FAQ

FAQ

Which languages are supported?

99 languages via the multilingual models (tiny, base, small, …). The .en models are English-only but slightly faster/more accurate for English. Set the language explicitly with WhisperConfig(language = "fr"), or use "auto" for automatic detection.

Why is transcription slow?

Usually the model is too large for the device. Speed depends on:

• Model size — tiny ≫ base ≫ small in speed. Try a smaller model.
• Threads — set WhisperConfig(threads = N) to match the CPU (4 is a good default).
• Audio length — processing time scales with audio duration.
• Device — a budget CPU with small can run several times slower than real-time.

Rule of thumb: use base/base.en on phones; reserve small for short clips or high-end devices.

Does it need internet?

No. Everything runs on-device. No network, no cloud, no API keys. The only online step is downloading the model file once.

What audio format is required?

16 kHz, mono, 16-bit PCM WAV. Resample other formats first (e.g. ffmpeg -i in.mp3 -ar 16000 -ac 1 -c:a pcm_s16le out.wav).

How much RAM does it use?

Roughly the model size plus working memory — e.g. base needs a few hundred MB. Larger models need proportionally more; very large models may OOM on low-RAM devices.

Can I transcribe from the microphone in real time?

Not in the Free tier (file-based only). Real-time streaming + VAD are in the Pro build — see jokobee.com.

Which ABIs are supported?

Free tier: arm64-v8a only (>90% of modern devices). x86_64 (emulators, Chromebooks) is in the Pro build.

Is it really free?

Yes — MIT licensed, free on Maven Central and JitPack. The Pro build adds streaming, VAD, multi-ABI and quantized models.