aisha-org/navoiy-tts

Navoiy TTS

Navoiy TTS is an open Uzbek text-to-speech checkpoint built on CosyVoice2-0.5B. It generates highly natural, expressive speech that captures the specific phonetic nuances of the Uzbek language.

Note: This is an LLM component checkpoint, not a standalone model. Inference requires the upstream CosyVoice codebase and the base model.

Model Overview & Results

Unlike standard TTS models that can sound robotic or struggle with regional phonemes, Navoiy TTS is specifically fine-tuned for emotional prosody and natural pacing in Uzbek.

• Training Data: Trained on a high-quality dataset comprising ~600 hours of neutral speech and an additional 40 hours of heavily expressive/emotional speech.
• Capabilities: Supports zero-shot voice cloning and instruction-conditioned emotion generation (e.g., whispering, shouting, varying tones).
• Evaluation: Exhibits strong zero-shot robustness, significantly reducing mispronunciations of common Uzbek vocabulary and numbers compared to the base multilingual models.

Checkpoint Details

File	Description
emotion_600h_joint.pt	Primary expressive checkpoint. Recommended for most use cases.

Quick Start

For reproducible inference, use the pinned revisions below.

1. Setup Upstream Environment

git clone --recursive [https://github.com/FunAudioLLM/CosyVoice.git](https://github.com/FunAudioLLM/CosyVoice.git)
cd CosyVoice
git checkout 074ca6dc9e80a2f424f1f74b48bdd7d3fea531cc
git submodule update --init --recursive
python3 -m venv .venv && source .venv/bin/activate
pip install --upgrade pip setuptools wheel
pip install -r requirements.txt
cd ..

2. Download Models

First, clone the Navoiy repository and install its specific requirements:

git clone [https://huggingface.co/aisha-org/navoiy-tts](https://huggingface.co/aisha-org/navoiy-tts)
pip install -r navoiy-tts/requirements.txt

Next, create a Python script (e.g., download_base.py) to fetch the exact base model checkpoint needed:

# download_base.py
from huggingface_hub import snapshot_download

print("Downloading CosyVoice2-0.5B base model...")
snapshot_download(
    repo_id="FunAudioLLM/CosyVoice2-0.5B", 
    revision="eec1ae6c79877dbd9379285cf8789c9e0879293d", 
    local_dir="CosyVoice/pretrained_models/CosyVoice2-0.5B"
)
print("Download complete.")

Run the script:

python download_base.py

3. Run Inference

Requires a CUDA-enabled GPU.

python navoiy-tts/inference.py \
  --cosyvoice-dir CosyVoice \
  --base-model-dir CosyVoice/pretrained_models/CosyVoice2-0.5B \
  --checkpoint navoiy-tts/emotion_600h_joint.pt \
  --reference reference.wav \
  --text "Bugun siz bilan muhim bir xabarni baham ko'rmoqchiman." \
  --emotion warm \
  --output output.wav

Supported Emotion Presets: calm, happy, sad, angry, nervous, surprised, whisper, warm, tired, sarcastic. (Note: Output depends heavily on the reference voice and seed; treat these as style hints, not guaranteed classifications).

Text Normalization

The included uztts.normalize utility handles Cyrillic-to-Latin conversion, apostrophe canonicalization, and expansion of numbers/dates/units.

from uztts.normalize import normalize

text = "Bugun 16.07.2026, soat 14:30 da uchrashamiz."
print(normalize(text))

Demos

Generated using emotion_600h_joint.pt on an NVIDIA A10-12Q. Full generation configs, seeds, and RTF metrics are in demo/manifest.json.

Calm:

Navoiy TTS o'zbekcha matnni tabiiy va ravon ovozga aylantiradi.

Warm Agent:

Assalomu alaykum! Sizga qanday yordam bera olaman?

Happy:

Ajoyib yangilik! Bugungi natijalarimiz kutilganidan ham yaxshi bo'ldi.

Angry

Bu xatoni darhol tuzatish kerak. Bunday holat boshqa takrorlanmasin.

Surprised

Nahotki, loyiha bir kun ichida tayyor bo'lgan bo'lsa!

Numbers and dates

Bugun 16.07.2026, soat 14:30 da uchrashamiz. Manzilgacha 5.5 km qoldi.

Long-form Uzbek

Alisher Navoiy merosi o'zbek tili va adabiyotining bebaho xazinasidir. Navoiy TTS zamonaviy sun'iy intellekt yordamida tabiiy o'zbekcha nutq yaratadi.

Note: While generally robust, the current checkpoint may occasionally mispronounce complex words, numbers, dates or generate audio artifacts. We are actively refining the training data to resolve these edge cases.

Constraints & Responsible Use

• Failure Modes: Expect potential mispronunciations of highly specialized loanwords or complex mathematical formulas.
• Ethics & Safety: Voice cloning requires knowing, explicit consent from the speaker. Do not use this generated speech for safety-critical instructions, fraud, or impersonation. Clearly disclose synthetic speech to listeners.

License

Apache License 2.0. Derived from CosyVoice and CosyVoice2-0.5B.