VyvoUp

VyvoUp is a PyTorch library for training 24 kHz to 48 kHz speech bandwidth extension models on NVIDIA GPUs.

Install

VyvoUp requires Python 3.11 or newer.

uv sync --all-extras

Dataset

Use one Hugging Face dataset or local AudioFolder with these columns:

• id: unique sample ID
• input: mono 24 kHz audio
• target: mono 48 kHz audio with twice as many samples

Train

uv run --extra data audio-upscaler-train-hf \
  --config configs/training/efficient-hbr-cuda-v1.toml \
  --dataset path/to/dataset \
  --output artifacts/run

Inference

import torch

from audio_upscaler.models import load_efficient_hbr

model = load_efficient_hbr("checkpoint.pt", device="cuda")
audio_24k = torch.zeros(1, 1, 24_000, device="cuda")

with torch.inference_mode():
    audio_48k = model.forward_aligned(audio_24k)

Test

uv run pytest