thread/hyprvoice
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

docs: update readme with local and streaming transcription

Browse Source
This commit is contained in:
leonardotrapani
2026-02-01 02:12:12 +01:00
parent a105c8c5a8
commit 10ec54fc99
3 changed files with 144 additions and 13 deletions
Download Patch File Download Diff File Expand all files Collapse all files
README.md
+131 -11
View File
@@ -8,7 +8,10 @@ Press a toggle key, speak, and get instant text input. Built natively for Waylan
- **LLM post-processing**: Automatically cleans up transcriptions - removes stutters, fixes grammar, adds punctuation (enabled by default)
- **Wayland native**: Purpose-built for Wayland compositors - no legacy X11 dependencies or hacky workarounds
- **Real-time feedback**: Desktop notifications for recording states and transcription status
- **Multiple transcription backends**: OpenAI Whisper, Groq, Mistral Voxtral, and Eleven Labs Scribe (99 languages, excellent accuracy)
- **Multiple transcription backends**: OpenAI Whisper, Groq, Mistral Voxtral, ElevenLabs Scribe, and Deepgram Nova
- **Local transcription**: Offline transcription via whisper.cpp - no API keys, no cloud, complete privacy
- **Streaming transcription**: Real-time results with ElevenLabs, Deepgram, and OpenAI Realtime
- **57 language support**: Full multilingual support with language-model compatibility validation
- **Smart text injection**: Clipboard save/restore with direct typing fallback
- **Daemon architecture**: Lightweight control plane with efficient pipeline management
@@ -63,7 +66,7 @@ export PATH="$HOME/.local/bin:$PATH"
- **Wayland desktop** (Hyprland, Niri, GNOME, KDE, etc.)
- **PipeWire audio system** with tools
- **API key for transcription**: OpenAI, Groq, Mistral, or Eleven Labs API key (check each provider's pricing)
- **API key for transcription**: OpenAI, Groq, Mistral, ElevenLabs, or Deepgram API key (check each provider's pricing), OR whisper.cpp for local transcription (no API key required)
**System packages** (automatically installed with AUR package):
@@ -153,6 +156,25 @@ hyprvoice version
hyprvoice stop
```
### Model Management (Local Transcription)
```bash
# List all available models
hyprvoice model list
# List only transcription models
hyprvoice model list --type transcription
# List models for a specific provider
hyprvoice model list --provider whisper-cpp
# Download a local model
hyprvoice model download base.en
# Remove a downloaded model
hyprvoice model remove base.en
```
### Keybinding Pattern
Most setups use this toggle pattern in window manager config:
@@ -215,8 +237,8 @@ hyprvoice configure
The wizard guides you through all settings with a user-friendly interface:
- **Providers** - API keys for OpenAI, Groq, Mistral, ElevenLabs
- **Transcription** - Speech-to-text provider and model selection
- **Providers** - API keys for OpenAI, Groq, Mistral, ElevenLabs, Deepgram
- **Transcription** - Speech-to-text provider and model selection (cloud or local)
- **LLM** - Post-processing to clean up transcriptions (enabled by default)
- **Keywords** - Domain-specific terms for better accuracy
- **Injection** - How text is typed (ydotool, wtype, clipboard)
@@ -227,6 +249,97 @@ Configuration is stored in `~/.config/hyprvoice/config.toml`. Changes are applie
For manual configuration and detailed options, see [docs/config.md](docs/config.md).
## Local Transcription
For complete offline privacy, use whisper.cpp for local transcription - no API keys, no cloud, no data leaves your machine.
### Prerequisites
1. **Install whisper.cpp**: Build from source or install via package manager
```bash
# Arch Linux
yay -S whisper.cpp
# Build from source (recommended for CUDA/Metal support)
git clone https://github.com/ggerganov/whisper.cpp
cd whisper.cpp && make
sudo cp main /usr/local/bin/whisper-cli
```
2. **Download a model**:
```bash
# List available models
hyprvoice model list --provider whisper-cpp
# Download recommended model (142MB, English-only, fast)
hyprvoice model download base.en
# Or download multilingual model (142MB, 57 languages)
hyprvoice model download base
```
### Available Models
| Model | Size | Languages | Speed | Accuracy |
| ---------- | ----- | ----------- | -------- | -------- |
| tiny.en | 75MB | English | Fastest | Good |
| base.en | 142MB | English | Fast | Better |
| small.en | 466MB | English | Medium | Great |
| medium.en | 1.5GB | English | Slow | Excellent|
| tiny | 75MB | 57 langs | Fastest | Good |
| base | 142MB | 57 langs | Fast | Better |
| small | 466MB | 57 langs | Medium | Great |
| medium | 1.5GB | 57 langs | Slow | Excellent|
| large-v3 | 3GB | 57 langs | Slowest | Best |
**Recommendation**: Start with `base.en` for English or `base` for multilingual. Models ending in `.en` are English-only but slightly faster.
