A multi-room reverberant dataset for sound event localization and detection

1. INTRODUCTION

• SELD combines recognition of known sound classes, onset and offset detection, and spatial localization of active events.The paper notes potential use in assisted listening, scene visualization, immersive media, and spatial machine cognition.Source: §1 · p. 1
• Earlier SELD methods were limited, and separating sound-event detection from direction estimation created a data-association problem with overlapping events.Later work proposed joint deep-network approaches, using either discrete-angle classification or continuous-angle regression.Source: §1 · p. 1
• Classification is limited to trained angles and can require large, imbalanced classifiers, while regression handles unseen angles but had lower accuracy on seen locations in prior work.Performance for simultaneous instances of the same sound class had not been evaluated.Source: §1 · pp. 1–2
• Because earlier synthesized data used impulse responses from one environment, the authors employ responses from five rooms with varied dimensions and reverberant properties.They also provide identical scenes in four-channel Ambisonics and microphone-array formats.Source: §1 · p. 2
• The authors propose a DCASE 2019 SELD task with a multi-room reverberant dataset, fixed evaluation setup, and modified SELDnet benchmark.The dataset includes varying numbers of overlapping sound events.Source: §1 · p. 2

2. MULTI-ROOM REVERBERANT DATASET

• The FOA and MIC datasets contain identical scenes in different four-channel formats, with 400 one-minute development recordings and 100 one-minute evaluation recordings.The development data are divided into four cross-validation splits of 100 recordings each.Source: §2 · p. 2

2.1. Real-life Impulse Response Collection

• Impulse responses were measured at 504 azimuth-elevation-distance combinations in five indoor environments with varied room properties.The collection also includes ambient noise recordings from those environments.Source: §2 · p. 2, §2.1 · p. 2

2.2. Dataset Synthesis

• Each synthesized recording spatializes isolated events with an impulse response from one environment, randomizes event start times, and adds ambient noise at an average 30 dB SNR.The isolated-event examples are divided into five sets; four support development splits and one supports evaluation.Source: §2.2 · p. 2
• MIC recordings select four Eigenmike channels for near-uniform tetrahedral coverage, while FOA recordings are produced by encoding the 32-channel signals.The FOA encoding filters are based on anechoic array-response measurements.Source: §2.2 · p. 3

2.3. Array Response

• The theoretical responses model each channel's directional sensitivity to source azimuth and elevation, using spherical harmonics for FOA and an analytical expansion for the tetrahedral array.The array expansion uses 30 terms with negligible modeling error up to 20 kHz; FOA is frequency-independent, while Eigenmike approximates that response to about 9 kHz.Source: §2.3 · p. 3

3. BASELINE METHOD

• The benchmark uses a convolutional recurrent neural network with convolutional layers followed by bidirectional recurrent layers and two output branches.The branches produce sound-event detection and direction-of-arrival estimates.Source: §3 · p. 3, Figure 1 · p. 3
• Frequency-axis max pooling reduces spectrogram dimensionality while leaving the temporal axis unchanged to preserve output time resolution.Source: §3 · p. 3
• The detection branch produces class probabilities per frame, while the localization branch estimates azimuth and elevation for each class at each frame.When multiple instances of one class occur in a frame, the model localizes one instance or oscillates between them.Source: §3 · p. 3
• Training combines cross-entropy detection loss with spherical-distance mean-square-error localization loss, weighting localization 50 times more.Source: §3 · p. 3

4.1. Evaluation Setup

• Participants report performance on the testing split of each of four cross-validation folds, with statistics pooled across folds rather than averaging fold metrics.The evaluation dataset permits participants to choose the training and validation files and ensemble size.Source: §4.1 · p. 4

4.2. Metrics

• SED is measured by F-score and error rate in one-second segments; DOA estimation uses frame-wise DOA error and frame recall.DOA error is the average angular error in degrees between predicted and reference directions.Source: §4.2 · p. 4
• DOA error matching uses the Hungarian algorithm to pair estimated and reference directions by spherical distance.Source: §4.2 · p. 4
• Frame recall accounts for frames with unequal numbers of estimated and reference directions, and final challenge positions use cumulative minimum ranks across the four metrics.Source: §4.2 · p. 4

5. RESULTS

• SELDnet performs better on FOA fold 1 and MIC fold 3, suggesting the formats contain different spectral and spatial information.The authors suggest methods could potentially benefit from mutual information between the datasets.Source: §5 · p. 4
• FOA performs marginally better than MIC overall.Source: §5 · p. 4
• SELDnet performs better when there is no polyphony across datasets.Source: §5 · p. 4
• Training with five environments performs best in the first environment.Source: §5 · p. 4

• The dataset spatializes isolated sound events using real-life impulse responses from five rooms with different acoustic properties.Source: §6 · p. 4
• Identical sound scenes are provided as four-channel Ambisonic and microphone-array recordings, alongside a predefined four-fold cross-validation split.Source: §6 · p. 4
• A convolutional recurrent neural network SELD method supplies benchmark results, with performance considered across overlapping events and reverberant scenes.Source: §6 · p. 4, Table 3 · p. 4