Whisper models

Everything you need to know about choosing and configuring Whisper models.

Overview

Clanker Yap uses whisper.cpp models via the whisper-rs Rust bindings. These are GGML-format models optimized for local inference.

Model Files

Format

GGML (Gate Generalized Model Language) is a tensor format optimized for:

• Fast inference on CPU
• Quantization support (INT4, INT8)
• No GPU required

File Extension

.bin files contain:

• Model architecture metadata
• Quantized weights
• Vocabulary mapping

Storage Location

Default location: ~/.local/share/voice-transcribe/

Available Models

English-Only Models

*Word Error Rate on clean speech

Multilingual Models

Choosing a Model

Decision Factors

┌─────────────────────────────────────────────────────────────┐
│                    Model Selection Flow                      │
├─────────────────────────────────────────────────────────────┤
│                                                              │
│   English only? ──YES──► Use ".en" models (smaller)         │
│         │                                                        │
│         NO                                                       │
│         │                                                        │
│         ▼                                                        │
│   Need speed? ──YES──► Base or Small                          │
│         │                                                        │
│         NO                                                       │
│         │                                                        │
│         ▼                                                        │
│   Need accuracy? ──YES──► Medium or Large                     │
│                                                              │
└─────────────────────────────────────────────────────────────┘

Recommendations by Use Case

Downloading Models

Via Application UI

1. Open Clanker Yap
2. Click the Model toolbar button
3. Click "Download base.en"
4. Wait for completion (~148 MB)

Via Command Line

Using curl:

# Create directory
mkdir -p ~/.local/share/voice-transcribe

# Download base English model
curl -L "https://huggingface.co/ggml-org/whisper.cpp/resolve/main/ggml-base.en.bin?download=true" \
  -o ~/.local/share/voice-transcribe/ggml-base.en.bin

# Download small English model
curl -L "https://huggingface.co/ggml-org/whisper.cpp/resolve/main/ggml-small.en.bin?download=true" \
  -o ~/.local/share/voice-transcribe/ggml-small.en.bin

Using wget:

wget -O ~/.local/share/voice-transcribe/ggml-base.en.bin \
  "https://huggingface.co/ggml-org/whisper.cpp/resolve/main/ggml-base.en.bin?download=true"

HuggingFace Mirrors

Official source: https://huggingface.co/ggml-org/whisper.cpp/tree/main

Alternative mirrors may be available. Search for "ggml whisper" repositories on GitHub for community mirrors.

Model Validation

Verify Download

Check file size and hash:

# Check file size (should be ~148 MB for base.en)
ls -lh ~/.local/share/voice-transcribe/ggml-base.en.bin

# Verify SHA256 (optional)
sha256sum ~/.local/share/voice-transcribe/ggml-base.en.bin

Test Model in App

1. Open Clanker Yap
2. Ensure model path is set correctly in Model settings
3. Hold hotkey and speak
4. Verify transcription appears

Error: "Model not found"

This means the model file doesn't exist at the configured path.

Solutions:

1. Verify file exists: ls -la ~/.local/share/voice-transcribe/
2. Check Settings model_path matches actual file location
3. Re-download if corrupted

Custom Model Paths

Setting Custom Path

1. Click Model in toolbar
2. Click "Edit"
3. Enter full path to your model file
4. Click "Save"

Example Custom Paths

Model Performance Tips

Improving Transcription Quality

1. Speak clearly - Whisper handles some noise but benefits from clear audio
2. Use headphones - Prevents feedback loops
3. Close background apps - Reduces CPU contention
4. Choose appropriate model - Larger models handle accents better

Optimizing Speed

1. Use smaller models - base is fastest
2. Close other applications - More CPU available for inference
3. Keep model in memory - First transcription may be slower
4. Use SSD storage - Faster model loading

Model Updates

New Model Versions

Whisper.cpp periodically releases improved models:

• ggml-base.bin → improved base model
• ggml-large-v2 → v3 migration

To update:

1. Download new model
2. Update model path in Settings
3. Old model can be deleted

Version Compatibility

Clanker Yap v0.1.0 is compatible with:

• All whisper.cpp GGML models (v1.0+)
• Both .en and multilingual variants
• INT8 and INT4 quantized models

Troubleshooting Models