Downloads · 30 days
0
azimxxm/rubaistt-v2-medium-ggml
rubaistt-v2-medium-ggml is a automatic speech recognition model from azimxxm. Use it when you need speech turned into text. The card lists the license as apache-2.0.
A ready-to-run whisper.cpp conversion of islomov/rubaisttv2medium, a Whisper-medium fine-tune for Uzbek (Latin script) by Sardor Islomov (~17 % WER / 5.5 % CER on the author's test set). Same weights, same Apache-2.0…
Downloads · 30 days
0
Access
Public
Updated Sep 24, 2026
Repo size
539 MB
Likes
0
Public
Click a slice to open those files.
.bin539 MB · 100%
From the Hugging Face model README
A ready-to-run whisper.cpp conversion of islomov/rubaistt_v2_medium,
a Whisper-medium fine-tune for Uzbek (Latin script) by Sardor Islomov (~17 % WER / 5.5 % CER on the author's test set).
Same weights, same Apache-2.0 license — only the file format changed, so whisper-cli / whisper-server load it directly
with no Python, no torch, no 3 GB download.
| file | size | how |
|---|---|---|
ggml-rubaistt-medium-q5_0.bin | 516 MB | convert-h5-to-ggml.py (whisper.cpp) → f16 → whisper-quantize … q5_0 |
On synthesized Uzbek test sentences q5_0 produced exactly the same text as q8_0, so only q5_0 is published.
brew install whisper.cpp
curl -L -o ggml-rubaistt-medium-q5_0.bin \
https://huggingface.co/azimxxm/rubaistt-v2-medium-ggml/resolve/main/ggml-rubaistt-medium-q5_0.bin
whisper-cli -m ggml-rubaistt-medium-q5_0.bin -f audio16k.wav -l uz -nt
# keep it warm: whisper-server -m ggml-rubaistt-medium-q5_0.bin -l uz --port 8178
Always pass -l uz: Whisper's auto-detect mistakes Uzbek for Arabic-script languages. The model transliterates Russian
into Latin letters, so use a general Whisper model for other languages.
Example (audio synthesized with a neural Uzbek voice):
Jarvis, iOS ilovada login ekranini Swift bilan yoz, keyin testlarni ishga tushir va GitHub'ga push qil. →
jarvis ios ilovada login ekranini swift bilan yoz, keyin testlarni ishga tushir va githubga push qil.
Built for claude-code-voice (ovoz) — talk to Claude Code in Uzbek and hear it answer. Credit for the model goes to the original author; this repo only hosts the converted file.