> ## Documentation Index
> Fetch the complete documentation index at: https://docs.voicetypr.com/llms.txt
> Use this file to discover all available pages before exploring further.

# AI Models

> Understanding VoiceTypr's AI models, including Whisper and Parakeet support, model sizes, hardware acceleration, and management.

## Overview

VoiceTypr supports multiple AI transcription engines with various model sizes to balance speed and accuracy based on your needs. All models run locally on your device for complete privacy.

## Supported Engines

<CardGroup cols={3}>
  <Card title="Whisper" icon="robot">
    OpenAI's Whisper models provide excellent accuracy across 99+ languages with multiple size options.
  </Card>

  <Card title="Parakeet" icon="bird">
    NVIDIA Parakeet models optimized for Apple Silicon, offering fast transcription using the Neural Engine.
  </Card>

  <Card title="Soniox" icon="cloud">
    Cloud-based speech recognition API offering fast, accurate transcription without local model downloads.
  </Card>
</CardGroup>

## Model Types

VoiceTypr distinguishes between local and cloud models:

```typescript theme={null}
type SpeechModelEngine = 'whisper' | 'parakeet' | 'soniox';
type ModelKind = 'local' | 'cloud';

interface LocalModelInfo {
  kind: 'local';
  name: string;              // Internal identifier
  display_name: string;      // User-friendly name
  engine: SpeechModelEngine; // Which engine runs this model
  size: number;              // Download size in bytes
  speed_score: number;       // Speed rating (1-10)
  accuracy_score: number;    // Accuracy rating (1-10)
  recommended: boolean;      // Is this a recommended model?
  downloaded: boolean;       // Is it downloaded locally?
  requires_setup: boolean;   // Does it need configuration?
  url: string;              // Download URL
  sha256: string;           // Checksum for verification
}
```

## Whisper Models

### Available Sizes

Whisper models come in multiple sizes, each with different speed and accuracy tradeoffs:

<Tabs>
  <Tab title="Tiny">
    **Whisper Tiny** (\~75 MB)

    * **Speed Score**: 10/10 ⚡
    * **Accuracy Score**: 6/10
    * **Best for**: Quick drafts, testing, low-power devices
    * **Languages**: Multilingual

    The smallest and fastest model. Good for quick notes when accuracy isn't critical.
  </Tab>

  <Tab title="Base">
    **Whisper Base** (\~150 MB)

    * **Speed Score**: 9/10 ⚡
    * **Accuracy Score**: 7/10
    * **Best for**: Balanced everyday use
    * **Languages**: Multilingual

    A good balance of speed and accuracy for most use cases.
  </Tab>

  <Tab title="Small">
    **Whisper Small** (\~500 MB) - Recommended ⭐

    * **Speed Score**: 7/10
    * **Accuracy Score**: 8/10
    * **Best for**: General purpose transcription
    * **Languages**: Multilingual

    The recommended default. Excellent accuracy with acceptable speed.
  </Tab>

  <Tab title="Medium">
    **Whisper Medium** (\~1.5 GB)

    * **Speed Score**: 5/10
    * **Accuracy Score**: 9/10
    * **Best for**: High-accuracy transcription, technical content
    * **Languages**: Multilingual

    Better accuracy than Small, but slower. Good for professional use.
  </Tab>

  <Tab title="Large">
    **Whisper Large** (\~3 GB)

    * **Speed Score**: 3/10
    * **Accuracy Score**: 10/10
    * **Best for**: Maximum accuracy, complex audio
    * **Languages**: Multilingual

    The most accurate model but slowest. Best for difficult transcription tasks.
  </Tab>
</Tabs>

### English-Only Models

Whisper also offers English-only variants (`.en` suffix) that are optimized for English:

```typescript theme={null}
"tiny.en"   // Tiny English-only
"base.en"   // Base English-only  
"small.en"  // Small English-only
"medium.en" // Medium English-only
```

<Info>
  English-only models are smaller, faster, and more accurate **for English** compared to their multilingual counterparts.
</Info>

## Parakeet Models

Parakeet models are available on **macOS only** and leverage Apple's Neural Engine for hardware acceleration.

