Skip to main content
NISQA (Non-Intrusive Speech Quality Assessment) provides deep learning-based speech quality prediction without requiring a reference signal. It can predict MOS scores and multiple perceptual dimensions.

​
Model Setup

nisqa_model_setup(
    nisqa_model_path=None,
    use_gpu=False
)

​
Parameters

​
nisqa_model_path
str
required
Path to the NISQA model checkpoint (.tar file). The checkpoint must contain:
  • Model state dictionary
  • Model arguments and configuration
  • Architecture specification (NISQA, NISQA_DIM, or NISQA_DE)
​
use_gpu
bool
default:"false"
Whether to use GPU for computation. Raises RuntimeError if GPU is not available but use_gpu=True

​
Returns

​
model
NISQA model
Loaded NISQA model with the following attributes:
  • model.args - Model configuration arguments
  • model.device - Device (“cuda” or “cpu”)
  • Model can be NISQA (basic), NISQA_DIM (dimensional), or NISQA_DE (double-ended)

​
Raises

  • ValueError: If model path is not provided or checkpoint is invalid
  • RuntimeError: If GPU requested but not available
  • NotImplementedError: If model type is not recognized

​
Metric Calculation

nisqa_metric(
    nisqa_model,
    pred_x,
    fs
)

​
Parameters

​
nisqa_model
NISQA model
required
The NISQA model loaded from nisqa_model_setup()
​
pred_x
numpy.ndarray
required
Audio signal to evaluate (1D array)
​
fs
int
required
Sampling rate of the input audio in Hz. Audio will be resampled to 48 kHz if needed

​
Returns

​
metrics
dict
Dictionary containing NISQA scores with keys prefixed by nisqa_. Typical metrics include:
  • nisqa_mos - Overall MOS prediction
  • nisqa_noi - Noisiness dimension (if NISQA_DIM model)
  • nisqa_dis - Distortion dimension (if NISQA_DIM model)
  • nisqa_col - Coloration dimension (if NISQA_DIM model)
  • nisqa_loud - Loudness dimension (if NISQA_DIM model)

​
Usage Example

import numpy as np
from versa import nisqa_model_setup, nisqa_metric

# Setup NISQA model
nisqa_model = nisqa_model_setup(
    nisqa_model_path="path/to/nisqa.tar",
    use_gpu=True
)

# Load your audio
audio = np.random.random(16000)  # Replace with actual audio
fs = 16000

# Calculate NISQA scores
scores = nisqa_metric(
    nisqa_model=nisqa_model,
    pred_x=audio,
    fs=fs
)

print(f"NISQA MOS: {scores['nisqa_mos']:.2f}")
if 'nisqa_noi' in scores:
    print(f"Noisiness: {scores['nisqa_noi']:.2f}")
    print(f"Distortion: {scores['nisqa_dis']:.2f}")
    print(f"Coloration: {scores['nisqa_col']:.2f}")
    print(f"Loudness: {scores['nisqa_loud']:.2f}")

​
Model Variants

​
NISQA (Basic)

Predicts overall MOS score for speech quality.

​
NISQA_DIM (Dimensional)

Predicts MOS along with perceptual dimensions:
  • Noisiness: Amount of background noise
  • Distortion: Signal distortion level
  • Coloration: Frequency response coloration
  • Loudness: Perceived loudness

​
NISQA_DE (Double-Ended)

Double-ended variant that can use reference signal if available.

​
Technical Details

Model Input: NISQA expects audio at 48 kHz sampling rate. Audio is automatically resampled if provided at a different rate.
Model Architecture: NISQA uses CNN features with temporal dependencies captured via self-attention or LSTM layers.

​
Model Download

Download pre-trained NISQA models from: 
# Example download location
wget https://github.com/gabrielmittag/NISQA/raw/master/weights/nisqa.tar

​
Performance Characteristics

AspectDetails
Sampling Rate48 kHz (auto-resampled)
Input TypeSingle-channel audio
Output Range1-5 (MOS scale)
Reference RequiredNo (non-intrusive)
Model TypeDeep neural network

Build docs developers (and LLMs) love

Get started for freeTalk to us