# SoundRisk Dataset

## Overview

SoundRisk is a multimodal dataset for studying financial communication signals, transcript characteristics, and market-related observations from publicly available earnings calls.

The dataset integrates earnings call transcripts, speaker information, structured annotations, and aligned financial time-series information. It is designed for research on multimodal data analysis, natural language processing, speech analysis, temporal modeling, and empirical financial communication studies.

The dataset is organized into two complementary components:

1. `soundrisk_nasdaq_earningscall_stockprice_transcripts`
2. `soundrisk_transcripts_descriptive_stat_analysis_result`

---

# Dataset Components

## 1. Earnings Call Transcript and Market Dataset

Directory:

```
soundrisk_nasdaq_earningscall_stockprice_transcripts/
```

This dataset contains processed earnings call records, transcript segments, metadata, and aligned market information.

### Dataset Statistics

- Companies: 283 publicly listed companies
- Earnings call sessions: 1,795
- Language: English
- Earnings call period: January 2021 – June 2025
- Stock price history: Daily market observations
- Speaker roles: CEO, CFO, CXO, Analyst, Host, and other participants when available

### Data Contents

Each company directory may contain:

- Earnings call metadata files
- Transcript files
- Speaker information
- Role annotations
- Stock price records
- Derived alignment information

### Transcript Information

Transcript records include:

| Field | Description |
|---|---|
| `uid` | Unique earnings call identifier |
| `ticker` | Company ticker |
| `event_time` | Earnings call timestamp |
| `segment_index` | Transcript segment order |
| `speaker` | Speaker identifier from transcript |
| `speaker_role` | Speaker role category |
| `transcripts` | Transcript text segment |
| `token_count` | Segment token count |

### Annotation Information

The dataset may include derived categorical annotations from transcript and acoustic processing pipelines.

These annotations represent computational outputs and should not be interpreted as verified human psychological states, intentions, personality traits, or emotional ground truth.

### Market Information

Aligned financial information may include:

- Historical volatility indicators
- Daily returns
- Future return windows
- Realized volatility measurements

Market variables represent historical observations and are not intended for autonomous trading or investment decisions.

---

# 2. Transcript Descriptive Statistical Analysis Dataset

Directory:

```
soundrisk_transcripts_descriptive_stat_analysis_result/
```

This dataset contains statistical summaries and analytical results generated from the processed transcript corpus.

### Included Information

Examples include:

- Transcript-level descriptive statistics
- Segment-level statistics
- Speaker-role statistics
- Temporal distribution analysis
- Annotation distribution summaries
- Communication pattern measurements

The dataset provides aggregated views of the transcript collection and supports exploratory analysis, benchmarking, and reproducibility of statistical studies.

---

# Data Processing and Front-End Processing Pipeline

The dataset is constructed through a multimodal processing pipeline that integrates
speech processing, transcript analysis, and financial time-series alignment.

The speech front-end processing component follows a physics-gated streaming
multi-task speech analysis framework, which performs reliable speech representation
extraction, temporal segmentation, and affect-related acoustic feature estimation.

The front-end processing methodology is described in:

```bibtex
@article{chen2026physics,
  title={Physics-Gated Streaming Multi-Task Speech Front-End for Reliability-Weighted Teleconference Affect Aggregation},
  author={Chen, Xiaoliang and Chang, Le and Cai, Yixuan and Huang, Yunhe},
  journal={IEEE Access},
  volume={14},
  pages={70747--70762},
  year={2026},
  doi={10.1109/ACCESS.2026.3687928},
  publisher={IEEE}
}
```
---

# Data Quality Control

Quality checks include:

- Missing file detection
- JSON and transcript format validation
- Duplicate identifier checking
- Annotation consistency checking
- Timestamp validation
- Market alignment verification

Records with incomplete information are retained where possible with explicit missing-value indicators.

---

# Intended Uses

The dataset is suitable for:

- Multimodal financial communication analysis
- Speech and transcript modeling
- Natural language processing research
- Temporal sequence modeling
- Market communication studies
- Statistical analysis of corporate disclosures
- Benchmark development for multimodal learning systems

---

# Limitations

Users should consider the following limitations:

- The dataset covers a defined group of companies and time periods.
- Public earnings calls may differ in recording quality and availability.
- Transcript quality depends on the original source material.
- Speaker role assignments may contain errors.
- Derived annotations are computational estimates rather than human-labeled ground truth.
- Market observations do not represent all possible financial responses.
- Results based on this dataset should account for temporal and company-level dependencies.

---

# Privacy and Ethical Considerations

Earnings calls are public corporate communications. However, transcript records may contain publicly available speaker names and professional roles.

Users should:

- Respect applicable data usage requirements.
- Avoid inferring private attributes from public communication data.
- Avoid using the dataset for employment, insurance, credit, surveillance, or other high-impact decisions.
- Avoid speaker identification, voice cloning, or identity inference applications.

---

# License and Redistribution

The dataset may contain components originating from different sources, including transcripts, financial records, metadata, and derived analytical outputs.

Users are responsible for:

- Checking applicable source licenses.
- Following copyright and database usage requirements.
- Ensuring appropriate redistribution and usage permissions.

---

# Reproducibility

The dataset structure is designed to support reproducible analysis workflows.

Recommended practice:

- Preserve dataset versions.
- Record preprocessing configurations.
- Document feature extraction methods.
- Report filtering criteria and missing-data handling procedures.

---

# Dataset Organization

Recommended directory structure:

```
SoundRisk/
│
├── README.md
│
├── soundrisk_nasdaq_earningscall_stockprice_transcripts/
│
└── soundrisk_transcripts_descriptive_stat_analysis_result/
```

---

# Contact

For questions regarding dataset structure, corrections, or data usage, please refer to the dataset repository information.