### Configuration
```toml
[transcription]
provider = "whisper-cpp"
model = "base.en" # or "base" for multilingual
language = "" # empty for auto-detect
threads = 0 # 0 = auto (NumCPU - 1)
```
## Streaming Transcription
For real-time transcription results as you speak, use streaming providers. Text appears progressively instead of waiting for the entire recording to finish.
### Streaming Providers
| Provider | Models | Latency | Languages |
| ---------- | -------------------------- | ---------- | --------- |
| ElevenLabs | scribe_v1-streaming, scribe_v2-streaming | ~150ms | 57 langs |
| Deepgram | nova-3, nova-2 | ~100ms | 40+ langs |
| OpenAI | gpt-4o-realtime-preview | ~200ms | 57 langs |
### Configuration
```toml
# ElevenLabs streaming
[providers.elevenlabs]
api_key = "..."
[transcription]
provider = "elevenlabs"
model = "scribe_v2-streaming"
# Deepgram streaming
[providers.deepgram]
api_key = "..."
[transcription]
provider = "deepgram"
model = "nova-3"
```
**Note**: Streaming models show partial results while recording. Final text is accumulated and injected when you toggle off.
### Service Management
The systemd user service is automatically installed with the AUR package:
@@ -251,7 +364,8 @@ journalctl --user -u hyprvoice.service -f
- **Socket**: `~/.cache/hyprvoice/control.sock` - IPC communication
- **PID file**: `~/.cache/hyprvoice/hyprvoice.pid` - Process tracking
- **Config**: `~/.config/hyprvoice/config.toml` - User settings (planned)
- **Config**: `~/.config/hyprvoice/config.toml` - User settings
- **Models**: `~/.local/share/hyprvoice/models/whisper/` - Downloaded whisper models
## Development Status
@@ -261,10 +375,15 @@ journalctl --user -u hyprvoice.service -f
| Recording workflow | ✅ | Toggle recording via PipeWire |
| Audio capture | ✅ | Efficient PipeWire integration |
| Desktop notifications | ✅ | Status feedback via notify-send |
| OpenAI transcription | ✅ | HTTP API integration |
| OpenAI transcription | ✅ | HTTP API + Realtime streaming |
| Groq transcription | ✅ | Fast Whisper API with transcription and translation |
| Mistral transcription | ✅ | Voxtral API for European languages |
| ElevenLabs transcription | ✅ | Scribe API with 99 language support |
| ElevenLabs transcription | ✅ | Scribe batch + streaming (90+ languages) |
| Deepgram transcription | ✅ | Nova-3 streaming (40+ languages) |
| Local transcription | ✅ | whisper.cpp with model download management |
| Streaming support | ✅ | Real-time results with ElevenLabs, Deepgram, OpenAI |
| Model management | ✅ | `hyprvoice model list/download/remove` CLI |
| Language validation | ✅ | Model-language compatibility checking |
| LLM post-processing | ✅ | OpenAI/Groq text cleanup (enabled by default) |
| Text injection | ✅ | Clipboard + wtype/ydotool with fallback |
| Configuration system | ✅ | TOML-based user settings with hot-reload |
@@ -272,8 +391,6 @@ journalctl --user -u hyprvoice.service -f
| Unit test coverage | ✅ | Comprehensive test suite (100% pass) |
| CI/CD Pipeline | ✅ | Automated builds and releases via GitHub Actions |
| Installation (AUR etc) | ✅ | AUR package with automated dependency installation |
| Light dictation models | ⏳ | Alternatives to whispers for light and fast dictation |
| whisper.cpp support | ⏳ | Local model inference |
**Legend**: ✅ Complete · ⏳ Planned
@@ -498,13 +615,16 @@ hyprvoice/
│ ├── bus/ # IPC (Unix socket) + PID management
│ ├── config/ # Configuration loading and validation
│ ├── daemon/ # Control daemon (lifecycle management)
│ ├── deps/ # Dependency checking (whisper-cli, ffmpeg)
│ ├── injection/ # Text injection (clipboard + wtype + ydotool)
│ ├── language/ # Language codes and provider-specific mappings
│ ├── llm/ # LLM post-processing adapters (OpenAI, Groq)
│ ├── models/whisper/ # Whisper model info and download management
│ ├── notify/ # Desktop notification integration
│ ├── pipeline/ # Audio processing pipeline + state machine
│ ├── provider/ # Provider registry and capability detection
│ ├── provider/ # Provider registry with Model metadata
│ ├── recording/ # PipeWire audio capture
│ ├── transcriber/ # Transcription adapters (OpenAI, Groq, Mistral, ElevenLabs)
│ ├── transcriber/ # Batch and streaming adapters (OpenAI, Groq, Mistral, ElevenLabs, Deepgram, whisper-cpp)
│ └── tui/ # Interactive configuration wizard
├── go.mod # Go module definition
└── README.md