notable-speech-models
Uzbek speech-recognition models in whisper.cpp ggml format, used by the .notable app for fully offline, on-device transcription. The app downloads one of these files once.
| File | Source model | Quantisation | Size (bytes) |
|---|---|---|---|
ggml-uz-base-q5_1.bin |
navai-uz/whisper-base-uzbek | q5_1 | 59707642 |
ggml-uz-small-q5_1.bin |
navai-uz/whisper-small-uzbek | q5_1 | 190085504 |
Credit
These files are unmodified weights of the NavAI Uzbek Whisper fine-tunes, only converted
from the Hugging Face format to whisper.cpp ggml (convert-h5-to-ggml.py, whisper.cpp v1.8.3)
and quantised to q5_1 (quantize). All credit for the models goes to
NavAI (https://huggingface.co/navai-uz), who fine-tuned them from OpenAI Whisper
(MIT) on public Uzbek corpora. Please see the original model cards for training data,
evaluation and limitations.
Licence
Apache License 2.0, as the original weights (see LICENSE). Attributions required by the
training-data licences and the OpenAI Whisper base model are reproduced in NOTICE.
Notes
- Output is Latin-script Uzbek, typically lowercase, without punctuation, with numbers spelled out. Works best on audio without timestamps in chunks of up to ~25 s.
- Run with whisper.cpp
-l uz. Not compatible with GGUF-based runtimes.