MISP-QEKS: A Large-Scale Dataset with Multimodal Cues for Query-by-Example Keyword Spotting
ACM MM 2025, Dublin, Ireland
DOI: https://doi.org/10.1145/3746027.3758268

Overview

MISP-QEKS is the first large-scale tri-modal benchmark for open-vocabulary Query-by-Example Keyword Spotting (QEKS).

Unlike traditional keyword spotting datasets that:

focus on fixed keyword sets
rely on clean audio-only recordings
lack OOV evaluation

MISP-QEKS provides:

Fully aligned Text–Audio–Visual keyword clips
Real-world noise simulation
In-Vocabulary (IV) and Out-of-Vocabulary (OOV) evaluation splits
610,000 enrollment–query pairs
9,830+ distinct keywords

This dataset enables robust, multimodal, open-vocabulary keyword spotting research under realistic acoustic conditions.

Task Definition

As shown in the figure, given:

An enrollment example (text/audio/video)
A query clip (audio/video)

The system predicts:

Whether both samples contain the same keyword
A probability score

This supports:

Open-vocabulary keyword spotting
Cross-modal matching
Robust detection under noise

Dataset Construction Pipeline

As shown in the figure, MISP-QEKS is constructed from sentence-level audio-visual-text data via:

Phone-level forced alignment
Word-level cropping
Real-world noise simulation
Enrollment–query pair construction

This pipeline enables large-scale, synchronized multimodal keyword samples suitable for open-vocabulary QEKS research.

Dataset Statistics

Total duration: 193.606 hours
Keywords: 9,830
Enrollment–Query pairs: 610,000
- 122,000 positive
- 488,000 negative
Positive:Negative ratio = 1:4

Data Splits

Split	Duration (h)	Keywords	Pairs	Positive	Negative
Train	157.756	8,357	500,000	100,000	400,000
Dev	3.245	2,247	10,000	2,000	8,000
Eval-seen	15.300	2,174	50,000	10,000	40,000
Eval-blind	17.305	1,445	50,000	10,000	40,000

Evaluation protocol:

Eval-seen → In-Vocabulary (IV)
Eval-blind → Out-of-Vocabulary (OOV)
Speaker-independent split

Keyword Frequency Distribution

The keyword frequency distribution demonstrates:

Strong coverage across high-frequency and mid-frequency words.
Long-tail behavior suitable for evaluating generalization.

This design supports robust training while preserving realistic lexical imbalance.

Noise Characteristics and Quality Distribution

To emulate realistic acoustic environments, clean clips are mixed with real-world background noise at SNR levels {+5, 0, −5, −10} dB.

Speech Quality Distribution

Most PESQ scores lie between 1.5 and 2.5.
Most STOI values fall between 65% and 85%.

This confirms that the dataset spans a broad spectrum of realistic noise conditions.

Repository Structure and File Description

The repository contains the following files:

Data Archives

train.zip Contains the training split (157.756 hours, 500,000 enrollment–query pairs across 8,357 keywords).
dev_seen.zip Development split for hyperparameter tuning on In-Vocabulary (IV) keywords.
Keywords partially overlap with the training set.
dev_unseen.zip Development split for Out-of-Vocabulary (OOV) validation.
Keywords do not appear in the training set.
eval_seen.zip In-Vocabulary (IV) evaluation split.
Keywords appear in the training set and are used for standard evaluation.
eval_unseen.zip
Out-of-Vocabulary (OOV) evaluation split.
Keywords are not seen during training and are used to assess generalization.

Noise and Metadata

noise.zip
Real-world background noise recordings used for acoustic simulation.
snr_map.zip Mapping file indicating the signal-to-noise ratio (SNR) assigned to each noisy sample.

Baseline Checkpoint

train/model/

Contains the official 10-epoch checkpoint of the XEQ-Matcher baseline described in the ACM MM 2025 paper.

This checkpoint can be used directly for evaluation or reproduction of reported results.

Official implementation: https://github.com/coalboss/MISP-QEKS

Pretrained Feature Extractors

model/

Contains pretrained feature extraction models required by the baseline system.

These models are used as frozen encoders for:
- Audio feature extraction (e.g., Whisper-Tiny encoder)
- Visual feature extraction (CNN-ResNet backbone)
- Text processing (G2P model)
These components are necessary to reproduce the reported baseline performance.

Downloads last month: 31