mucemi/orpheus-ke-lora

Orpheus Kenyan English TTS (LoRA)

Fine-tuned Orpheus 3B for Kenyan English text-to-speech using a custom G2P (Grapheme-to-Phoneme) lexicon.

Quick Start

from snac import SNAC
from inference import load_ke_model, synthesise
import soundfile as sf

snac_model = SNAC.from_pretrained("hubertsiuzdak/snac_24khz")
model, tokenizer, lexicon = load_ke_model("mucemi/orpheus-ke-lora")
audio = synthesise("Habari, how are you today?", model, tokenizer, lexicon, snac_model)
sf.write("output.wav", audio, 24000)

Requirements

pip install unsloth snac transformers soundfile accelerate

Details

• Base model: Orpheus 3B (canopylabs/orpheus-3b-0.1-ft)
• Method: LoRA (r=32, alpha=32)
• Audio codec: SNAC 24kHz
• Quantisation: 4-bit (bitsandbytes)
• Training: ~45 min on Kaggle T4 GPU