NavaiSTT-2v Medium: A Fully Open-Source Uzbek Speech-to-Text Model

About NavaiSTT v2

This is the second version of the NavaiSTT model, representing a classic Whisper medium model fine-tuned for the Uzbek language. This release prioritizes transparency, reproducibility, and community access.

The key difference from v1 is that v2 is fully open-sourced. Due to conflicts with data partners, the v1 model and its proprietary 500-hour dataset were removed. For v2, I have curated a new set of publicly available and permissively licensed datasets. Critically, both the final dataset and the training scripts will be open-sourced, making the entire fine-tuning process fully repeatable by anyone in the community.

Special attention was again given to Tashkent dialect audio, resulting in strong performance for this common dialect. Future work will focus on incorporating more regional dialects to improve coverage across Uzbekistan.

Model Details

Training Data & Methodology

This model was fine-tuned on approximately 475 hours of diverse Uzbek audio data. The dataset was composed of 50% human-transcribed and 50% pseudo-transcribed material (using Gemini 2.5 Pro).

The training corpus included:

• Common Voice 17 dataset (filtered)
• USC (filtered)
• Google fleurs (filtered)
• Podcasts Tashkent Dialect Youtube Uzbek Speech Dataset
• News Youtube Uzbek Speech Dataset
• IT Youtube Uzbek Speech Dataset

A custom script was used to filter the datasets based on Word Error Rate (WER) and similarity checks to ensure data quality. This filtering script will also be open-sourced as part of our commitment to a fully repeatable process.

Open-Sourced Datasets for Full Reproducibility

Training and Filtering Code

The complete training pipeline, including data filtering and model fine-tuning scripts, is available for anyone who wants to reproduce or improve upon NavaiSTT-2v:

🚀 View Training Code on GitHub

Performance Comparison

Listen to the sample below and compare NavaiSTT-2v Medium transcription accuracy with other leading services and our previous model:

Thoughts on Open Source vs. Commercial

Let's be honest - most startups don't need open source AI models unless they are facing specific legal restrictions. Gemini 1.5, 2, 2.5 Pro delivers remarkably human-like performance, especially for transcription where it captures nuanced speech with accuracy that rivals human ears. When you hear someone speaking, this model understands exactly as you would.

However, when handling sensitive personal data that cannot be shared with foreign third parties, such as private audio recordings containing confidential information, regionally-hosted services or self-hosted open source alternatives (like NavaiSTT) provide necessary compliance and data sovereignty.

Implementation Example

Integrating NavaiSTT v2 into your application is straightforward using the Transformers library. Here's a Python code example:

import torch
import torchaudio
from transformers import WhisperProcessor, WhisperForConditionalGeneration

def transcribe_audio(audio_path, model_path):
    """Simple function to transcribe audio with a fine-tuned Whisper model"""
    # Load model and processor
    processor = WhisperProcessor.from_pretrained(model_path)
    model = WhisperForConditionalGeneration.from_pretrained(model_path)

    # Move to GPU if available
    device = "cuda" if torch.cuda.is_available() else "cpu"
    model.to(device)

    # Load and preprocess audio
    waveform, sample_rate = torchaudio.load(audio_path)
    if sample_rate != 16000:
        waveform = torchaudio.functional.resample(waveform, sample_rate, 16000)

    # Convert to mono if needed
    if waveform.shape[0] > 1:
        waveform = waveform.mean(dim=0, keepdim=True)

    # Process audio
    input_features = processor(
        waveform.squeeze().numpy(),
        sampling_rate=16000,
        return_tensors="pt"
    ).input_features.to(device)

    # Generate transcription
    with torch.no_grad():
        predicted_ids = model.generate(input_features)

    # Decode
    transcription = processor.batch_decode(predicted_ids, skip_special_tokens=True)[0]
    return transcription

# Example usage
if __name__ == "__main__":
    MODEL_PATH = "islomov/navaistt_v2_medium"
    # Make sure to create a dummy audio file named "some_audio.wav" for this to run
    audio_file = "some_audio.wav"

    text = transcribe_audio(audio_file, MODEL_PATH)
    print(f"Transcription: {text}")