Yammer

Local speech-to-text with LLM correction for Linux. Speak naturally, get polished text typed into any application.

Features

• Voice Activity Detection - Automatically detects when you start/stop speaking
• Speech-to-Text - Local Whisper model for accurate transcription
• LLM Text Correction - Fixes homophones, adds punctuation, removes filler words
• Floating Overlay - Transparent window with real-time waveform visualization
• Universal Input - Types into any application via xdotool
• 100% Local - No cloud services, all processing on-device
• GPU Accelerated - Optional CUDA support for faster transcription

Requirements

• Linux with X11 (Wayland not yet supported)
• 4GB RAM minimum, 8GB recommended
• ~2GB disk for models

# Check your session type
echo $XDG_SESSION_TYPE  # Should output: x11

Quick Start

1. Install Dependencies

# Ubuntu/Debian
sudo apt install clang libclang-dev libasound2-dev xdotool libayatana-appindicator3-dev nodejs npm

# Arch Linux
sudo pacman -S clang alsa-lib xdotool libayatana-appindicator nodejs npm

# Fedora
sudo dnf install clang clang-devel alsa-lib-devel xdotool libayatana-appindicator-gtk3-devel nodejs npm

2. Build

git clone https://github.com/andynu/yammer.git
cd yammer
cargo build --release

First build takes 10-15 minutes (compiles Whisper and llama.cpp).

3. Download Models

cargo run --release --bin yammer download-models

Downloads Whisper (~141MB) and TinyLlama (~1.6GB) to ~/.cache/yammer/models/.

4. Run

# GUI mode
cd yammer-app && npm install && npm run tauri dev

# CLI mode (for testing)
cargo run --release --bin yammer dictate

Usage

GUI Mode

The floating overlay shows:

• Microphone status and state indicator
• Real-time waveform visualization
• Transcribed and corrected text

Press the global hotkey (configurable) to start/stop dictation. Text is automatically typed into the focused application.

CLI Commands

# Live dictation
yammer dictate

# List audio devices
yammer list-devices

# Record audio
yammer record --duration 5 --output test.wav

# Transcribe a file
yammer transcribe test.wav

# Test voice activity detection
yammer vad-test --duration 30

# Correct text with LLM
yammer correct "your transcribed text here"

Project Structure

yammer/
├── yammer-core/     # Shared types, config, model management
├── yammer-audio/    # Audio capture, VAD, resampling
├── yammer-stt/      # Whisper speech-to-text
├── yammer-llm/      # LLM text correction
├── yammer-output/   # Text output via xdotool
├── yammer-cli/      # CLI interface
└── yammer-app/      # Tauri desktop app

Architecture

Microphone → Audio Capture → VAD → Resampler (16kHz)
                                       ↓
                                   Whisper → LLM Correction → xdotool → Active Window
                                       ↓
                               Tauri UI (waveform, status, text)

Configuration

VAD sensitivity can be adjusted:

# More sensitive (quiet speech)
yammer dictate --threshold 0.005

# Less sensitive (loud/clear speech only)
yammer dictate --threshold 0.02

Performance

Component	CPU	GPU (CUDA)
Whisper (5s audio)	500ms-1.5s	250-750ms
LLM correction	100-500ms	-
VAD latency	~200ms start, ~400ms end	-

Troubleshooting

Build fails with "stdbool.h not found"

sudo apt install clang libclang-dev

"No audio devices found"

sudo apt install libasound2-dev
cargo clean && cargo build --release

Window not transparent Requires X11 with a compositor (standard on GNOME/KDE).

Transcription slow Use --release builds. Consider enabling CUDA if you have an NVIDIA GPU.

Documentation

• Audio Testing Guide
• UI Testing Guide
• Architecture & Onboarding

License

MIT - See LICENSE and THIRD_PARTY_NOTICES.md for details.

Credits

• whisper.cpp - Speech recognition
• llama.cpp - LLM inference
• Tauri - Desktop app framework
• cpal - Audio I/O