VOICe: Sound Event Detection Mixtures for Domain Adaptation

VOICe contains 1,449 mixtures of three distinct sound events: baby crying, glass breaking, and gunshot. The dataset was created by Shayan Gharib of Tampere University to support research in domain adaptation for sound event detection. It includes mixtures with background noise from three acoustic scenes and at two signal-to-noise ratios.

Use Cases

• Developing domain adaptation models based on mixtures across different acoustic scenes (vehicle, outdoors, indoors).
• Evaluating sound event detection robustness based on signal-to-noise ratio variations (-3 dB and -9 dB).
• Training models for generalizable audio event recognition based on the three specific sound event classes.
• Researching domain shift between clean audio and audio with background noise based on the provided mixtures.

Strengths

• Contains 1,449 total audio mixtures, providing a defined scale for experiments.
• Focuses on three specific, distinct sound events (baby crying, glass breaking, gunshot) for controlled study.
• Designed for a clear research purpose: domain adaptation across three acoustic scenes and between noisy/clean conditions.
• Includes 1,242 mixtures with controlled background noise and 207 mixtures without noise, enabling specific comparisons.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment for large-scale training.
• Last update date is unknown; freshness unverified.

Provenance

Source

Tampere University

Collection Method

Artificially mixed sound events with background noise from defined acoustic scenes.