Libraries & Frameworks
Runtime Components
AI Models
Model Attribution and Modifications
CosyVoice2 includes its language model, flow decoder, HiFT vocoder, CAM++ speaker encoder and speech tokenizer. Unitra converts the synthesis models to GGUF/ONNX as applicable and quantizes the language model without retraining.
Preset voice reference clips were generated offline with Qwen3-TTS 12Hz 0.6B CustomVoice. The generation model's weights are not distributed with Unitra.
The PP-OCRv6 small detection and recognition models are the publisher's ONNX exports, used without retraining or weight conversion. The Korean recognition pack uses PP-OCRv3. Lingua includes the 1.8.0 library and its bundled 1.3.0 language models.
The DNSMOS P.835 speech-quality model is created by Microsoft and published as part of the Deep Noise Suppression Challenge under CC-BY-4.0. Unitra uses sig_bak_ovr.onnx without modification, renamed on disk, to assess recorded speech while learning a voice profile. No endorsement by Microsoft is implied.