https://github.com/yixin0829/push-to-talk
Ultra-fast, customizable AI voice dictation in any active app on Windows (MacOS and Linux coming soon)
https://github.com/yixin0829/push-to-talk
cerebras deepgram dictation dictation-tool speech-to-text voice-dictation voice-input whisper
Last synced: about 1 month ago
JSON representation
Ultra-fast, customizable AI voice dictation in any active app on Windows (MacOS and Linux coming soon)
- Host: GitHub
- URL: https://github.com/yixin0829/push-to-talk
- Owner: yixin0829
- License: mit
- Created: 2025-07-26T18:32:05.000Z (10 months ago)
- Default Branch: main
- Last Pushed: 2026-03-08T19:11:55.000Z (3 months ago)
- Last Synced: 2026-03-08T22:22:55.662Z (3 months ago)
- Topics: cerebras, deepgram, dictation, dictation-tool, speech-to-text, voice-dictation, voice-input, whisper
- Language: Python
- Homepage: https://yixin0829.github.io/push-to-talk/
- Size: 3.54 MB
- Stars: 22
- Watchers: 1
- Forks: 6
- Open Issues: 16
-
Metadata Files:
- Readme: README.md
- Changelog: CHANGELOG.md
- Contributing: CONTRIBUTING.md
- License: LICENSE
- Agents: AGENTS.md
Awesome Lists containing this project
README
# PushToTalk - AI Refined Speech-to-Text Dictation
[](https://codecov.io/gh/yixin0829/push-to-talk)
A Python application that provides push-to-talk speech-to-text functionality with AI speech to text transcription, smart text refinement, and automatic text insertion into the active window on Windows, MacOS (coming soon), and Linux (coming soon). **Now features a persistent GUI configuration interface with real-time status management and easy application control.**
**Watch 60s Demo:** https://www.youtube.com/watch?v=PIqToMHP-c4
**How to download (FOR FREE):** https://www.youtube.com/watch?v=CjbcRMQE3jE
## 🏆 Why PushToTalk?
**Compared to Wispr Flow ($14.40/month) and Superwhisper ($8.49/month), PushToTalk delivers:**
| Feature | PushToTalk | Others |
|---------|-----------|--------|
| **Cost** | **FREE** + pay-as-you-go API | $102-216/year subscription |
| **STT Speed** | **Ultra-fast** (Deepgram Nova-3 and Cerebras) | Standard latency |
| **Customization** | Multi-provider choice, custom glossary, custom prompts | Limited vendor lock-in |
| **Privacy** | Full inference provider control, choose your API including Ollama | Proprietary ecosystem |
**For power users, developers, and privacy-conscious teams: PushToTalk is the most customizable, fastest, and nearly free alternative to existing AI dictation tools.**
## Features
- **🎯 GUI Interface**: Integrated configuration control and application status monitoring in one window
- **🔄 Live Config Sync**: GUI edits instantly push updates to the running background service—no restart required
- **📚 Custom Glossary**: Add domain-specific terms and acronyms to improve transcription accuracy
- **✨ Multi-Provider Text Refinement**: Improves transcription quality using OpenAI GPT, Cerebras, Gemini, or custom OpenAI-compatible endpoints
- **🤖 Multi-Provider Speech-to-Text**: Choose between OpenAI Whisper or Deepgram for accurate transcription
- **🎤 Push-to-Talk Recording**: Hold a customizable hotkey to record audio (platform-aware defaults)
- **📝 Auto Text Insertion**: Automatically inserts refined text into the active window
## Release Demos
- v0.6.0: https://www.youtube.com/watch?v=A6Lyk7w0Lzs
- v0.5.0: https://app.canvid.com/share/fi_01KDGEDBA86537CP18E9XDDJY8
- v0.3.0 - v0.4.0: https://www.loom.com/share/71ecc05d4bb440ecb708d980505b9000
- v0.2.0: https://www.loom.com/share/fbabb2da83c249d3a0a09b3adcc4a4e6
## Roadmap
See [issues](https://github.com/yixin0829/push-to-talk/issues) for more details.
