--- language: - uz license: creativeml-openrail-m library_name: transformers tags: - automatic-speech-recognition - asr - speech-recognition - whisper-asr - mute-speech - whispered-speech - uzbek-stt - uzbek-asr - speech-to-text - ai-uzbek pipeline_tag: automatic-speech-recognition metrics: - wer - cer --- # 🎙️ Tunar-STT-Medium: High-Precision Uzbek Speech Recognition (Whispered & Silent Voice Optimized) [![Hugging Face Model](https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-Model-blue)](https://huggingface.co/openlimi/Tunar-stt-medium) [![License: OpenRAIL-M](https://img.shields.io/badge/License-OpenRAIL--M-green.svg)](https://huggingface.co/spaces/CompVis/stable-diffusion-license) [![Language: Uzbek](https://img.shields.io/badge/Language-Uzbek%20(uz)-red)](https://en.wikipedia.org/wiki/Uzbek_language) **Tunar-STT-Medium** is a cutting-edge, state-of-the-art Uzbek Automatic Speech Recognition (ASR) model specifically engineered to tackle both standard high-fidelity voice inputs and extremely low-amplitude speech patterns, such as **whispered, quiet, or near-silent voices (MuteVoice)**. --- ## ⚡ Key Advantages & Tunar-Tiny Comparison - **Whispered Speech Transcription**: Unlike standard ASR systems, this model is highly optimized to detect, capture, and precisely transcribe speech even when the speaker is pichirlab gapirganda (whispering). - **Massive Performance Boost**: Compared to the lighter **Tunar-stt-tiny** model, Tunar-STT-Medium features a larger model capacity, vastly improved contextual comprehension, and significantly lower Word Error Rate (WER) on noisy, complex, or low-volume recordings. - **Robust Accents & Dialects Support**: Offers deep acoustic understanding of standard Uzbek phonemes as well as various regional dialectical nuances. --- ## 🛠️ Performance & Fine-Tuning Details This model is fine-tuned with a rich corpus of native Uzbek datasets, carefully mixing clean studio speech, noisy phone calls, and whispered voice segments. It utilizes standard 16 kHz sampling and outputs clean text formatted directly in standard Uzbek transcription styles. --- ## 🚀 Quick Start & Usage Make sure to load and test the model using standard transformers code: ```python import torch import soundfile as sf from transformers import AutoModelForCTC, AutoProcessor repo_id = "openlimi/Tunar-stt-medium" # Load model and processor # processor = AutoProcessor.from_pretrained(repo_id) # model = AutoModelForCTC.from_pretrained(repo_id).eval() ``` --- ## 🌐 SEO Keywords & Search Engine Optimization This model is highly indexed and optimized for search queries such as: *Uzbek STT, Uzbek ASR, O'zbekcha nutqni matnga aylantirish, Pichirlash ovozini aniqlash, Whisper Uzbek speech, Uzbek MuteVoice, Tunar STT Medium, OpenLimi STT Uzbek, AI Uzbek voice transcriber, Sun'iy intellekt o'zbekcha STT.*