Quyashbek/karakalpak-omnivoice

Karakalpak OmniVoice - Fine-tuned TTS Model

A high-quality Text-to-Speech model fine-tuned for the Karakalpak language using the powerful OmniVoice architecture.

Model Description

This model was fine-tuned on approximately 6 hours of Karakalpak speech data to support the development of speech technologies for the Karakalpak language (a Turkic language spoken in Uzbekistan and Karakalpakstan).

Key Features

• Supports Karakalpak language with natural prosody
• Built on OmniVoice (Qwen3-0.6B backbone)
• Fine-tuned for voice cloning and high-quality synthesis
• Suitable for education, research, and real-world applications

Training Details

Parameter	Value
Base Model	k2-fsa/OmniVoice
Dataset	~6 hours Karakalpak speech
Training Steps	5,000
Final Eval Loss	4.055
Training Time	~51 minutes
Framework	Accelerate + PyTorch 2.6

Training Infrastructure

• GPUs: 2× NVIDIA L40S (96GB total VRAM)
• Institution: New Uzbekistan University
• Lab: Xalq Banki AI Lab

Usage

import soundfile as sf
import torch
import numpy as np
from omnivoice import OmniVoice

# ================== CHANGE ONLY IF NEEDED ==================
model_name = "Quyashbek/karakalpak-omnivoice"   # Your HF model
text = "Algoritm keń mániste tek kompyuterge tiyisli termin bolmay"
# ==========================================================

print("🔄 Loading model from Hugging Face...")
model = OmniVoice.from_pretrained(model_name)

print(f"✅ Model loaded! Sampling rate = {model.sampling_rate} Hz")

print("🎤 Generating speech...")
audio = model.generate(
    text=text,
    speaker="default",
    temperature=0.7,
    top_p=0.9,
    seed=42
)

# Fix audio format
if isinstance(audio, torch.Tensor):
    audio = audio.squeeze().cpu().numpy()

audio = np.asarray(audio).squeeze().astype(np.float32)

# Save
output_file = "karakalpak_output.wav"
sf.write(output_file, audio, model.sampling_rate, subtype='PCM_16')

print(f"✅ Success! Audio saved as → {output_file}")
print(f"   Duration: {len(audio) / model.sampling_rate:.2f} seconds")