## Requirements
- [uv](https://docs.astral.sh/uv/) (Python package manager)
- **At least one STT API key**:
- OpenAI Whisper: https://platform.openai.com/docs/api-reference/introduction
- Deepgram: https://deepgram.com/
- **Text refinement (optional)**:
- OpenAI GPT: https://platform.openai.com/docs/api-reference/introduction
- Cerebras: https://cerebras.ai/
- Google Gemini: https://ai.google.dev/
- Custom OpenAI-compatible endpoints (e.g., Ollama, Together AI)
- Microphone access (for recording)
## Quick Start (GUI Application)
### For End Users (Recommended)
1. **Download and launch**:
- Download `PushToTalk.exe` from releases
- Double-click to launch the configuration interface
2. **One-window setup and control**:
- **Welcome section** explains the application at the top
- **Configure your settings** in the organized sections below
- **Click "Start Application"** to begin - the GUI stays open
- **Monitor status** with real-time indicators (green = running, gray = stopped)
- **View active settings** displayed when running
- **Easy control** with "Stop Application" button to terminate
- **Tweak settings live**—changes made while running apply instantly
3. **Daily usage**:
- GUI provides persistent control and status monitoring
- Use your configured hotkeys to record and transcribe
- Tune preferences without stopping the service; hotkeys update in real time
- Start/stop the service anytime from the GUI
### For Developers
1. **Clone the repository**:
```bash
git clone
cd push-to-talk
```
2. **Install dependencies**:
```bash
uv sync
```
3. **Run the GUI application**:
```bash
uv run python main.py # Normal mode (logs to file only)
uv run python main.py --debug # Debug mode (logs to console and file)
```
The `--debug` flag enables console logging for real-time debugging and troubleshooting.
## GUI Configuration Interface
The application features a comprehensive, persistent configuration GUI with organized sections:
### Welcome & Status
- **Real-Time Status**: Visual indicators show current application state
- **Gray circle + "Ready to start"**: Application stopped
- **Green circle + "Running - Use your configured hotkeys"**: Application running
- **Active Settings Display**: Shows current hotkeys and enabled features when running
- **Live Updates Banner**: Status automatically refreshes when settings change mid-session
### Live Configuration Updates
The application features a sophisticated real-time configuration system that applies changes instantly while running:
#### How It Works
- **Variable Tracing**: Every GUI field (text boxes, checkboxes, dropdowns) automatically detects changes using Tkinter variable traces
- **Smart Debouncing**: Rapid typing is intelligently handled with configurable delay to prevent excessive updates
- **Selective Reinitialization**: Only components affected by changes are reinitialized (e.g., hotkey changes → restart hotkey service)
- **Service Continuity**: Critical services like hotkey detection automatically restart after updates
#### Technical Features
- **Instant Propagation**: Editing any field triggers a debounced update to the running PushToTalk service
- **Persistent Storage**: Changes are automatically saved to JSON file asynchronously for permanent persistence
- **Callback Support**: Optional listeners receive validated Pydantic configuration models whenever values change
- **Glossary Sync**: Glossary edits are copied before rebuilds to prevent UI/model divergence
- **Safe Programmatic Updates**: GUI refreshes suspend traces to avoid infinite callback loops
- **Thread-Safe Saves**: Non-blocking background saves with deduplication prevent file conflicts
#### Example Scenarios
- **Hotkey Change**: Type "ctrl+alt+space" → Only final result triggers one hotkey service restart
- **Non-Critical Change**: Toggle "Audio Feedback" → Updates instantly without restarting core components
