Synthetic Medical Speech Dataset for Clinical ASR Fine-Tuning

A synthetic medical speech dataset contains 101,475 audio-text pairs totaling 184.1 hours of 16 kHz mono speech. It was generated by IntelMedica using the Kokoro-82M TTS system with 19 voices across three English accent groups, focusing on clinical and nursing terminology. The dataset version was noted in April 2026.

Use Cases

• Fine-tune Whisper-based ASR models on clinical terminology using the synthetic audio-text pairs.
• Evaluate ASR model performance on medical terms sourced from RxNorm API and FDA data referenced in the dataset.
• Train multi-accent medical ASR systems leveraging the 19 distinct synthetic voices across three English accent groups.
• Benchmark speech synthesis quality for medical domain terms using the Kokoro-82M TTS generated audio.
• Augment real medical speech datasets with the 184.1 hours of synthetic speech to improve model robustness.

Strengths

• 101,475 audio-text pairs provides a substantial corpus for training.
• 184.1 hours of speech offers extensive audio data for model fine-tuning.
• Focus on clinical and nursing terminology from specific sources like RxNorm and FDA adds domain relevance.

Limitations

• All data is synthetically generated, which may not capture nuances of real human medical speech.
• Limited to three English accent groups, potentially lacking global dialect coverage.
• Companion version v1 has more samples (125,500) and hours (~257), suggesting v2 is a focused subset.

Provenance

Source

IntelMedica.

Collection Method

Synthetically generated using the Kokoro-82M text-to-speech system.

Freshness

Dataset page was updated in April 2026.

Geography

English accent groups, specific regions not detailed.