aisha-org/navoiy-tts

Lison Navoiy TTS

Lison Navoiy TTS is Aisha's public Uzbek expressive text-to-speech model artifact built on CosyVoice2.

This repository contains the LLM/checkpoint component and demo assets. It is not a standalone one-file inference package; inference requires the upstream CosyVoice/CosyVoice2 runtime and compatible base model assets.

Model summary

Field	Value
Product family	Lison
Task	Uzbek text-to-speech
Base model	FunAudioLLM/CosyVoice2-0.5B
Language	Uzbek (uz)
License	Apache-2.0
Main checkpoint	emotion_600h_joint.pt
Repository visibility	public
Revision	912e6bb6663f11e5dec5de830ca06ea8dd4c985f

Intended use

• Uzbek expressive TTS research and demos
• Voice-agent speech generation experiments
• CosyVoice2-based Lison model evaluation

Files

File	Size
emotion_600h_joint.pt	1.88 GiB
demo/long_form.wav	0.49 MiB
demo/numbers_dates.wav	0.42 MiB
demo/happy.wav	0.24 MiB
demo/calm_intro.wav	0.23 MiB
demo/angry.wav	0.23 MiB
demo/sad.wav	0.22 MiB
demo/warm_agent.wav	0.17 MiB
demo/surprised.wav	0.16 MiB
LICENSE	0.01 MiB

Demo assets

The demo/ folder contains generated Uzbek WAV examples and a demo manifest. Use these for quick qualitative checks, not as a replacement for systematic MOS or production latency evaluation.

Limitations

• Requires CosyVoice/CosyVoice2 serving code and compatible dependencies.
• Production quality depends on prompt/reference audio, normalization, serving latency, and post-processing.
• Do not treat demo outputs as full benchmark coverage.

Related Lison datasets

• Canonical non-emotion corpus: aisha-org/uzbek-tts-corpus-v1
• Separate emotion dataset: aisha-org/uzbek-emotion-tts