Ayush2312/multimodal

Spatial Audio Encoder Training Dataset (SAET)

A high-fidelity synthetic dataset designed for training audio encoders to perceive and reason about 3D soundscapes. The dataset maps binaural/stereo audio cues to precise spatial trajectories and semantic labels.

🎧 Dataset Summary

This dataset contains 10-second stereo scenes (44.1kHz) synthesized in a virtual 3D room. Each scene features 1-3 moving sound sources with ground-truth trajectory metadata sampled at 10Hz.

📊 Dataset Generation Progress (Current State)

Stage	Description	Progress	Details
1. Extraction	Mono event extraction from AudioSet-Strong	✅ Complete	224 events extracted from 70/216 segments.
2. Synthesis	3D Spatial Scene Synthesis (Target: 10k)	🔄 ~75%	7,500+ scenes generated.
3. Reasoning	QnA Pair Generation	⏳ Pending	High-level reasoning tasks (7 categories).

📐 Spatial Metadata Specification

Each audio sample is accompanied by a dense JSON metadata file (in data/scene_metadata/) and a summary entry in metadata.jsonl.

Coordinate System

• Origin: Bottom-left-front corner of the room $[0, 0, 0]$.
• Room Dimensions: $10m \times 8m \times 3m$ (Length $\times$ Width $\times$ Height).
• Listener (Mic) Position: Fixed at center $[5.0, 2.0, 1.6]$.
• Azimuth: $0^\circ$ is directly in front (+Y), $+90^\circ$ is Right (+X), $-90^\circ$ is Left (-X). Range: $[-180^\circ, 180^\circ]$.
• Distance: Euclidean distance from the microphone center in meters.

Motion Dynamics

Sources follow one of five deterministic motion profiles:

• Static: Source remains at a fixed 3D point.
• Approach: Source moves linearly towards the listener.
• Recede: Source moves linearly away from the listener.
• Lateral: Source moves across the field of view (e.g., Left-to-Right).
• Arc: Source moves in a circular path around the listener, maintaining relatively constant distance but shifting azimuth.

🧠 Reasoning Q&A Pairs (Stage 3)

A subset of scenes includes 7 question-answer pairs generated by an LLM (DeepSeek-R1-Distill-Qwen-7B) focusing on:

1. Lateral Trajectory: Directional changes (Left-to-Right, Right-to-Left).
2. Radial Change: Distance shifts (Approaching, Receding).
3. Comparative: Which source is closer/farther?
4. Temporal: Entry/Exit timings (Early, Middle, Late).
5. Relative Motion: Inter-source spatial relationships.
6. Natural Perception: Qualitative descriptions of sound movement.
7. Choreography: Overall spatial pattern recognition.

🔊 Audio Simulation Details

• Engine: PyRoomAcoustics (Image Source Method).
• Reverberation: 2nd order reflections simulated with a frequency-independent absorption coefficient of $0.25$.
• Source Events: 224 high-variety mono events extracted from 70/216 AudioSet-Strong segments, rigorously filtered for quality (Duration $\geq$ 3.0s, CLAP semantic similarity score $\geq$ 0.45).
• Format: 2-channel Stereo, 16-bit PCM, 44.1kHz.

🛠️ Data Columns (metadata.jsonl)

Column	Type	Description
audio	Audio	Path to the stereo .wav file.
scene_id	int	Unique ID matching the filename.
labels	list	Semantic classes (e.g., Crowd, Siren, Engine).
num_events	int	Number of simultaneous sources in the scene.
motion_types	list	List of motion profiles for each source.

🎯 Use Cases

1. Spatial Audio Embedding: Training models like CLAP or Wav2Vec to create embeddings that cluster by spatial location or motion type.
2. Trajectory Inference: Predicting the azimuth/distance change of a source over time.
3. Source Separation: Decoupling multiple spatialized streams in a reverberant environment.

---

Reference: This dataset follows the methodology of "Spatial Audio Question Answering and Reasoning on Dynamic Source Movements" (2024).