AfroRadVoice-FR

Dataset Description

AfroRadVoice-FR is a French speech dataset composed of radiology report recordings, designed to support research in Automatic Speech Recognition (ASR) for African-accented French in medical contexts.

The dataset combines real recordings, synthetic speech, and augmented audio to address data scarcity and improve acoustic diversity in a specialized domain.

---

Motivation

Current ASR systems show strong performance in high-resource settings but fail to generalize effectively to African-accented French, particularly in specialized domains such as medical reporting.

This dataset aims to:

• Improve ASR robustness to African accents
• Enable research in low-resource medical speech recognition
• Support domain adaptation for clinical transcription systems

---

Dataset Composition

• Total samples: 562 audio recordings
• Total duration: ~4.54 hours
• Speakers: 26 (real + synthetic voices)

Data Types

• Real recordings: 177 samples (~76 minutes)
• Synthetic recordings: 317 samples
• Augmented recordings: 68 samples

Domain Coverage

Radiology reports including:

• Mammography
• Brain imaging
• Fractures
• Ultrasound
• Pediatric radiology
• etc

Splits

• Train/Validation: 487 samples
• Test: 75 samples

Data Fields

Each sample contains:

• file: relative path to audio file
• audio: waveform (Hugging Face Audio feature)
• text: normalized transcription

---

Data Collection Process

Text Source

The dataset is based on 150 radiology report conclusions collected from medical contexts.

Audio Collection

• Real recordings were collected through a web-based recording platform
• Speakers read predefined radiology reports
• Recording conditions were semi-controlled

Synthetic Data

Synthetic speech was generated using a neural text-to-speech system with voices adapted to African French accents.

Data Augmentation

To improve balance and robustness:

• Pitch shifting
• Time stretching
• Noise injection

were applied to underrepresented speakers.

---

Preprocessing

Audio

• Resampled to 16 kHz
• Normalized amplitude

Text

• Lowercased
• Cleaned (basic normalization)
• Removal of irrelevant formatting

---

Anonymization and Privacy

• No personal or demographic data is collected
• No identifiable patient information is present
• Text content consists only of generic radiology report conclusions

Synthetic samples are identifiable via filename prefixes (e.g., synth_).

---

Intended Uses

• ASR model training and evaluation
• Research on African-accented French speech
• Domain adaptation for medical transcription

---

Limitations

• Limited dataset size (~4.5 hours)
• Restricted to radiology domain
• Limited number of real speakers
• Partial reliance on synthetic data
• Does not represent all African French accents

---

Ethical Considerations

This dataset is intended for research purposes only.

It should not be used in clinical decision-making systems without proper validation and regulatory compliance.

---

Licensing and Access

• License: Creative Commons Attributions 4.0
• Access: Gated dataset (controlled access for usage tracking)

---

Future Work

• Expansion of real speaker data
• Inclusion of broader medical domains
• Improved accent diversity
• Detailed metadata documentation

---