VoicePOC

Local-First Voice Engine & Proving Ground for macOS

VoicePOC is a standalone, local-first conversation loop engineered for macOS Apple Silicon. It forms the audio and speech pipeline powering Jarvis — delivering real-time wake-word detection, speech-to-text (STT), streaming LLM orchestration, and text-to-speech (TTS) synthesis with zero cloud round-trips.


Architecture & Audio Pipeline

VoicePOC operates as a 4-stage pipeline running local models on the Apple Neural Engine and GPU:

VoicePOC Local Audio Pipeline & System Topology

  1. Voice Activity Detection (VAD): Continuous low-power audio monitoring using Silero VAD over onnxruntime to filter ambient noise before triggering speech recognition.
  2. Speech-to-Text (STT): High-speed local speech recognition utilizing WhisperKit compiled directly for Apple Silicon hardware acceleration.
  3. Local LLM Streaming: Asynchronous streaming event loop interfacing with a local Ollama server instance via high-throughput HTTP/IPC streams.
  4. Speech Synthesis (TTS): Low-latency audio output processing using native AVSpeechSynthesizer or local neural TTS engines.

Technical Specifications

Component Technology Target Hardware
Language & Toolchain Swift 6.0 macOS Apple Silicon (M1/M2/M3/M4)
Speech Recognition WhisperKit (Argmax OSS) Apple Neural Engine / CoreML
VAD Runtime Silero VAD (ONNX Runtime 1.20) CPU / GPU Acceleration
LLM Interface Ollama Local Streaming API Apple Silicon Unified Memory
Audio Engine AVAudioEngine & CoreAudio Native macOS Audio Stack

System Requirements

  • Operating System: macOS 15.0+ or macOS 26.0+ (Apple Silicon aarch64-apple-darwin)
  • Developer Tools: Xcode 16.0+ with Swift 6.0 toolchain & xcodegen
  • Dependencies: Local ollama daemon installed (brew install ollama)

Build & Installation

Option 1: SwiftPM Library & CLI Build

# Build the core VoicePOCKit framework
swift build -c release

# Run integration tests
swift test

Option 2: Xcode App Bundle via XcodeGen

# Generate the native Xcode project bundle
xcodegen generate

# Build and launch the standalone macOS application
xcodebuild -project VoicePOC.xcodeproj -scheme VoicePOC -configuration Debug build

Quality Gates & Verification

# 1. Compiler & Linter Verification
swift build --target VoicePOCKit

# 2. Run Test Suite
swift test --filter VoicePOCTests

License & Credits

  • License: MIT License. See LICENSE for details.
  • Open Source Dependencies: WhisperKit by Argmax, ONNX Runtime by Microsoft, Swift Log by Apple.
S
Description
Local-first macOS voice POC: WhisperKit STT + Ollama LLM + AVSpeech TTS, hotkey-driven conversation loop. Sibling to MetalPOC + Jarvis.
Readme MIT
286 KiB
0 Stars 2 Watchers 0 Forks
Languages
Swift 98.3%
Shell 1.7%