Audient

Overview

Audient is a privacy-first desktop application that captures, transcribes, and analyzes audio using local AI models. All processing happens on-device — your audio never leaves your machine. Built with Electron + React for the UI and FastAPI for the backend, it leverages Llama 3.2, Faster Whisper, and Sherpa-ONNX for fully offline intelligence.

Features

• Privacy-First – 100% local processing, zero data sent to external servers
• Multi-Source Recording – Capture system audio and microphone simultaneously with real-time level monitoring
• Speech-to-Text – High-accuracy transcription powered by Faster Whisper
• Speaker Diarization – Identify and separate different speakers using Sherpa-ONNX
• AI Chat – Interact with local Llama 3.2 LLM about your transcripts
• Smart Summaries – Generate AI-powered summaries of recordings
• Agent System – Create and manage custom LLM-powered agents for specialized audio analysis
• Workflow Automation – Build and execute multi-step processing pipelines
• JSON Storage – Human-readable, portable flat-file data storage (no database required)
• Cross-Platform – Desktop app for Windows, macOS, and Linux

Architecture

• Frontend: Electron + React + Vite (Desktop UI)
• Backend: FastAPI + Python 3.12+ (REST API)
• LLM: Llama 3.2 (3B Instruct, Q4_K_M quantized) via llama-cpp-python
• Speech-to-Text: Faster Whisper
• Speaker Diarization: Sherpa-ONNX (Pyannote segmentation + WeSpeaker embeddings)
• Storage: JSON-based flat files for portability
• Packaging: PyInstaller (backend), electron-builder (frontend)

Local AI Models

| Model | Purpose | |-------|---------| | Llama 3.2 | Chat, summaries, agent reasoning | | Faster Whisper | High-accuracy speech-to-text | | Sherpa-ONNX | Speaker diarization (who spoke when) |