Guides › Switching Whisper Model Sizes for OpenCode
Switching Whisper Model Sizes for OpenCode
Whisper.cpp offers multiple model sizes. The base model is fast but less accurate. Larger models improve accuracy at the cost of speed and memory.
1
Base model (142MB)
Default. Fast transcription (~1s for 5s audio). Best for: quick corrections, low-resource machines, testing. Accuracy: adequate for clear speech.
2
Small model (466MB)
Recommended upgrade. Good transcription speed (~2s). Best for: daily use, mild accents, typical coding vocabulary. Dramatically better than base.
3
Medium model (1.5GB)
High accuracy. Slower transcription (~4s). Best for: strong accents, technical terminology, noisy environments. Uses ~2GB RAM during transcription.
4
Large model (3GB)
Maximum accuracy. Slowest (~8s). Best for: critical transcriptions, heavy accents, specialized vocabulary. Requires ~4GB free RAM.
Tips
• Download a model with: MODEL=small bash scripts/install-whisper.sh. The script downloads the model file to ~/.local/bin/.
• Set the model via opencode.json: { "whisperModel": "small" } or env var: WHISPER_MODEL=...
• OpenAI API (whisper-1) is more accurate than even the large model and requires no local resources.
Troubleshooting
Models are large files. Ensure sufficient disk space: base 142MB, small 466MB, medium 1.5GB, large 3GB. If downloads fail, check your internet connection and disk space.