- **API Key Change**: Update OpenAI key → Only transcription/refinement components reinitialize
### Speech-to-Text Settings
- **STT Provider Selection**: Choose between OpenAI or Deepgram (default: Deepgram)
- **API Key**: Secure entry with show/hide functionality (dynamically shows OpenAI or Deepgram field based on provider)
- **Model Selection**: Choose provider-specific models:
- **OpenAI**: whisper-1, gpt-4o-transcribe, gpt-4o-mini-transcribe
- **Deepgram**: nova-3 (default and recommended), nova-2, base, enhanced, whisper-medium
### Text Refinement Settings
- **Refinement Provider**: Choose between OpenAI, Cerebras, Gemini, or Custom (default: Cerebras)
- **Refinement Model**: Provider-specific models:
- **OpenAI**: gpt-5, gpt-5-mini, gpt-5-nano, gpt-4.1, gpt-4.1-mini, gpt-4.1-nano, gpt-4o-mini, gpt-4o
- **Cerebras**: llama-3.3-70b (default), qwen-3-235b-a22b-instruct-2507, qwen-3-32b, llama3.1-8b
- **Gemini**: gemini-3-flash-preview (default), gemini-3-pro-preview, gemini-2.5-flash-preview-05-20, gemini-2.5-pro-preview-06-05
- **Custom**: llama3, mistral, mixtral, gemma (or any model name for your endpoint)
### Hotkey Configuration
- **Push-to-Talk Hotkey**: Hold to record (default: Ctrl+Space+^ on Windows/Linux, Cmd+Space+^ on macOS)
- **Toggle Recording Hotkey**: Press once to start/stop (default: Ctrl+Cmd on Windows/Linux, Cmd+Cmd on macOS)
- **Record Button**: Click "Record" and press keys one at a time to capture hotkey combinations
- **Validation**: Prevents duplicate hotkey assignments and ensures hotkeys are different
### Custom Glossary
- **Domain-Specific Terms**: Add specialized vocabulary, acronyms, and proper names
- **Easy Management**: Add, edit, and delete glossary terms through the GUI
- **Search Functionality**: Quickly find and manage existing terms
- **Automatic Integration**: Glossary terms are automatically included in transcription refinement
### Custom Refinement Prompt
- **Customizable System Prompt**: Create your own text refinement instructions
- **Glossary Placeholder**: Use `{custom_glossary}` placeholder to include glossary terms in your prompt
- **Copy Default Buttons**: Start from default prompts (with or without glossary) as templates
- **Reference Section**: View default prompts in collapsible reference panel
- **Live Updates**: Changes apply immediately to running application
## How to Use
1. **Build**: Build the application using `build_script\build.bat` if first time running on Windows
2. **Launch**: Double-click the built `PushToTalk.exe` or run `uv run python main.py`
3. **Configure**: Use the integrated setup interface with welcome guidance
4. **Start**: Click "Start Application" - GUI stays open with status indicators
5. **Monitor**: Watch real-time status and active settings display
6. **Use**: Background operation with your configured hotkeys
7. **Control**: Use "Stop Application" button to terminate, or restart anytime
## Building the Application
```bash
build_script\build.bat
```
This creates `dist\PushToTalk.exe` - a standalone GUI application.
## Configuration
The application supports both GUI and file-based configuration with automatic environment variable fallback:
### API Key Management
The application supports three ways to provide API keys (checked in this order):
1. **GUI**: Enter API keys directly in the configuration interface (stored in `push_to_talk_config.json`)
2. **Environment Variables**: Set `OPENAI_API_KEY`, `DEEPGRAM_API_KEY`, `CEREBRAS_API_KEY`, or `GOOGLE_API_KEY`
3. **Configuration File**: Manually edit `push_to_talk_config.json`
The required API key depends on your selected providers:
- **STT Provider**: OpenAI requires `openai_api_key`, Deepgram requires `deepgram_api_key`
- **Refinement Provider**: OpenAI requires `openai_api_key`, Cerebras requires `cerebras_api_key`, Gemini requires `gemini_api_key`, Custom requires `custom_api_key`
Environment variables are checked automatically if GUI or config file values are empty.
