LungTTA: Text-to-Audio Generation of Synthetic Lung Sounds for Respiratory Health

Abstract: Respiratory audio analysis is still limited by data scarcity, as real recordings are difficult to collect and often involve privacy and clinical constraints, which makes it harder to train robust machine learning models. We introduce LungTTA, a text-to-audio framework based on a latent diffusion model, which generates respiratory sounds such as cough, breathing, and phonation from structured prompts. The model is fine-tuned on 116,660 publicly available recordings and includes a retrieval-based memory component together with watermarking for traceability. We evaluate the generated audio using Fréchet Audio Distance (FAD), Kullback–Leibler (KL) divergence, and Inception Score (IS), and also introduce PRISM (Pulmonary Respiratory Integrity & Similarity Metric) a domain aware metric designed to capture respiratory signal structure. LungTTA achieves a FAD of 2.72, KL of 0.50, IS of 1.22, and PRISM of 0.23, compared to Stable Audio Open (6.73, 0.67) for FAD and KL, Make-An-Audio (1.54) for IS, and RespAgent (0.24) for PRISM. In human evaluation, LungTTA achieves 80.91 (Overall Quality, OVL) and 75.13 (Relevance to Text, REL), compared to RespAgent (59.27, 58.97) and EZAudio (55.24, 52.69), while expert assessment yields 58.33 (OVL), 44.44 (REL), and 38.89 (Clinical Relevance for Assessment, CRA), compared to RespAgent (56.94, 43.06, 36.11) and EZAudio (36.11, 29.17, 33.33). In a downstream COVID-19 cough classification task, LungTTA improves performance under a VGGish-based setting, increasing AUC from 0.7331 (no augmentation) and 0.7631 (classical augmentation) to 0.7701 using LungTTA. These results demonstrate that LungTTA-generated synthetic respiratory audio can be used as an effective data augmentation method.

Submission Type: Regular submission (no more than 12 pages of main content)

Assigned Action Editor: ~Hao_Tang1

Submission Number: 8456