DictatorFlow - Cross-Platform Voice-to-Text in Zig

DictatorFlow is a voice-to-text platform that ships as a single 23MB native binary for Linux, Windows, and macOS. It evolved from VoiceType, a Go-based Linux-only CLI tool, into a full cross-platform application rewritten in Zig with local GPU inference, multi-provider cloud fallback, and a web demo.

The Original: VoiceType (Go)

VoiceType was a Linux voice-to-text tool built in Go with a Python VAD (Voice Activity Detection) chunking layer. About 5137 lines total across Go and Python. You pressed Ctrl+Super+H, spoke, and it transcribed via a multi-provider fallback chain – trying several cloud APIs in sequence until one returned a result.

It worked well but had limitations: Linux-only, required a Python runtime for VAD, and the Go+Python split made deployment awkward. The GC pauses in Go also introduced occasional latency spikes during real-time audio processing.

Why Zig

Zig solves every pain point from the Go version:

• No GC – deterministic memory management, no pause spikes during audio capture
• Cross-platform from one codebase – Linux, Windows, macOS from the same source
• Direct C interop – call ONNX Runtime and Raylib C APIs without bindings overhead
• Single binary deployment – one 23MB file, no runtime dependencies
• Comptime – compile-time code generation for platform-specific audio backends

The Zig Rewrite

6358 lines of Zig across 21 source files. The major components:

• Raylib GUI overlay – translucent recording indicator, mode display, transcript history
• Cross-platform audio capture – ALSA on Linux, CoreAudio on macOS, WASAPI on Windows
• Platform-specific hotkeys – X11 key grabbing on Linux, RegisterHotKey on Windows
• VAD in Zig – no more Python dependency, Silero VAD via ONNX Runtime

Local GPU Inference

DictatorFlow runs NVIDIA’s Parakeet CTC 0.6B model locally via ONNX Runtime with CUDA acceleration. Build with -Dgpu to enable NVIDIA support.

1

zig build -Dgpu -Doptimize=ReleaseFast

This gives you fully private, offline transcription with no API calls. The ONNX Runtime CUDA execution provider handles GPU memory management and kernel dispatch.

Multi-Provider Cloud Fallback

When cloud is available, DictatorFlow uses a ranked fallback chain:

1

Gemini (quality) -> Groq (speed) -> Fal -> Local ONNX

The provider ranking is dynamic. Each utterance gets bucketed by audio length, and the system tracks per-bucket latency statistics. Short utterances might prefer Groq for speed while longer ones route to Gemini for accuracy. If a provider fails or times out, the next in the chain fires immediately.

WER Optimization

Starting from a baseline Word Error Rate of 1.614, DictatorFlow reached 0.041 – a 98% improvement. The key techniques:

• Spectral normalization – audio preprocessing that normalizes frequency energy distribution before inference
• VAD chunking – tuned silence thresholds to split audio at natural phrase boundaries
• Auto-learning – continuous WER-driven parameter optimization that adjusts preprocessing and chunking parameters based on measured accuracy

The auto-learning loop evaluates transcription accuracy against ground truth samples and tunes parameters like silence threshold duration, spectral normalization strength, and chunk overlap.

Voice Modes

Four distinct modes, each with its own hotkey:

Transcribe mode is the daily driver – press the hotkey, speak, release, and the text appears at your cursor. Command mode interprets speech as system commands. Research mode sends queries to AI research backends for longer-form answers.

Voice Activation

DictatorFlow supports always-on listening with local VAD. Configure wake words like “flow”, “hey”, “command”, or “research” to trigger specific modes without touching the keyboard. All activation detection runs locally – no cloud calls until actual transcription begins.

Architecture

1
2
3

Desktop Client (Zig) -> Cloud API (Go) -> Providers (Gemini/Groq/Fal)
                    \-> Local ONNX (Parakeet CTC 0.6B)
Web App (React) -----> Cloud API (Go)

The cloud API is a Go service that handles provider routing, rate limiting, and the web demo. The React web app at dictatorflow.com lets you try transcription in the browser via POST /demo/transcribe. The desktop client can operate fully offline with local ONNX or connect to the cloud API for the multi-provider fallback chain.

Try It

The web demo at dictatorflow.com runs in any modern browser. For the full desktop experience with hotkeys, voice activation, and local GPU inference, grab the Zig binary from the releases page.