### Via GUI (Recommended)
- Launch the application to access the integrated configuration interface
- **All settings** validated and saved automatically to `push_to_talk_config.json`
- **Real-time status** shows application state with visual indicators
- **Auto-sync**: Edits instantly update the running background service and any registered callbacks
- API keys are validated on startup; application will show clear error messages if required keys are missing
### File-Based Configuration
The application creates a `push_to_talk_config.json` file. Example configuration file:
```json
{
"stt_provider": "deepgram",
"openai_api_key": "",
"deepgram_api_key": "your_deepgram_key",
"stt_model": "nova-3",
"refinement_provider": "cerebras",
"refinement_model": "llama-3.3-70b",
"cerebras_api_key": "your_cerebras_key",
"gemini_api_key": "",
"custom_api_key": "",
"custom_endpoint": "",
"sample_rate": 16000,
"chunk_size": 1024,
"channels": 1,
"hotkey": "ctrl+space+^",
"toggle_hotkey": "ctrl+cmd",
"enable_text_refinement": true,
"enable_logging": true,
"enable_audio_feedback": true,
"debug_mode": false,
"custom_glossary": ["API", "OAuth", "microservices", "PostgreSQL"],
"custom_refinement_prompt": ""
}
```
#### Configuration Parameters
| Parameter | Type | Default | Description |
|-----------|------|---------|-------------|
| `stt_provider` | string | `"deepgram"` | Speech-to-text provider. Options: `openai`, `deepgram`. Determines which transcription service to use. |
| `openai_api_key` | string | `""` | Your OpenAI API key for Whisper services. Required when using OpenAI provider. Can be set via GUI, config file, or `OPENAI_API_KEY` environment variable. |
| `deepgram_api_key` | string | `""` | Your Deepgram API key for transcription services. Required when using Deepgram provider. Can be set via GUI, config file, or `DEEPGRAM_API_KEY` environment variable. |
| `stt_model` | string | `"nova-3"` | STT Model for speech-to-text. For OpenAI: `whisper-1`, `gpt-4o-transcribe`, `gpt-4o-mini-transcribe`. For Deepgram: `nova-3`, `nova-2`, `base`, `enhanced`, `whisper-medium`. |
| `refinement_provider` | string | `"cerebras"` | Text refinement provider. Options: `openai`, `cerebras`, `gemini`, `custom`. Determines which AI service refines transcribed text. |
| `refinement_model` | string | `"llama-3.3-70b"` | Refinement Model for text refinement. For OpenAI: `gpt-5`, `gpt-5-mini`, `gpt-5-nano`, `gpt-4.1`, `gpt-4.1-mini`, `gpt-4.1-nano`, `gpt-4o-mini`, `gpt-4o`. For Cerebras: `llama-3.3-70b`, `qwen-3-235b-a22b-instruct-2507`, `qwen-3-32b`, `llama3.1-8b`, `gpt-oss-120b`. For Gemini: `gemini-3-flash-preview`, `gemini-3-pro-preview`, `gemini-2.5-flash-preview-05-20`, `gemini-2.5-pro-preview-06-05`. For Custom: any model name. |
| `cerebras_api_key` | string | `""` | Your Cerebras API key for text refinement. Required when using Cerebras provider. Can be set via GUI, config file, or `CEREBRAS_API_KEY` environment variable. |
| `gemini_api_key` | string | `""` | Your Google Gemini API key for text refinement. Required when using Gemini provider. Can be set via GUI, config file, or `GOOGLE_API_KEY` environment variable. |
| `custom_api_key` | string | `""` | Your custom API key for text refinement. Required when using Custom provider. |
| `custom_endpoint` | string | `""` | Custom API endpoint URL for OpenAI-compatible APIs (e.g., `http://localhost:11434/v1` for Ollama). Required when using Custom provider. |
| `sample_rate` | integer | `16000` | Audio sampling frequency in Hz. 16kHz is optimal for speech recognition. |
| `chunk_size` | integer | `1024` | Audio buffer size in samples. Determines how much audio is read at once (affects latency vs performance). |
| `channels` | integer | `1` | Number of audio channels. Use `1` for mono recording (recommended for speech). |
| `hotkey` | string | `"ctrl+space+^"` | Hotkey combination for push-to-talk. Platform-aware defaults: Windows/Linux `ctrl+space+^`, macOS `cmd+space+^`. See [Hotkey Options](#hotkey-options) for examples. |
| `toggle_hotkey` | string | `"ctrl+cmd"` | Hotkey combination for toggle recording mode. Press once to start, press again to stop. Platform-aware defaults: Windows/Linux `ctrl+cmd`, macOS `cmd+cmd`. |
| `enable_text_refinement` | boolean | `true` | Whether to refine transcribed text using AI. Disable for faster processing without refinement. |
| `enable_logging` | boolean | `true` | Whether to enable detailed logging to `push_to_talk.log` file using loguru. |
