πŸŽ™οΈ Vosen-TTS-Small: High-Fidelity Uzbek Voice Cloning & Speech Synthesis (LoRA v4)

Hugging Face Model License: OpenRAIL-M Language: Uzbek

Vosen-TTS-Small is a state-of-the-art, highly optimized Text-to-Speech (TTS) model designed specifically for the Uzbek language, featuring high-fidelity zero-shot voice cloning capabilities. Trained on extensive high-quality dataset iterations using Low-Rank Adaptation (LoRA), this model delivers exceptionally natural-sounding Uzbek speech, accurate word accentuations, and robust prosody.


🌟 Key Features

  • Native Uzbek Pronunciation: Fully adjusted phoneme processing and prosody maps for maximum naturalness in northern, standard, and regional Uzbek accents.
  • Zero-Shot Voice Cloning: Instantly clone any target voice using only a short audio sample.
  • Highly Optimized LoRA Adapter: Built upon lightweight and responsive structures to guarantee rapid inference even on consumer-grade hardware.
  • High-Fidelity Audio output: Native 24kHz/44.1kHz speech generation limits background noise and audio artifacts.

βš™οΈ Model Specifications & Hyperparameters

To maintain the optimal balance between the base model's stability and the custom Uzbek voice's unique timbre, please closely adhere to the following setup parameters during inference:

1. LoRA Scale Settings (LoRA kuchi)

  • Recommended Scale: 0.9 to 1.0
  • Note: Setting the scale below 0.8 may cause a loss in voice texture identity, while going above 1.0 might cause sound clipping or synthesis distortion.

2. Voice Cloning Audio Requirements

  • Minimum Prompt Audio Duration: 6.0 seconds
  • Recommended Duration: 8.0 to 15.0 seconds
  • Input Quality: High-quality mono audio recorded at 16kHz or 24kHz without background music, echo, or overlapping speech.

πŸš€ Quick Start & Usage

Ensure you have the required dependencies installed:

pip install torch torchaudio transformers

Python Implementation

import torch
import torchaudio
# Import your pipeline code or adaptation architecture below
# from openlimi_vosen import VosenTTSPipeline 

# 1. Load the Model & LoRA weights
repo_id = "openlimi/Vosen-tts-small"
# pipeline = VosenTTSPipeline.from_pretrained(repo_id)

# 2. Setup Hyperparameters
lora_scale = 0.95  # Recommended range: 0.9 - 1.0
reference_audio_path = "path_to_your_6_second_voice_sample.wav"  # Minimum 6 seconds!
text_to_synthesize = "Vosen TTS yordamida o'zbek tilida yuqori sifatli nutq hosil qilish juda oson."

# 3. Generate Speech
# audio_output = pipeline.generate(
#     text=text_to_synthesize,
#     reference_audio=reference_audio_path,
#     lora_scale=lora_scale
# )

# 4. Save the generated speech
# torchaudio.save("uzbek_output.wav", audio_output, sample_rate=24000)

🌐 SEO Keywords & Search Optimization

This model is heavily indexed and optimized for search terms including: Uzbek TTS, O'zbekcha matnni nutqqa o'tkazish, Uzbek Speech Synthesis, Uzbek Voice Cloning, OpenLimi Vosen TTS, Uzbek LoRA Voice, AI Uzbek voice, zero-shot voice cloning Uzbek, sun'iy intellekt ovoz, o'zbekcha ovoz klonlash, TTS-Uzbek.


πŸ“„ License & Attribution

This model is released under the CreativeML OpenRAIL-M license. You are free to use, modify, and build upon this model for personal, research, and commercial applications, provided you adhere to the safety guidelines of speech generation models.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support