### Available Models

<Tabs>
  <Tab title="Parakeet 1.1B">
    **Parakeet 1.1B** (\~1.3 GB)

    * **Speed Score**: 8/10
    * **Accuracy Score**: 8/10
    * **Languages**: Multilingual (100+ languages)
    * **Hardware**: Apple Neural Engine

    Multilingual support with good performance on Apple Silicon.
  </Tab>

  <Tab title="Parakeet 1.1B v2">
    **Parakeet 1.1B v2** (\~1.3 GB)

    * **Speed Score**: 9/10
    * **Accuracy Score**: 9/10
    * **Languages**: English only
    * **Hardware**: Apple Neural Engine

    English-only variant optimized for Apple Neural Engine.
  </Tab>
</Tabs>

<Warning>
  Parakeet models are macOS-exclusive and require an Apple Silicon Mac (M1, M2, M3, or newer). They will not run on Intel Macs or other platforms.
</Warning>

## Soniox Cloud Models

Soniox is a **cloud-based** speech recognition service that provides fast, accurate transcription without requiring local model downloads.

### Overview

Unlike Whisper and Parakeet which run entirely on your device, Soniox processes audio in the cloud:

* **No downloads required**: No disk space needed for models
* **Fast transcription**: Cloud processing with optimized infrastructure
* **Requires internet**: Audio is sent to Soniox API for processing
* **API key required**: You need a Soniox account and API key

### Setup

To use Soniox models:

<Steps>
  <Step title="Get API Key">
    Sign up at [soniox.com](https://soniox.com) and obtain an API key
  </Step>

  <Step title="Configure in VoiceTypr">
    Add your API key in Settings → Models → Soniox Configuration
  </Step>

  <Step title="Select Soniox Model">
    Choose a Soniox model from the Models tab
  </Step>
</Steps>

### Available Models

Soniox offers several optimized models:

* **stt-async-v3**: Latest asynchronous model with best accuracy
* **stt-streaming**: Real-time streaming transcription
* **stt-multilingual**: Support for multiple languages

<Info>
  Check the [Soniox documentation](https://soniox.com/docs) for the latest available models and language support.
</Info>

### Privacy Considerations

<Warning>
  When using Soniox, audio is sent to Soniox servers for processing. This differs from Whisper and Parakeet which process entirely offline on your device.

  Only use Soniox if you're comfortable with cloud-based processing of your audio.
</Warning>

### Performance

Soniox typically provides:

* **Speed**: Very fast, limited by network latency
* **Accuracy**: High accuracy comparable to Whisper Large
* **Cost**: Based on Soniox API pricing

### API Key Storage

Soniox API keys are stored securely in your system keychain:

* **macOS**: Keychain Access
* **Windows**: Credential Manager

The key is never stored in plain text.

### Validation

VoiceTypr validates your Soniox API key by:

```typescript theme={null}
await invoke('validate_and_cache_soniox_key', {
  apiKey: 'your-soniox-api-key'
});
```

The validation checks against Soniox's `/v1/models` endpoint to verify the key is active.

## Hardware Acceleration

VoiceTypr automatically uses hardware acceleration when available for maximum performance.

### macOS

* **Whisper**: Uses **Metal GPU acceleration** via Apple's Metal Performance Shaders
* **Parakeet**: Uses **Apple Neural Engine** for ultra-fast inference
* **Requirements**: macOS 13.0+ (Ventura or later)

### Windows

* **Whisper**: Supports **GPU acceleration** via DirectML
* **Compatible GPUs**: NVIDIA, AMD, and Intel GPUs
* **Fallback**: Automatically uses CPU if GPU unavailable or drivers missing

<Tip>
  On Windows, ensure your graphics drivers are up to date for 5-10x faster transcription:

  * [NVIDIA Drivers](https://www.nvidia.com/drivers)
  * [AMD Drivers](https://www.amd.com/support)
  * [Intel Drivers](https://www.intel.com/content/www/us/en/support/products/80939/graphics.html)
</Tip>

## Model Management

### Downloading Models

<Steps>
  <Step title="Open Models Tab">
    Click the VoiceTypr menubar icon and go to the **Models** tab.
  </Step>

  <Step title="Browse Available Models">
    Models are organized into two sections:

    * **Available to Use**: Already downloaded and ready
    * **Available to Setup**: Need to be downloaded first
  </Step>

  <Step title="Download a Model">
    Click the **Download** button on any model. Progress is shown in real-time.
  </Step>

  <Step title="Verify and Activate">
    After download, the model is verified using SHA-256 checksum and automatically becomes available for use.
  </Step>
</Steps>

