LocalVQE

LocalVQE is a set of compact neural models designed for enhancing voice quality in real-time audio streams. The tool addresses acoustic echo cancellation, noise suppression, and dereverberation for 16 kHz speech, making it suitable for scenarios where clear voice communication is essential. It operates efficiently on commodity CPUs and is optimized for streaming use, featuring a causal architecture with a 256-sample hop and approximately 16 ms latency.

The platform supports audio-to-audio processing and is implemented using PyTorch, with published model weights available. Inference can be performed using the GGML C++ engine on GGUF files, and build instructions are provided via its GitHub repository. The repository also includes source code, build systems, and tests, supporting integration and deployment for developers and researchers interested in speech enhancement tasks. A specific model variant, GTCRN-AEC, is mentioned as being tailored for lower-power CPUs, indicating attention to resource efficiency.

0 license, allowing for open use and modification. The tool is attributed to authors Richard Palethorpe and Claude (Anthropic). The Hugging Face page serves as a host for the model weights, and a demonstration is accessible online. LocalVQE is positioned as a streaming, CPU-tuned derivative of DeepVQE, focusing on real-time voice quality enhancement through neural network techniques.