# ml-switchboard-affect

**Repository Path**: mirrors_apple/ml-switchboard-affect

## Basic Information

- **Project Name**: ml-switchboard-affect
- **Description**: No description available
- **Primary Language**: Unknown
- **License**: Not specified
- **Default Branch**: main
- **Homepage**: None
- **GVP Project**: No

## Statistics

- **Stars**: 0
- **Forks**: 0
- **Created**: 2025-09-25
- **Last Updated**: 2026-02-28

## Categories & Tags

**Categories**: Uncategorized

**Tags**: None

## README

# Switchboard Affect (SWB-Affect)
The Switchboard Affect dataset contains perceptual emotion annotations for 10,000 publicly available audio segments. This amounts to 25 hours of speech. The source audio files can be acquired from LDC (https://catalog.ldc.upenn.edu/LDC97S62) and we include a script to extract audio segments from the corpus.

## Annotation Descriptions 
Each segment was annotated independently by 6 graders who passed a training and certification process. The set of labels includes both categorical and dimensional emotions, and we provide a detailed format (annotations from each grader) as well as a consensus format (annotations aggregated for each segment). 

### Emotion Labels 
- **Categorical emotions**.
  
  Selections of primary and secondary emotions from the following set:

    |            |            |            |
    |------------|------------|------------|
    | Anger      | Contempt   | Disgust    | 
    | Sadness    | Fear       | Surprise   |  
    | Happiness  | Tenderness | Calmness   |
    | Neutral    | Other

- **Dimensional emotions**.
  
  Ratings for valence, activation, and dominance ranging from 1 to 5:
    |            |     1      |     5      |
    |------------|------------|------------|
    | Valence    | negative   | positive   | 
    | Activation | drained    | energetic  |  
    | Dominance  | weak       | strong     |

### Data Files 
- `labels_detailed.csv` includes annotations from each annotator, unaggregated. 
- `labels_consensus.csv` includes consensus annotations aggregated for each segment. For consensus on categorical emotions, 50%+ of graders need to agree on a primary or secondary emotion. For consensus on dimensional emotions, we take the mean of ratings by all annotators. 

## Segment Extraction Script 
The script `extract_segments.py` reads in the raw audio and metadata from the LDC corpus and saves .wav files for each segment. `[LDC_DIR]` refers to the folder that contains raw audio files (in subfolder `swb1_LDC97S62`) and segment metadata (in subfolder `ms98_transcriptions`). `[SEG_DIR]` refers to the folder in which you want to save the segments. 

To extract the segments, run the following from this directory
```
python3 extract_segments.py --ldc_dir [LDC_DIR] --seg_dir [SEG_DIR] 
```

## Citation 
If you find the SWB-Affect dataset or this code useful in your research, please cite the following paper:
```
@misc{romana2025,
    author       = {Amrit Romana AND Jaya Narain AND Tien Dung Tran AND Andrea Davis AND Jason Fong AND Ramya Rasipuram AND Vikramjit Mitra},
    title        = {Switchboard-Affect: Emotion Perception Labels from Conversational Speech},
    howpublished = {ACII 2025},
}
```