Orphues_Q4_K_M

This repository contains a quantized GGUF version (Q4_K_M) of an Orpheus-based Text-to-Speech model, fine-tuned using Unsloth for efficient training and inference.

⚠️ Note:
This model is NOT trained or fine-tuned on Persian (Farsi) data.
It is intended for non-Persian (primarily English) text-to-speech use cases.

---

🔊 Model Overview

• Base architecture: Orpheus TTS (LLM-based Audio Generation)
• Backbone: LLaMA-style Transformer
• Fine-tuning framework: Unsloth
• Quantization: Q4_K_M (GGUF)
• Inference target: llama.cpp / GGUF-compatible runtimes
• Language support: Non-Persian (English-focused)

---

🧠 About Orpheus TTS

Orpheus is a modern LLM-based Text-to-Speech architecture that generates audio by predicting audio tokens instead of mel-spectrograms.
It enables:

• Natural prosody
• Expressive speech
• Low-latency generation
• Zero-shot or few-shot voice adaptation (depending on setup)

This model follows that paradigm and has been adapted via fine-tuning.

---

🛠️ Fine-Tuning Details

• Fine-tuning method: LoRA-based fine-tuning via Unsloth
• Training focus: Improving speech naturalness and stability
• Tokenizer: Original LLaMA-compatible tokenizer
• No Persian data used

---

⚙️ Quantization Details

• Format: GGUF
• Quantization type: Q4_K_M
• Optimized for:
  • Reduced VRAM usage
  • Fast CPU/GPU inference
• Recommended runtime: llama.cpp (latest version)

---

🚀 Usage Example (llama.cpp)

./main \
  -m Orphues_Q4_K_M.gguf \
  --temp 0.7 \
  --ctx-size 4096