To support the development and evaluation of the Physics-Informed Acoustic Model (PIAM), we construct a multimodal dataset categorized as follows:

1. Sound Earnings Call Data

Number of Companies: 283 NASDAQ-listed corporations
Number of Recordings: 1,795 earnings call sessions
Total Audio Duration: Approximately 1,780 hours
Time Range: January 22, 2021 – June 29, 2025
Speaker Roles Labeled: CEO, CFO, CXO (and others where applicable)
Transcripts: Available, but used primarily for alignment — acoustic signal is the primary modality
Audio Quality: High-fidelity recordings, minimally preprocessed to preserve nonlinear acoustic features
Annotations:
- Speaker segmentation
- Role labeling
- Temporal alignment with stock movement data

2. Stock Price Time Series Data

Companies Covered: Corresponding to the 283 firms in the earnings call dataset
Time Range: November 1, 1999 – July 10, 2025
Granularity: Daily resolution
Fields Included:
- Open price
- High price
- Low price
- Close price
- Volume
Usage: Time-align financial trends with vocal features extracted from earnings calls

3. Multimodal Alignment

Temporal Resolution: Precise alignment between vocal dynamics and high-frequency financial movements
Cross-Modality Mapping:
- Executive-level acoustic sequences → Acoustic State Labels (ASLs)
- Financial records → Volatility and return metrics
Applications:
- Modeling nonlinear vocal instability as a predictor of future volatility
- Benchmarking against GARCH and textual sentiment models
- Developing interpretable, ethical AI systems for financial insight and risk forecasting

This dataset serves as the backbone for studying the intersection of speech-based cognitive stress indicators and market volatility, enabling robust, real-time, and interpretable predictions grounded in physical and economic realities.

Downloads last month: 2