Comfort Noise Generation for Low Bit-Rate Audio Coding

The patent segments the audio signal into two separate sources: speech signal and background noise. The encoder separately codes speech parameters and noise parameters, allowing independent optimization of each source. This segmentation enables the decoder to reconstruct both components and combine them, maintaining audio quality while operating at low bit-rates by only transmitting essential parameters for each source rather than full audio data.

The patent introduces a dual-mode coding structure that acts as an intermediary between single-source coding and full-bandwidth transmission. The system uses voice activity detection to determine whether to activate speech-only mode or speech-plus-noise mode, providing a flexible intermediate solution that adapts to signal conditions and maintains quality across varying bit-rate requirements.

The patent extracts the background noise component from the mixed audio signal and codes it separately from the speech signal. By identifying and separating noise parameters through noise estimation algorithms, the system can represent noise independently, allowing the decoder to reconstruct the original noise characteristics without applying aggressive noise reduction that would distort speech or create artifacts.

The patent changes the representation parameters by coding noise characteristics (such as spectral shape, level, and temporal properties) separately from speech parameters. This parameter-based approach allows flexible control over noise reconstruction, enabling the system to maintain natural noise characteristics while enhancing speech intelligibility through selective noise management rather than aggressive filtering.

The patent segments the parameter coding into speech-specific parameters and noise-specific parameters, each with their own quantization schemes. This allows optimization of quantization precision for each parameter type based on its perceptual importance and variability, improving overall efficiency while maintaining quality. Critical speech parameters can use finer quantization while less critical noise parameters use coarser quantization.

The patent employs adaptive quantization strategies where the precision of parameter coding is dynamically adjusted based on signal conditions, bit-rate availability, and perceptual importance. By changing quantization parameters adaptively and using entropy coding techniques, the system achieves efficient compression while minimizing perceptible artifacts from quantization.