sonusai 1.3.2

Framework for building deep neural network models for sound, speech, and voice AI

SonusAI

Aaware Sound and Voice Machine Learning Framework.

SonusAI is a Python framework for building deep neural network models for sound, speech, and voice AI, with a performance-critical core implemented in Rust.

Key Features

• Performance Core: High-performance signal processing and data management in Rust.
• Data Augmentation: Advanced noise mixing, pitch shifting, and time-stretching.
• Model Support: Training and evaluation support for RCNNs, binary, and multi-label classification.
• Mixture Generation: Powerful engine for creating complex audio mixture databases.

Getting Started

Refer to the Development Guide for installation and setup instructions.

Prerequisites

• Python 3.13+ (Python 3.13 and 3.14 supported)
• Rust Toolchain
• uv

Quick Build

./build_local.sh

Documentation

• CLI Reference: Detailed overview of all sonusai commands.
• Development Guide: Setup, build, and test instructions.
• Data Formats & Dimensions: Technical specifications for data processing.
• Architecture: Details on the Rust/Python integration.
• Plugins: Summary of the plugin and extension system.

Current Performance Investigation Focus

• Active performance work is focused on genmixdb (mixture database generation) to establish bottlenecks before runtime refactoring.
• Real-world benchmark scenarios are currently centered on configurations under examples/ including mcgsp-valid, vctk28n2paper, voicebank, and voicebank8kof.
• Results are intended to guide both SonusAI core changes and downstream data-loading experience in sonusai_torchl projects.

Command Overview

   calc_metric_spenh            Run speech enhancement and analysis
   doc                          Documentation
   evaluate                     Evaluate model performance
   genft                        Generate feature and truth data
   genmix                       Generate mixture and truth data
   genmixdb                     Generate a mixture database
   gentcst                      Generate target configuration from a subdirectory tree
   keras_onnx                   Convert a trained Keras model to ONNX
   keras_predict                Run Keras predict on a trained model
   keras_train                  Train a model using Keras
   lsdb                         List information about a mixture database
   mkmanifest                   Make ASR manifest JSON file
   mkwav                        Make WAV files from a mixture database
   onnx_predict                 Run ONNX predict on a trained model
   plot                         Plot mixture data
   post_spenh_targetf           Run post-processing for speech enhancement targetf data
   torchl_onnx                  Convert a trained Pytorch Lightning model to ONNX
   torchl_predict               Run Lightning predict on a trained model
   torchl_train                 Train a model using Lightning
   tplot                        Plot truth data
   vars                         List custom SonusAI variables

Use sonusai help <command> for detailed information on any command.