MedMosaic: Large-Scale Medical Audio Benchmark Released
A team of researchers has introduced MedMosaic, a dataset aimed at evaluating language and audio reasoning models within authentic clinical settings. Gathering medical audio data poses challenges due to privacy laws and significant annotation expenses. MedMosaic includes a variety of audio formats, such as physiological sounds linked to conditions, synthetic speech with artifacts, and actual clinical dialogues of differing lengths. The dataset comprises 46,701 question-answer pairs in multiple-choice, sequential multi-turn, and open-ended styles. It facilitates thorough assessment of multi-hop reasoning and answer creation. This benchmark evaluates 13 different audio and language models.
Key facts
- MedMosaic is a medical audio QA dataset for benchmarking language and audio reasoning models.
- Dataset includes physiological sounds, synthetic voices, and real clinical conversations.
- Contains 46,701 question-answer pairs in multiple-choice, multi-turn, and open-ended formats.
- Benchmarks 13 audio and language models.
- Addresses challenges of privacy regulations and high annotation costs in medical audio.
Entities
—