nardi/cosyvoice3-albanian-llm

CosyVoice3 — Albanian LLM (merged LoRA)

Albanian text-to-speech LLM checkpoint for the CosyVoice3 pipeline. This is the language-model part of FunAudioLLM/Fun-CosyVoice3-0.5B-2512 (a Qwen2-0.5B with an extended vocabulary: 151,924 text tokens + 6,561 speech tokens + special tokens), merged with a LoRA fine-tuned for Albanian speech synthesis.

Training data

• 55 hours of Albanian news speech only (VOA valid Albanian, 20,711 clips).
• LoRA: r=32, alpha=64, dropout=0.05, on all attention + MLP projections.

Usage

This checkpoint is meant to be used together with the CosyVoice3 acoustic stack (flow-matching DiT + HiFi-GAN). The easiest way is the Gradio demo / Space — it downloads the acoustic models from the official CosyVoice3 release and this repo for the LLM:

from transformers import AutoModelForCausalLM, AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained("nardi/cosyvoice3-albanian-commonvoice-llm")
model = AutoModelForCausalLM.from_pretrained("nardi/cosyvoice3-albanian-commonvoice-llm", dtype="bfloat16")
# -> feed text + prompt speech tokens, generate speech tokens, then decode
#    with flow.pt + hift.pt (see the Space's app.py for the full pipeline).

Token layout (see cosyvoice3_metadata.json): text < 151,924 | speech tokens 151,924 … 158,484 | <|sos|>=158,485 <|eos1|>=158,486 <|task_id|>=158,487.

License & ethics

Apache-2.0, following the base model. Voice cloning requires the explicit consent of the voice owner — do not use for impersonation or deception.