Unveiling Lombard speech flavors: A comprehensive categorization based on perceived intelligibility beyond loudness

The Lombard effect involves involuntary vocal adjustments in noisy environments to enhance speech intelligibility. This study categorizes Lombard speech into distinct "flavors" based on perceived intelligibility beyond loudness, focusing on how non-loudness acoustic modifications collectively improve speech intelligibility. A Lombard dataset was collected from 10 speakers across 11 noise levels (30-80 dBA) and two noise types: steady speech-shaped noise (SSN) and dynamic babble noise. Iterative paired-sample t tests were employed to assess the intelligibility of loudness-controlled Lombard speech across various noise conditions, identifying three distinct Lombard flavors for SSN and four for babble: for SSN, LF.0 (30-50 dBA), LF.1 (55-65 dBA), and LF.2 (70-80 dBA); for babble, LF.0 (30-50 dBA), LF.1 (55-60 dBA), LF.2 (65-70 dBA), and LF.3 (75-80 dBA). LF.0 represents plain speech in quiet, while LF.1 to LF.3 reflect progressively stronger adaptations yielding higher perceived intelligibility. Acoustic analysis showed that higher Lombard flavors exhibited significantly increased fundamental frequency (F0), alpha ratio, formant frequencies (F1/F2), vowel duration, and vowel ratio for each noise type. These findings provide insights into intelligibility-enhancing mechanisms in Lombard speech beyond loudness and inform speech technologies, such as telecommunications and hearing aids, aimed at optimizing speech intelligibility while reducing reliance on increased loudness.