firdavsus/whisper_small_uz_ru

Whisper Small - Fine-Tuned for Uzbek & Russian (uz-ru)

Model Description

whisper_small_uz_ru is a fine-tuned version of the automatic speech recognition (ASR) model openai/whisper-small. This model has been specifically optimized to handle code-switching, multilingual environments, and balanced acoustic processing for both the Uzbek (uz) and Russian (ru) languages.

By leveraging targeted audio filtering, text normalization rules, and balanced data sampling, this model bridges the performance gap often found in lower-resource language variants while preserving strong transcription stability in Russian.

Key Performance Metrics

• Average WER (Word Error Rate): 19.0% across evaluated Uzbek and Russian joint conversational/speech datasets.

---

Intended Uses & Limitations

Primary Tasks

• Automatic Speech Recognition (ASR): Transcribing raw audio input files into punctuated and properly cased text.
• Bilingual Conversations: Highly resilient against mixed audio streams containing both Uzbek and Russian phrases (frequent in urban regions like Tashkent).

Limitations

• Audio Sample Quality: Performs best on audio inputs sampled at or resampled to 16kHz. Highly degraded, low-bitrate telephonic streams might see higher error rates.
• Hallucination in Silent Blocks: Like most Whisper-based fine-tunes, long silent intervals or persistent background ambient noise may trigger repetitive phrase hallucinations if voice activity detection (VAD) is not utilized.

---

How to Use

You can easily run inference using this model via the Hugging Face transformers pipeline API.

Basic Transcription Example

import torch
from transformers import pipeline

# Initialize the speech recognition pipeline
pipe = pipeline(
    task="automatic-speech-recognition",
    model="firdavsus/whisper_small_uz_ru",
    chunk_length_s=30,
    device="cuda" if torch.cuda.is_available() else "cpu"
)

# Transcribe an audio file (Uzbek or Russian)
audio_file_path = "path_to_your_audio.mp3"
result = pipe(audio_file_path, generate_kwargs={"task": "transcribe"})

print("Transcription:")
print(result["text"])