This is the official implementation of "DNSMOS Pro: A Reduced-Size DNN for Probabilistic MOS of Speech". DNSMOS Pro is a model that takes as input a speech clip, and outputs a Gaussian mean opinion score (MOS) distribution.

Authors: Fredrik Cumlin, Xinyu Liang
Emails: [email protected], [email protected]

There are three pretrained DNSMOS Pro ready to be used, trained on three datasets respectively. For inference, one can do the following (all paths are relative to this directory):

import numpy as np
import torch

import utils  # Python file containing the STFT.

model = torch.jit.load('runs/NISQA/model_best.pt', map_location=torch.device('cpu'))
samples = np.ones(160_000)
# Defaults in `utils.stft` correspond to training values.
spec = torch.FloatTensor(utils.stft(samples))
with torch.no_grad():
    prediction = model(spec[None, None, ...])
mean = prediction[:, 0]
variance = prediction[:, 1]
print(f'{mean=}, {variance=}')

The mean can be used as a scalar prediction of MOS. Recommended input duration is 10 s, and should be of 16 kHz sample rate.

Installation with pip:

pip install -r requirements.txt
pip install torch==2.1.0+cu118 --index-url https://download.pytorch.org/whl/cu118

VCC2018

BVCC

NISQA

The framework is Gin configurable; specifying model and dataset is done with a Gin config. See examples in configs/*.gin.

Example launch:

python train.py --gin_path "configs/vcc2018.gin" --save_path "runs/VCC2018"