Case Study

Medical ASR Saudi Arabic

Category:

Speech & Healthcare AI

Impact:

6 months | $425k Seed Funding

Background

Healthcare professionals often spend significant time documenting patient visits, leading to increased administrative burden and reduced focus on patient care. For Arabic-speaking regions—particularly in dialect-rich contexts like Saudi Arabia existing speech recognition systems fail to capture medical terminology and dialectal nuances accurately. Sahl.ai required a custom ASR (Automatic Speech Recognition) system to convert spoken medical consultations into precise, structured clinical notes.

Project Goals

Develop an ASR system tailored for Saudi Arabic dialects in medical contexts
Reduce documentation burden for healthcare professionals
Build scalable pipelines for data collection, training, and evaluation
Apply advanced fine-tuning techniques (LoRA, QLoRA) for performance optimization
Generate and integrate synthetic speech data to enhance low-resource datasets
Improve model robustness with audio cleaning and self-supervised learning

Our Approach

Model Development & Training

Built end-to-end training and evaluation pipelines for ASR models. Conducted 200+ experiments across datasets to refine accuracy and robustness.

Fine-Tuning Techniques

Applied LoRA and QLoRA fine-tuning, optimizing large ASR models for medical speech.

Data Augmentation & Expansion

Generated synthetic speech data to expand training resources. Developed self-supervised ASR methods, doubling productivity in data annotation and collection.

Data Collection & Annotation

Supervised teams for speech dataset collection and medical annotation. Designed audio cleaning pipelines to improve transcription quality.

Deployment & MVP Delivery

Delivered a functional MVP, enabling the startup to validate the product with investors.

Key Results

MVP delivered, securing $425k seed funding for Sahl.ai
Dialect-specific ASR system tailored for Saudi Arabic medical speech
200+ experiments run to optimize accuracy and robustness
Significant dataset expansion through synthetic data generation
Productivity in annotation doubled via self-supervised techniques
Improved input quality with advanced audio cleaning pipelines

Technologies Used

ASR Frameworks (Wav2Vec2.0, Whisper, HuBERT)

LoRA & QLoRA

Python & PyTorch

Self-Supervised Learning

Synthetic Data Generation

Audio Processing Libraries