Model-At-checkPoint-20000
This is a fully independent, merged deployment of the VibeVoice acoustic model fine-tuned for Egyptian Arabic at checkpoint 20000. It includes the integrated Qwen tokenizer configs and base weights side-by-side.
📦 Repository Structure
model.safetensors: Independent merged model weights.voices/: Reference speech audio samples.- Tokenizer and configuration JSON files.
🚀 How to Use (Inference API / Colab)
import torch
from peft import PeftModel
# Load this repository directly using standard HuggingFace or VibeVoice modules:
# model = VibeVoiceForConditionalGeneration.from_pretrained("MohammedEhab20/Model-At-checkPoint-20000", trust_remote_code=True)
🎛️ Runtime Parameters (CFG Scale)
The Classifier-Free Guidance (CFG) scale is a runtime parameter and not baked into these weights. You can dynamically adjust it during inference calls:
- Higher CFG (e.g.,
5.0): Strict alignment with the prompt. - Lower CFG (e.g.,
3.5): More natural flow and creativity.
- Downloads last month
- 8
import torch, soundfile as sf, librosa, numpy as np from vibevoice.processor.vibevoice_processor import VibeVoiceProcessor from vibevoice.modular.modeling_vibevoice_inference import VibeVoiceForConditionalGenerationInference # Load voice sample (should be 24kHz mono) voice, sr = sf.read("path/to/voice_sample.wav") if voice.ndim > 1: voice = voice.mean(axis=1) if sr != 24000: voice = librosa.resample(voice, sr, 24000) processor = VibeVoiceProcessor.from_pretrained("MohammedEhab20/Model-At-checkPoint-20000") model = VibeVoiceForConditionalGenerationInference.from_pretrained( "MohammedEhab20/Model-At-checkPoint-20000", torch_dtype=torch.bfloat16 ).to("cuda").eval() model.set_ddpm_inference_steps(5) inputs = processor(text=["Speaker 0: Hello!\nSpeaker 1: Hi there!"], voice_samples=[[voice]], return_tensors="pt") audio = model.generate(**inputs, cfg_scale=1.3, tokenizer=processor.tokenizer).speech_outputs[0] sf.write("output.wav", audio.cpu().numpy().squeeze(), 24000)