agnostic and coding-agnostic, providing consistent accurate assessments regardless of the specific transmission conditions or signal characteristics.

The patent implements dynamic models that adapt to different transmission conditions and signal characteristics. The objective models can dynamically adjust their analysis parameters, frequency resolution, and processing strategies based on the input signal properties and transmission conditions, providing optimized assessment for each specific scenario rather than using fixed parameters.