README.md

---
dataset_info:
  features:
  - name: audio
    dtype:
      audio:
        sampling_rate: 16000
  - name: speaker_id
    dtype: string
  - name: gender
    dtype: string
  - name: emotion
    dtype: string
  - name: transcript
    dtype: string
  - name: ipa
    dtype: string
  splits:
  - name: train
    num_bytes: 1116848970
    num_examples: 3000
  download_size: 1043404503
  dataset_size: 1116848970
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
license: apache-2.0
task_categories:
- audio-classification
- automatic-speech-recognition
language:
- fa
tags:
- ser
- speech-emotion-recognition
- asr
- farsi
- persian
pretty_name: Modified ShEMO
---

# Dataset Card for Modified SHEMO

## Dataset Summary

This dataset is a corrected and modified version of the **Sharif Emotional Speech Database (ShEMO)** named [**modified_shemo**](https://github.com/aliyzd95/modified-shemo). The original dataset contained significant mismatches between audio files and their corresponding transcriptions. This version resolves those issues, resulting in a cleaner and more reliable resource for Persian Speech Emotion Recognition (SER) and Automatic Speech Recognition (ASR).

### Curation and Correction

The original ShEMO dataset suffered from incorrectly named transcription files, leading to a high baseline Word Error Rate (WER). We addressed this by:
1.  Using an ASR system with a 4-gram language model to identify and re-align mismatched audio-text pairs.
2.  Correcting 347 files that had high error rates.

This modification process significantly improved the dataset's quality, **reducing the overall WER from 51.97% to 30.79%**. More details are available at this [GitHub repository](https://github.com/aliyzd95/modified-shemo) and in this [paper](https://www.researchgate.net/publication/365613775_A_Persian_ASR-based_SER_Modification_of_Sharif_Emotional_Speech_Database_and_Investigation_of_Persian_Text_Corpora).


## Dataset Statistics

Below are the detailed statistics for this dataset, calculated using an analysis script.

### Emotion Distribution
```text
==================================================
           Emotion Distribution
==================================================
   Emotion  Percentage (%)
   neutral           38.66
     happy           06.76
       sad           12.03
     angry           34.73
  surprise           06.73
      fear           01.06
--------------------------------------------------
```

### Overall Statistics
```text
============================================================
        Overall Statistics for Modified SHEMO Dataset
============================================================
📊 Total Number of Files: 3000
⏰ Total Dataset Duration: 3.43 hours
🗣️ Total Number of Speakers: 87
------------------------------------------------------------
```

### Gender-Emotion Breakdown
```text
==================================================
           Gender-Emotion Breakdown
==================================================
gender     female  male
emotion                
anger         449   593
fear           20    12
happiness     113    90
neutral       346   814
sadness       224   137
surprise      111    91
--------------------------------------------------
```

## Supported Tasks

* **Speech Emotion Recognition:** The dataset is ideal for training models to recognize various emotions from speech. The `emotion` column is used for this task.
* **Automatic Speech Recognition:** With precise transcriptions available in the `transcript` column, the dataset can be used to train ASR models for the Persian language.

## Languages

The primary language of this dataset is **Persian (Farsi)**.

## Dataset Structure

### Data Instances

An example from the dataset looks as follows:
```json
{
  "audio": {
    "path": "/path/to/F21N05.wav",
    "array": ...,
    "sampling_rate": 16000
  },
  "speaker_id": "F21",
  "gender": "female",
  "emotion": "neutral",
  "transcript": "مگه من به تو نگفته بودم که باید راجع به دورانت سکوت کنی؟",
  "ipa": "mæge mæn be to nægofte budæm ke bɑyæd rɑdʒeʔ be dorɑnt sokut koni"
}
```

### Data Fields

* `audio`: A dictionary containing the audio file's array and its sampling rate (set to 16kHz in this dataset).
* `speaker_id`: A unique identifier for each speaker (e.g., `F01` or `M23`).
* `gender`: The gender of the speaker (`female` or `male`).
* `emotion`: The emotion label for each audio file (`angry`, `fear`, `happy`, `sad`, `surprise`, `neutral`).
* `transcript`: The precise orthographic transcription of the utterance in Persian.
* `ipa`: The phonetic transcription of the utterance according to the IPA standard.

### Data Splits

The dataset does not have predefined `train`, `validation`, and `test` splits by default. Users can easily create their own splits using the `.train_test_split()` method from the `datasets` library.

## How to Use

To load the dataset, use the `datasets` library:

```python
from datasets import load_dataset

dataset = load_dataset("aliyzd95/modified_shemo")

print(dataset["train"][0])
```

## Additional Information

### Citation Information

If you use this dataset in your research, please cite the original paper:

```bibtex
@misc{https://doi.org/10.48550/arxiv.2211.09956,
  doi = {10.48550/ARXIV.2211.09956},
  url = {https://arxiv.org/abs/2211.09956},
  author = {Yazdani, Ali and Shekofteh, Yasser},
  keywords = {Audio and Speech Processing (eess.AS), Artificial Intelligence (cs.AI), Sound (cs.SD), FOS: Electrical engineering, electronic engineering, information engineering, FOS: Electrical engineering, electronic engineering, information engineering, FOS: Computer and information sciences, FOS: Computer and information sciences, I.2, 68T10 (Primary) 68T50, 68T07 (Secondary)},
  title = {A Persian ASR-based SER: Modification of Sharif Emotional Speech Database and Investigation of Persian Text Corpora},
  publisher = {arXiv},
  year = {2022},
  copyright = {arXiv.org perpetual, non-exclusive license}
}
```