Model Download

Model Download

The ggml model is not bundled in the AAR — models range from 75 MB to 1.5 GB, far too large to ship inside a library. You download the one that fits your needs.

Which model?

Model	Size	Speed	Quality	Languages	Download
tiny.en	~75 MB	⚡⚡⚡ fastest	★★	English only	ggml-tiny.en.bin
tiny	~75 MB	⚡⚡⚡ fastest	★★	99 languages	ggml-tiny.bin
base.en	~142 MB	⚡⚡ fast	★★★	English only	ggml-base.en.bin
base	~142 MB	⚡⚡ fast	★★★	99 languages	ggml-base.bin
small	~466 MB	⚡ slower	★★★★	99 languages	ggml-small.bin

Recommendation: start with base (multilingual) or base.en (English only) — the best speed/quality balance on a phone. Drop to tiny.en for the fastest response on low-end devices; move up to small when accuracy matters more than latency.

Larger models = better transcription but slower and more RAM. On a mid-range phone, small can take several times real-time. Test on your target hardware.

All models

The full catalogue (including quantized variants) lives on Hugging Face: ggml-org/whisper.cpp.

How to ship the model

Three common strategies:

1. First-run download — download from your own server / Hugging Face on first launch, store in getExternalFilesDir("models"). Keeps the APK small.
2. Bundle in assets — put it in src/main/assets/models/ and load with Whisper.loadModelFromAsset(context, "models/ggml-base.bin"). Increases APK size.
3. Dev / testing — adb push to the app's files dir (see Quick Start).

Quantized models (Pro)

q4_0 / q5_1 / q8_0 quantized models are smaller and faster with a small quality trade-off. Quantized-model support is part of the Pro build — see jokobee.com.