### Download Progress Tracking

Download progress is tracked in real-time:

```typescript theme={null}
downloadProgress: Record<string, number>  // modelName -> percentage (0-100)
```

You can cancel an in-progress download at any time:

```typescript theme={null}
cancelDownload(modelName: string)
```

### Model Verification

All downloaded models are verified using SHA-256 checksums to ensure integrity:

```typescript theme={null}
verifyingModels: Set<string>  // Models currently being verified
```

<Note>
  Model verification happens automatically after download. If verification fails, the download is considered corrupted and must be retried.
</Note>

### Deleting Models

To free up disk space, you can delete models you no longer need:

<Steps>
  <Step title="Navigate to Models">
    Open the **Models** tab in VoiceTypr.
  </Step>

  <Step title="Find Downloaded Model">
    Locate the model in the "Available to Use" section.
  </Step>

  <Step title="Delete">
    Click the delete/trash icon on the model card.
  </Step>
</Steps>

<Warning>
  If you delete the currently active model, VoiceTypr will clear your model selection. You'll need to select and download a new model before transcription will work.
</Warning>

### Model Selection

To switch between downloaded models:

1. Go to the **Models** tab
2. Click on any downloaded model to select it
3. The selected model is saved in settings:

```typescript theme={null}
current_model: string;           // Model name
current_model_engine: 'whisper' | 'parakeet' | 'soniox';
```

## Choosing the Right Model

Use this guide to select the best model for your needs:

<AccordionGroup>
  <Accordion title="I want the fastest possible transcription">
    **Recommended**: Whisper Tiny or Tiny.en

    * Fastest inference times
    * Good for quick notes and drafts
    * Trade-off: Lower accuracy
  </Accordion>

  <Accordion title="I need the best accuracy">
    **Recommended**: Whisper Large or Medium

    * Highest accuracy scores
    * Best for professional transcription
    * Trade-off: Slower processing

    Consider enabling [AI Enhancement](/features/ai-enhancement) to further improve output quality.
  </Accordion>

  <Accordion title="I want balanced performance (recommended)">
    **Recommended**: Whisper Small or Small.en ⭐

    * Excellent balance of speed and accuracy
    * Default recommended model
    * Good for most daily use cases
  </Accordion>

  <Accordion title="I have an Apple Silicon Mac">
    **Recommended**: Parakeet 1.1B v2 (English) or Parakeet 1.1B (Multilingual)

    * Optimized for Apple Neural Engine
    * Faster than Whisper on M-series chips
    * Great accuracy
  </Accordion>

  <Accordion title="I transcribe non-English languages">
    **Recommended**: Whisper Small, Medium, or Large (multilingual variants)

    * Support for 99+ languages
    * Avoid `.en` suffix models (English-only)
    * Parakeet 1.1B also supports 100+ languages
  </Accordion>
</AccordionGroup>

## Disk Space Requirements

Ensure you have enough free disk space before downloading models:

| Model          | Size     |
| -------------- | -------- |
| Whisper Tiny   | \~75 MB  |
| Whisper Base   | \~150 MB |
| Whisper Small  | \~500 MB |
| Whisper Medium | \~1.5 GB |
| Whisper Large  | \~3 GB   |
| Parakeet 1.1B  | \~1.3 GB |

<Info>
  You only need **one model** to use VoiceTypr. The recommended Whisper Small model requires just 500 MB.
</Info>

## Model Storage Location

Downloaded models are stored in:

* **macOS**: `~/Library/Application Support/com.voicetypr.app/models/`
* **Windows**: `%APPDATA%\com.voicetypr.app\models\`

## Troubleshooting

### Download Failed

1. Check your internet connection
2. Ensure you have enough free disk space
3. Try canceling and restarting the download
4. Check firewall/antivirus isn't blocking the download

### Verification Failed

If SHA-256 verification fails:

1. Delete the corrupted model
2. Retry the download
3. Check for disk errors if it continues failing

### Model Not Appearing

If a downloaded model doesn't appear:

1. Click "Refresh Models" in the Models tab
2. Restart VoiceTypr
3. Check the model storage location manually

### Slow Transcription

If transcription is slower than expected:

* **macOS**: Ensure you're using a model compatible with Metal acceleration
* **Windows**: Update your GPU drivers for hardware acceleration
* Try a smaller/faster model (Tiny or Base)
* Close resource-intensive applications