CosyVoice3 — Badini Kurdish TTS (usage files)

Reference voices and inference script for the Badini Kurdish CosyVoice3 fine-tune. The fine-tuned model is in this repo: model_all/llm.pt (trained on all three speakers). The unchanged base-model files (flow, hift, tokenizer) are fetched automatically from computeram/cosyvoice3-badini-tts; the old llm.pt there is not used.

CosyVoice3 is a zero-shot voice-cloning model: every generation needs a short reference clip and its exact transcript. "Emotion" simply means which real recording of the speaker is used as the reference. This repo bundles them:

Voice	Gender	Emotions	Files
hayfa	female	natural, angry, sad, happy, fear, surprise	prompt-hayfa (natural), prompt-<emotion>
abdullah	male	angry, sad, happy, fear, surprise	prompt-abdullah-<emotion>
bafreen	female	angry, sad, happy, fear, surprise	prompt-bafreen-<emotion>

Each clip is a .wav (44.1 kHz, mono, 3–7 s) with its exact transcript in the matching .txt. Abdullah and Bafreen have no neutral ("natural") recording.

• infer.py — loads the weights and synthesizes speech in the chosen voice and emotion

Usage

git clone --recursive https://github.com/FunAudioLLM/CosyVoice.git
pip install -r CosyVoice/requirements.txt huggingface_hub

python infer.py "text to speak" --voice abdullah --emotion sad \
    --cosyvoice-dir ./CosyVoice --out output.wav

To use another voice, pass --prompt-wav and --prompt-text-file for a clip of that speaker and its exact transcript.