balacoon/ecapa

This is an ECAPA model traced into a jit for simple use.

Usage:

from huggingface_hub import hf_hub_download
import torch
import soundfile as sf

# Download the model from repo
model_path = hf_hub_download(
    repo_id="balacoon/ecapa",
    filename="ecapa.jit",
    repo_type="model",
    local_dir="./",
)

# load model
utmos_model = torch.jit.load(model_path).to(torch.device("cuda"))
# load audio
wav, sr = sf.read(
    "rms_arctic_a0001.wav",
    dtype="int16"
)
assert sr == 16000
# run inference
x = torch.tensor(wav).unsqueeze(0).cuda()
x_len = torch.tensor([x.shape[1]], device=x.device)
emb = utmos_model(x, x_len)