# SpInt v1.0: A Spanish Speech Intelligibility Dataset

## Overview

SpInt is a speech intelligibility dataset designed for the evaluation of objective intelligibility metrics on Spanish speech.

The dataset contains:

* Complex spectral masks describing the effect of speech enhancement processing.
* Noise signals.
* Intelligibility labels obtained from human listening experiments.
* Metadata describing each stimulus and listener response.
* Reconstruction script.

The dataset was developed using speech material from the Spanish Matrix Test (OLSA corpus) and includes listening-test results obtained under different noisy conditions and speech enhancement systems.

To respect the licensing conditions of the Spanish Matrix Test corpus, the original clean speech recordings are not distributed as part of SpInt.

---

## Further Documentation

A detailed description of the dataset design, listening-test methodology, speech enhancement systems, objective intelligibility metric evaluation framework, and experimental results can be found in:

> Iván López-Espejo and Jesper Jensen,
> "An Objective Intelligibility Metric Evaluation on Spanish Speech",
> Proceedings of IberSPEECH 2026.

Users interested in the rationale behind the dataset construction and experimental protocol are encouraged to consult the accompanying publication.

---

## Dataset Contents

The dataset contains the following components:

### Metadata

**SpInt.json**

Metadata describing every stimulus, including:

* stimulus file,
* associated clean-speech file,
* associated noise file,
* enhancement mask,
* SNR condition,
* enhancement processing type,
* listener response,
* intelligibility score,
* participant information.

### Noise Signals

**Noise/**

Time-domain noise signals associated with each stimulus.

### Complex Masks

**Masks/**

Complex-valued spectral masks stored in NumPy `.npz` format.

### Reconstruction Script

**Generate_Stimuli.py**

Script that reconstructs the speech stimuli from:

* Spanish Matrix Test speech material,
* released noise signals,
* released masks.

---

## Why Are the Clean Speech Recordings Not Included?

The speech material used in SpInt originates from the Spanish Matrix Test corpus, whose intellectual property belongs to Hörzentrum Oldenburg.

To respect the licensing conditions governing that corpus, SpInt distributes only:

* enhancement masks,
* noise signals,
* metadata,
* reconstruction script, and
* intelligibility labels.

Users must obtain the original Spanish Matrix Test speech recordings directly from Hörzentrum Oldenburg before reconstructing the stimuli.

---

## Obtaining the Spanish Matrix Test Speech Material

The Spanish Matrix Test corpus is not included in this repository.

Researchers interested in reconstructing the stimuli should contact Hörzentrum Oldenburg:

* Dr. Kirsten Wagener (wagener@hz-ol.de)
* Dr. Melanie Zokoll (zokoll@hz-ol.de)
* Dr. Melanie Krüger (krueger@hz-ol.de)

Access to the corpus remains subject to Hörzentrum Oldenburg's licensing conditions.

---

## Reconstruction Method

For each stimulus, the reconstruction script generates:

Y(f,t) = M(f,t) · [ X(f,t) + N(f,t) ]

where:

* X(f,t): STFT of the original clean speech signal,
* N(f,t): STFT of the released noise signal,
* M(f,t): released complex spectral mask,
* Y(f,t): reconstructed stimulus.

The released masks and noise signals allow users to reconstruct high-fidelity approximations of the stimuli used during the listening experiments. Minor numerical differences may arise due to floating-point precision and implementation details of the signal processing libraries.

### STFT Parameters

| Parameter     | Value       |
| ------------- | ----------- |
| Sampling rate | 16000 Hz    |
| Window        | Hann        |
| Window length | 512 samples |
| Overlap       | 256 samples |

---

## Software Versions

The dataset was generated and validated using:

* Python 3.12.9
* NumPy 2.1.3
* SciPy 1.15.1
* SoundFile 0.13.1

---

## Citation

If you find the SpInt dataset useful in your research, please cite the following publication:

> Iván López-Espejo and Jesper Jensen,
> "An Objective Intelligibility Metric Evaluation on Spanish Speech",
> Proceedings of IberSPEECH 2026.

Please also cite any relevant publications associated with the Spanish Matrix Test corpus where appropriate.

---

## License

The SpInt dataset distributes only:

* metadata,
* noise signals,
* complex spectral masks,
* reconstruction script, and
* intelligibility labels.'

The original Spanish Matrix Test speech recordings are not included and remain subject to the licensing conditions established by Hörzentrum Oldenburg.

Users are responsible for complying with all applicable licensing terms governing the Spanish Matrix Test corpus.

---

## Contact

For questions regarding SpInt, please contact:

Dr. Iván López-Espejo
University of Granada
iloes@ugr.es