whisqa: a PyPI package for the WhiSQA Non-Intrusive Speech Quality Prediction Model

This project is a fork of the original WhiSQA project repackaged for PyPI.

WhiSQA is a non-intrusive MOS (ITU-T P.835) speech quality predictor.

Original repo: https://github.com/leto19/WhiSQA

• pip install whisqa — fully installable package; no manual cloning or dependency wrangling.
• Bundled weights — the frozen Whisper encoder weights (~350 MB per checkpoint) are stripped down to head-only files of 11 MB and 13 MB that ship inside the wheel. The strip script is kept in the repo for full reproducibility.
• Flexible input — predict() accepts file paths, NumPy arrays, and Torch tensors, making it easy to integrate into existing pipelines without writing audio to disk.
• Auto-resampling — any sample rate is accepted; audio is resampled to 16 kHz automatically with a suppressible warning.
• CLI — score a file from the terminal with whisqa speech.wav.
• Simple API — predict() and load_model() with type hints, docstrings, and unit tests.

pip install whisqa

First-run note: load_model() downloads openai/whisper-small (~240 MB) from HuggingFace Hub and caches it in ~/.cache/huggingface/. The WhiSQA head weights ship inside the package and require no download.

import whisqa

# File path — WAV or any format supported by libsndfile
whisqa.predict("speech.wav")
# → {'mos': 3.82}

# NumPy array
import numpy as np
whisqa.predict(np.zeros(16000, dtype=np.float32), sample_rate=16000)

# Torch tensor — drop straight in from your existing pipeline
import torch
whisqa.predict(torch.zeros(1, 16000), sample_rate=16000)

# All five P.835 dimensions
whisqa.predict("speech.wav", model_type="multi")
# → {'mos': 3.82, 'noisiness': 4.10, 'coloration': 3.55, 'discontinuity': 4.20, 'loudness': 3.90}

# Efficient repeated inference — load once, score many
model = whisqa.load_model("single")
for f in my_files:
    print(whisqa.predict(f, model=model))

# CLI
whisqa speech.wav
whisqa speech.wav --model multi
whisqa speech.wav --no-warn-resample

If you use WhiSQA in your work, please cite the original authors:

@inproceedings{close2025whisqa,
  title     = {{WhiSQA}: Non-Intrusive Speech Quality Prediction Using {Whisper} Encoder Features},
  author    = {Close, George and Hong, Kris Y. and Hain, Thomas and Goetze, Stefan},
  booktitle = {Speech and Computer -- 27th International Conference, {SPECOM} 2025,
               Szeged, Hungary, October 13--15, 2025, Proceedings, Part {I}},
  editor    = {Karpov, Alexey and Gosztolya, G{\'a}bor},
  series    = {Lecture Notes in Computer Science},
  volume    = {16187},
  pages     = {39--51},
  publisher = {Springer},
  year      = {2025},
  doi       = {10.1007/978-3-032-07956-5_3},
}