VoiceCC

Voice-to-Command for AI coding CLI — Project Overview

What It Is

VoiceCC is a voice-to-command system that lets you control AI coding CLI (CLI) by speaking. You speak, the system transcribes via Whisper (local GPU), interprets the voice command, and injects it into the terminal session where AI coding CLI runs — both on WSL (tmux) and Windows (PowerShell/Terminal).

It supports direct text commands, special keys (escape, enter, tab), and intelligent interaction with AI coding CLI prompts (option selection, confirmations, custom input, multiselect) — completely hands-free.

Architecture

graph TB
    User["User\nVoice Input"] -->|"hotkey"| VE

    subgraph Windows
        VE["VoiceEmitter\nAudio Capture + Whisper STT\n+ Hotkey + System Tray"]
        WW["WindowsWorker\nTerminal Detection\n+ PostMessage Injection"]
        BR["Browser\nDashboard (React)"]
    end

    subgraph WSL / Linux
        HUB["SignalR Hub :5050\nMessage Broker + Filter Pipeline\n+ Choice Detection + Static Dashboard"]
        TW["TmuxWorker\nSession Monitor\n+ Tmux Injector"]
        TMUX["tmux session 'cc'\nAI coding CLI"]
    end

    VE -->|"SignalR\nvoice command"| HUB
    WW -->|"SignalR\nsession state + results"| HUB
    HUB -->|"route command\nto worker"| TW
    HUB -->|"route command\nto worker"| WW
    TW -->|"send-keys"| TMUX
    WW -->|"PostMessage\nWM_CHAR"| Terminal["Windows Terminal"]
    BR -->|"SignalR"| HUB
    HUB -->|"events + status"| BR
  

Development Time Estimate

How long would a super senior .NET/React developer take to build this from scratch, writing all code by hand?

Breakdown

Scenario Comparison

Where The Time Really Goes

The Windows Pipeline (Phase 4) dominates because:

• PostMessage injection requires reverse-engineering undocumented Win32 behavior
• The lParam scan code requirement is discovered only through trial-and-error
• UIAutomation for Windows Terminal is poorly documented — TextPattern access varies by terminal version
• Detecting AI coding CLI's state from the status bar requires reading terminal content without stealing focus
• Dual SignalR connection architecture (emitter + worker roles in one process) is non-trivial

The Intelligent Message (Phase 6) is the second time sink:

• AI coding CLI keybinding research must be done manually — no documentation exists, you test each prompt type by hand
• Screen parsing must handle 5 different prompt types with different separators and selection indicators
• Cross-platform differences (tmux uses '❯' U+276F, Windows uses '>' 0x3E) only surface during testing
• Multilanguage voice mapping (8 languages, number words, action words) requires meticulous testing

The Whisper STT (Phase 2) hides complexity:

• Whisper.net stable (1.9.0) doesn't support beam search — you need 1.9.1-preview1, discovered after hours
• WithProbabilities() shows 0% confidence unless enabled during processor creation, not after
• Language prompts must be IN the target language (not "speak Italian" but actual Italian vocabulary)
• Short audio (<1s) causes hallucinations ("thank you", "yeah") that need dedicated filtering

Actual Development Timeline (with AI coding CLI)

The project was built with AI coding assistant writing ~90% of the code, directed by a senior developer who designed architecture and made all decisions. Starting from absolute zero: no environment, no tools installed, just the idea. The 6 days include everything: environment setup, research, architecture, development, testing, deployment.

Day-by-Day Breakdown

Estimate vs Reality

Key Observations

1. Architecture + contracts + boilerplate = 10–17x speedup. AI generates DTOs, interfaces, events, DI registrations, SignalR hub scaffolding, React components almost instantly. What takes a human days of typing, AI produces in minutes.
2. Day 3 was superhuman: 46 commits, 5 features, 2 bugfixes in 14 hours. The entire Windows pipeline (UIAutomation + PostMessage + WindowsWorker + Hub routing fix) plus the message filter pipeline were built in a single day. This volume of correct, tested code is physically impossible without AI — it's not typing speed, it's the sheer breadth of simultaneous domain knowledge (Win32, SignalR, tmux, TDD).
3. Whisper.net + CUDA still required human debugging. Discovering that beam search needs a preview NuGet, that WithProbabilities() must be set during creation, that language prompts need actual vocabulary — these are the kind of undocumented gotchas where AI can research but the human must verify through real execution.
4. POC → production in hours, not days. PostMessage injection POC was validated and turned into a full WindowsWorker + integrated into VoiceEmitter in a single afternoon. A human would need the POC day plus 3–4 more days for the production implementation.
5. Tests were "free" — written alongside code, not after. With TDD driven by AI, 245 tests were produced as a natural byproduct of development. A human writing 245 tests by hand would need 5–6 dedicated days. With AI, testing cost was essentially zero — absorbed into every feature.
6. Total human effort was ~62 hours over 6 days, starting from absolute zero. The human work was: environment setup, architecture decisions, prompting AI, reviewing output, manual testing with real microphone + tmux + Windows Terminal, and debugging Whisper/CUDA behavior. The estimated manual equivalent is 62 working days (~500 hours). Effective acceleration: ~10x.