| `enable_audio_feedback` | boolean | `true` | Whether to play audio cues when starting/stopping recording. Provides immediate feedback for hotkey interactions. |
| `debug_mode` | boolean | `false` | Whether to enable debug mode. When enabled, recorded audio files are saved to timestamped debug directories (e.g., `debug_audio_20231215_143022_456/`) with recording metadata for troubleshooting. |
| `custom_glossary` | array | `[]` | List of domain-specific terms, acronyms, and proper names to improve transcription accuracy. Terms are automatically included in text refinement prompts. |
| `custom_refinement_prompt` | string | `""` | Custom system prompt for text refinement. Leave empty to use default prompts. Use `{custom_glossary}` placeholder to include glossary terms dynamically. |
#### Audio Quality Settings
- **sample_rate**:
- `16000` (16kHz) - Recommended for speech (Whisper optimized)
- `8000` (8kHz) - Lower quality but faster processing
- `44100` (44.1kHz) - CD quality (overkill for speech, slower)
- **chunk_size**:
- `512` - Lower latency, more CPU overhead
- `1024` - Balanced (recommended)
- `2048` - Higher latency, less CPU usage
- **channels**:
- `1` - Mono recording (recommended for speech)
- `2` - Stereo recording (unnecessary for speech-to-text)
### Hotkey Options
You can configure different hotkey combinations for both modes. Platform-specific defaults ensure compatibility:
**Push-to-talk hotkey** (hold to record):
- Windows/Linux: `ctrl+shift+^` (default), `ctrl+alt+r`, `f12`
- macOS: `cmd+shift+space` (default), `cmd+alt+r`, `f12`
**Toggle hotkey** (press once to start, press again to stop):
- Windows/Linux: `ctrl+shift+space` (default), `ctrl+shift+t`
- macOS: `cmd+shift+^` (default), `cmd+shift+t`
Both hotkeys support any combination from the `pynput` library. The application automatically uses platform-aware defaults based on your OS.
### Custom Glossary
The application supports custom glossary terms to improve transcription accuracy for domain-specific vocabulary:
- **Add Terms**: Use the GUI to add specialized vocabulary, acronyms, and proper names
- **Automatic Integration**: Glossary terms are automatically included in text refinement prompts
- **Smart Processing**: TextRefiner switches between prompts with and without glossary automatically
- **Persistent Storage**: Terms are saved in the configuration file and restored on startup
- **Easy Management**: Add, edit, delete, and search glossary terms through the GUI interface
**Example use cases**:
- Technical terms: "API", "OAuth", "microservices", "PostgreSQL"
- Company names: "Anthropic", "OpenAI"
- Acronyms: "CEO", "CTO", "AI/ML", "SaaS"
- Domain-specific vocabulary: Medical terms, legal terminology, etc.
### Audio Feedback
The application includes clean and simple audio feedback:
- **Recording Start**: A high-pitched beep sound that signals recording has begun
- **Recording Stop**: A lower-pitched confirmation beep that confirms recording completion
- **Non-Blocking**: Audio playback runs asynchronously to avoid interfering with recording or transcription
- **Configurable**: Can be toggled on/off via GUI or configuration JSON file
- **Cross-Platform**: Uses `playsound3` for audio playback with pre-recorded WAV files
## Architecture
```mermaid
sequenceDiagram
participant Main as Main Thread (GUI)
participant Hotkey as Hotkey Service Thread
participant Worker as Worker Thread
participant Audio as Audio Recording Thread
participant BG as Background Processing Thread
Note over Main,BG: Application Startup
Main->>Main: Initialize GUI
Main->>+Hotkey: Start Hotkey Service
Hotkey-->>Main: Service running
Main->>+Worker: Start Worker Thread
Worker-->>Main: Worker running
Note over Main,BG: Start Recording (User Presses Hotkey)
Hotkey->>Worker: Queue "START_RECORDING"
Worker->>Worker: Play start feedback (non-blocking)
Worker->>+Audio: Start audio recording
Audio-->>Worker: Recording started
Audio->>Audio: Capture audio loop
Note over Main,BG: Stop Recording (User Releases Hotkey)
Hotkey->>Worker: Queue "STOP_RECORDING"
Worker->>Worker: Play stop feedback (non-blocking)
Worker->>Audio: Stop recording
Audio-->>Worker: Audio file returned
Worker->>+BG: Spawn background thread
Note over Main,BG: Background Processing (Non-Blocking)
BG->>BG: Transcribe Audio (API)
BG->>BG: Refine Text (API)
BG->>BG: Insert Text (Clipboard)
BG->>BG: Cleanup temp files
Note over Main,BG: User Can Start New Recording Immediately
Hotkey->>Worker: Queue "START_RECORDING"
Worker->>+Audio: Start new recording
Note over BG: Previous transcription continues...
