README.md

---
language:
- bm
pretty_name: Transcription Scorer
version: 1.1.0
tags:
- audio
- speech
- evaluation
- human-feedback
- ASR
- reward-model
- Bambara
license: cc-by-sa-4.0
task_categories:
- automatic-speech-recognition
- reinforcement-learning
- audio-classification
annotations_creators:
- expert-annotated
language_creators:
- found
size_categories:
- 1K<n<10K
dataset_info:
- config_name: default
  audio_format: arrow
  features:
  - name: audio
    dtype: audio
  - name: duration
    dtype: float
  - name: text
    dtype: string
  - name: score
    dtype: float
  total_audio_files: 2153
  total_duration_hours: ~2
- config_name: partially-reviewed
  features:
  - name: audio
    dtype: audio
  - name: duration
    dtype: float64
  - name: text
    dtype: string
  - name: score
    dtype: float64
  splits:
  - name: train
    num_bytes: 600583588
    num_examples: 1000
  - name: test
    num_bytes: 116626924
    num_examples: 200
  download_size: 695513651
  dataset_size: 717210512
configs:
- config_name: partially-reviewed
  data_files:
  - split: train
    path: partially-reviewed/train-*
  - split: test
    path: partially-reviewed/test-*
---

# Transcription Scorer Dataset

The **Transcription Scorer** dataset was created to support research in reference-free evaluation of Automatic Speech Recognition (ASR) systems using **human feedback**. Unlike traditional evaluation metrics such as WER and its derivatives, this dataset reflects judgments of ASR outputs by human raters across multiple criteria, simulating the way a teacher grades students.

## ⚙️ What’s Inside

This dataset contains **1200 audio samples** (from diverse sources including music with lyrics) totaling 2.28 hours. It is made of short to meduim length segments each associated with:

- One **transcriptions** (drawn by selecting the best hypothesis of two Bambara ASR models)
- A **score** between 0 and 100 assigned by human annotators

| bucket (s) | partially‑reviewed |
| ---------- | --------------     |
| 0.6 – 15   | 965                |
| 15 – 30    | 235                |

### Sources:
- Transcriptions were generated by two ASR models: 
  - **Djelia-V1** (proprietary, access through API)
  - **Soloni** (open-source from [RobotsMali](https://huggingface.co/RobotsMali/soloni-114m-tdt-ctc-V0))
- Additional 81 transcriptions were intentionally **randomized/shuffled** to measure baseline judgment.

Most of the audios were collected by RobotsMali AI4D Lab with the [Office de Radio et Télévision du Mali](https://www.ortm.ml/) which gave us early access to a few archives of some of their past emissions in Bamanankan. But this dataset also include a few samples from [bam-asr-early](https://huggingface.co/datasets/RobotsMali/bam-asr-early). 

The evaluation was based on the [following criteria](https://docs.google.com/document/d/e/2PACX-1vRHFEAwU4C43NUHEY85auokgiG9dJgB0ApKwY41fwFGYn7xUSl1hXnk-CBp0_67c1C7mC7jXLzte3Mu/pub) but we also left room for a personnal subjective judgement so it also include some form of human preference feedback as the annotations were partially reviwed by professional Bambara linguists. So it is a Human feedback dataset but not based on preferences only, the score is actually designed to be a refective of the quality of the transcriptions enough to be used as a proxy metric.

## **Usage**

This dataset is intended for researchers and developers who face a label scarcity situation making traditional ASR evaluation metrics like WER impossible (which is especially relevent to low resource languges such as Bambara). By leveraging human-assigned scores, it enables the development of scoring models which outputs can be used as a proxy to transcription quality. Whether you're building evaluation tools or studying human feedback in speech systems, you might find this dataset useful if you face label scarcity.

- Developing **reference-free** evaluation metrics  
- Training **reward models** for RLHF-based fine-tuning of ASR systems
- Understanding how **human preferences** relate to transcription quality  


```python
from datasets import load_dataset

# Load the dataset into Hugging Face Dataset object
dataset = load_dataset("RobotsMali/transcription-scorer", "partially-reviewed")

```

## Data Splits

- **Train**: 1000 examples (~1.92h)
- **Test**: 200 examples (~0.37h)

This initial version is only **partially reviewed**, so you may contribute by opening a PR or a discussion if you find that some assigned scores are innacurate.

## Fields

- `audio`: raw audio
- `duration`: audio length (seconds)
- `transcription`: text output to be scored
- `score`: human-assigned score (0–100)

## Known Limitations / Issues 

- Human scoring may contain inconsistencies.
- Only partial review/consensus exists — **scores may be refined** in future versions.
- The dataset is very limited in context diversity and transcription variance, only two models were used to generate transcriptions for the same ~560 audios + 80 shuffled transcriptions for baseline estimation so it will benefit from additional data from different distribution.

## 🤝 Contribute

Feel something was misjudged? Want to improve score consistency? Add transcriptions from another model ? Please open a discussion  — we **welcome feedback and collaboration**.

---

## 📜 Citation

```bibtex
@misc{transcription_scorer_2025,
  title={A Dataset of human evaluations of Automatic Speech Recognition for low Resource Bambara language},
  author={RobotsMali AI4D Lab},
  year={2025},
  publisher={Hugging Face}
}
```

---