Reza2kn/Gooya-Koochik-v2.0-exp

Gooya Koochik v2.0-exp

Persian text in. Speech out. This is the merged, listener-preferred model, with the Negara v7.1 frontend and a default synthetic voice included. No adapters to choose or merge. No phonemes to prepare.

Quick start

Python 3.10+; authenticate with Hugging Face first because the repo is private.

pip install -U huggingface_hub
hf download Reza2kn/Gooya-Koochik-v2.0-exp --local-dir gooya
pip install -r gooya/requirements.txt
python gooya/gooya.py "سلام! خوش اومدید." --output hello.wav

That writes hello.wav. NVIDIA CUDA is used when available; otherwise it runs on CPU, which is slower. No reference recording is required.

Python

From inside the downloaded gooya directory:

from gooya import Gooya

model = Gooya.from_pretrained(".")
model.speak("قرار بود امروز بیای، چرا نیومدی؟", "output.wav")

To use your own reference voice:

model.speak(
    "سلام! خوش اومدید.",
    "output.wav",
    reference="my_voice.wav",
    reference_text="متنی که دقیقاً در فایل صدای مرجع گفته شده است.",
)

The bundled default voice is made from synthetic Orus development samples. The wrapper handles Persian phonemization, sentence splitting, and short pauses automatically. It uses 32 generation steps and speed 0.85 by default. Voice references should be used with the speaker's authorization.

What model is this?

This is the fully merged continuation-200 checkpoint, the version preferred in our listening comparisons. It is a Persian adaptation of OmniVoice, with approximately 613 million speech-model parameters plus its audio codec. Model weights are stored in BF16; the wrapper uses FP32 on CPU.

The preferred training mixture contained 33,689 synthetic clips plus 85 real-data clips, approximately 31.68 hours before repetition. Training and phonemization used only normalized Gemini transcripts. Negara v7.1 supplies the case-sensitive pronunciation representation internally.

Experimental: pronunciation, question intonation, and pacing can still be wrong. The model is not claimed to be production-ready. A later Nasle training run made intonation worse and is not included in these weights. The old-data-only control was not promoted either. Informal listening selected this release; small validation-loss differences did not reliably track prosody.

All earlier adapters, experiments, and detailed recipes are preserved on the adapter-history branch. They are not needed to load this model.

Built from OmniVoice, Negara v7.1, AvaSanj, and the Homorich, Nasle Mana, and Gooshkon data used in the archived recipes. Refer to the respective upstream model, codec, code, and dataset terms. See release.json for exact provenance and verification.