Omnivox

Omnivox is the local speech model used by a Gboard-style Android keyboard with offline dictation: an offline, on-device Whisper engine for an Android keyboard. It runs entirely on the phone through whisper.cpp + JNI, with no network calls and no dependency on Google's speech service.

The base weights are the standard ggml-base model from ggerganov/whisper.cpp (OpenAI Whisper base, MIT). This release packages them together with the project's local correction dictionaries so the whole offline pipeline can be reproduced in one download. The base model is multilingual, so Omnivox is not tied to a single language; Portuguese and English are the languages exercised by the keyboard.

Contents

File Purpose
ggml-base.bin Whisper base weights in ggml format (~141 MB)
common-words.txt High-frequency common words for the local corrector
pt-br-dictionary.txt pt-BR frequency dictionary (~3k words)
pt-br-lm.txt Local unigram/bigram language model used by the glide decoder

Usage

hf download giofolf/omnivox ggml-base.bin --local-dir .

In the keyboard, the model is copied into the app assets and loaded through VoiceEngine.java:

AudioRecord -> 16 kHz PCM -> libwhisper.so (JNI) -> transcript -> local dictionary correction

Training / adaptation

Omnivox uses the unmodified Whisper base checkpoint. Adaptation is handled at runtime by the local dictionary corrector (DictionaryCorrector.java in the source repo), which replaces low-confidence ASR words with high-frequency dictionary candidates using accent-insensitive edit distance.

License

Whisper base weights and whisper.cpp are MIT licensed (Copyright OpenAI / Georgi Gerganov). The correction dictionaries are part of the keyboard project (MIT). See the source repository for full notices.

Credits

  • OpenAI Whisper
  • ggerganov/whisper.cpp
  • Keyboard project contributors
Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support