Note over Main,BG: Application Shutdown
Main->>Hotkey: Stop service
Hotkey-->>Main: Service stopped
Main->>Worker: Queue "QUIT"
Main->>Main: Wait for background threads
Worker-->>Main: Worker stopped
```
### Threading Model
The application uses separate threads for responsiveness and non-blocking operation:
- **Main Thread**: GUI and configuration
- **Hotkey Service Thread**: Global keyboard detection (producer) → queues commands
- **Worker Thread**: Command consumer → handles start/stop recording and audio feedback
- **Audio Recording Thread**: Real-time PyAudio buffering
- **Background Processing Threads**: Per-recording daemon threads for transcription → refinement → text insertion
**Key Design**: Each recording spawns its own background processing thread, allowing users to start a new recording immediately after stopping the previous one. The transcription and refinement API calls run in parallel in the background, reducing perceived latency from 3-5 seconds to ~100ms. Hotkeys are detected instantly even during API calls because the hotkey service operates independently from the worker thread.
See [AGENTS.md](AGENTS.md) for detailed threading implementation.
## Troubleshooting
### GUI Application Issues
1. **Application won't start**:
- Make sure you're running as Administrator for hotkey detection
- Check that the executable isn't blocked by antivirus
- Try running from command line to see error messages
2. **Status indicators not updating**:
- The GUI should show real-time status changes when starting/stopping
- If stuck, try restarting the application
- Check `push_to_talk.log` for error details
3. **Start/Stop button not working**:
- Ensure all required fields are filled (API keys based on selected providers)
- For Deepgram STT: Need `deepgram_api_key`
- For OpenAI STT: Need `openai_api_key`
- For Cerebras refinement: Need `cerebras_api_key`
- For OpenAI refinement: Need `openai_api_key`
- For Gemini refinement: Need `gemini_api_key`
- For Custom refinement: Need `custom_api_key` and `custom_endpoint`
- Use "Test Configuration" to validate settings
- Check that no other instance is running
4. **Settings not saving**:
- Ensure the application has write permissions in its directory
- Check that the configuration file isn't marked as read-only
- Try running as Administrator
### Common Issues
1. **"Could not find PyAudio"** (Development):
- Install PyAudio: `uv add pyaudio`
- On Windows, you may need Visual C++ build tools
2. **Hotkey not working**:
- Run as administrator (required for global hotkey detection)
- Check if another application is using the same hotkey
- Try a different hotkey combination in the GUI
- Ensure the application shows "Running" status in the GUI
3. **API errors or missing credentials**:
- Use the "Test Configuration" button in the GUI to validate settings
- Verify your API keys are valid and have sufficient credits/usage quota
- Check that your accounts have access to the models you're using
- For Deepgram: Verify Deepgram API key and subscription plan
- For Cerebras: Verify Cerebras API key and account status
- For Gemini: Verify Google API key and Gemini API access
- For Custom: Verify endpoint URL is accessible and API key is valid
- Ensure internet connectivity
- Check `push_to_talk.log` for specific API error messages
4. **Text not inserting**:
- Make sure the target window is active and has a text input field
- Check Windows permissions for clipboard access
5. **GUI appearance issues**:
- Try restarting the application
- Check display scaling settings (recommended: 100-150%)
- Ensure Windows is up to date
### Logging
Logs are written to `push_to_talk.log` using loguru's enhanced formatting. The GUI application logs only to file for cleaner user experience.
## Testing
For comprehensive testing information, see [CONTRIBUTING.md](CONTRIBUTING.md#testing).
**Quick Start:**
```bash
uv run pytest tests/ -v # All tests
uv run pytest tests/ --cov=src --cov-report=html # With coverage
```