Sound Privacy: A Conversational Speech Corpus for Quantifying the Experience of Privacy

Pablo Pérez Zarazaga, Sneha Das, Tom Bäckström, Vishnu Vidyadhara Raju Vegesna, Anil Kumar Vuppala

Introduction

While digital privacy often centers on encryption and data security, the physical acoustic context of our spoken interactions is frequently overlooked. Whether interacting with a voice assistant or sharing a secret in public, our sense of privacy is fundamentally shaped by our surroundings. Researchers have introduced the Sound Privacy Corpus, a novel dataset designed to link real-world acoustic environments with the subjective human experience of privacy.

Visualizing the risk of eavesdropping in common physical environments.

The Problem: Beyond Speech Intelligibility

Traditional room acoustics research typically focuses on "speech intelligibility"—how clearly a message is heard—to optimize office productivity or performance spaces. However, there is a significant gap in understanding the subjective experience of privacy: how comfortable an individual feels sharing information in a specific setting. By quantifying this "feeling," future AI devices could "sense" high-risk environments (such as quiet, reverberant hallways) and proactively adapt their behavior to protect user information.

Methodology

The study employed a pragmatic, real-world approach to data collection across two distinct cultural contexts: Aalto University (Finland) and IIIT Hyderabad (India).

  1. Naturalistic Recordings: Using standard mobile phones, speech was recorded in five scenarios: Office (closed door), Office (open door), Cafeteria, Big Hall, and Outdoors.
  2. Controlled Content: To prevent conversational content from biasing privacy ratings, participants used scripted nonsense dialogues from the Harvard dataset.
  3. Subjective Labeling: Immediately following recordings, participants completed an onsite questionnaire. They rated their likelihood of sharing a secret and the perceived likelihood of an eavesdropper overhearing them.
The subjective rating scales used to quantify participants' sense of privacy.

To validate that the sense of privacy is encoded in the audio itself (rather than just visual cues), the researchers conducted an online study where new participants rated the environments based solely on the recorded audio clips.

The Sound Privacy Corpus

The resulting Sound Privacy Corpus is publicly available and includes:

  • Diverse Participants: 25 speakers from Finland and 60 from India.
  • Segmented Data: Audio is organized into 15-second blocks (4 sentences each) for ease of use in machine learning and listening tests.
  • Rich Metadata: Clips are annotated with location, device model, and subjective privacy ratings.

Key Insights and Results

The study revealed several counter-intuitive findings regarding how we perceive acoustic space:

  • Noise as a Privacy Shield: Noisy environments like cafeterias were often perceived as more private than quiet ones. The "babble" of background voices acts as a natural masker, leading speakers to feel their specific conversation is harder to isolate.
  • The Reverb Effect: High reverberation in large halls significantly decreased the sense of privacy. Even in empty rooms, the way sound bounces makes speakers feel "exposed."
  • Cultural Nuance: Perceptions varied by region; for instance, Indian participants generally rated cafeterias as more private than Finnish participants, suggesting that cultural expectations of public space influence acoustic comfort.
  • Audio Validation: Trends from the audio-only online questionnaire closely matched the onsite results, confirming that "privacy" is a feature that can be extracted directly from an audio signal.

Conclusion and Future Work

The Sound Privacy Corpus provides a foundation for developing "privacy-aware" AI. While the current study focused on acoustic features, future work aims to integrate visual cues and differentiate between various types of non-stationary noise (e.g., traffic vs. speech). Ultimately, this research paves the way for devices that automatically adjust sensitivity or output volume based on the detected privacy level of their environment.

Summary updated

Abstract

With the growing popularity of social networks, cloud services and online applications, people are becoming concerned about the way companies store their data and the ways in which the data can be applied. Privacy with devices and services operated by the voice are of particular interest. To enable studies in privacy, this paper presents a database which quantifies the experience of privacy users have in spoken communication. We focus on the effect of the acoustic environment on that perception of privacy. Speech signals are recorded in scenarios simulating real-life situations, where the acoustic environment has an effect on the experience of privacy. The acoustic data is complemented with measures of the speakers’ experience of privacy, recorded using a questionnaire. The presented corpus enables studies in how acoustic environments affect peoples’ experience of privacy, which in turn, can be used to develop speech operated applications which are respectful of their right to privacy.
BibTeX
@inproceedings{zarazaga2019sound,
  author = {Pablo Pérez Zarazaga and Sneha Das and Tom Bäckström and Vishnu Vidyadhara Raju Vegesna and Anil Kumar Vuppala},
  title = {{Sound Privacy: A Conversational Speech Corpus for Quantifying the Experience of Privacy}},
  booktitle = {Interspeech},
  year = {2019},
  doi = {10.21437/interspeech.2019-1172},
}
Figure