The Sound of Hypnotherapy: Identification of Paralinguistic Speech Features Using Acoustic Analysis and Deep Learning.
Albin Grataloup, Angélique Remacle
Journal of voice : official journal of the Voice Foundation October 16, 2026 DOI: 10.1016/j.jvoice.2026.09.066 (opens in new tab)
Study at a glance
AI-extracted from the abstract| Characteristics | Observational acoustic analysis with machine learning classification Peer reviewed |
|---|---|
| Sample size | 32 |
| Population | Recordings of 32 hypnotherapists speaking in hypnosis and control conditions during reading and spontaneous tasks |
| Measures | Praat acoustic parameters, mel spectrograms |
| Keywords | Acoustic analysis Audio classification Deep learning Feature importance Hypnotherapy Voice |
| Key findings | Hypnosis-condition speech in this sample showed a paralinguistic profile of lower voicing proportion, reduced intensity, and decreased low-frequency spectral energy. Deep learning classifiers reliably distinguished 12-second hypnosis from control segments for therapists represented in training, with F0 and F1-F3 as key predictors; spectrogram-based models performed better, suggesting additional acoustic cues. |
Abstract
Hypnosis is an altered state of consciousness characterized by focused attention and heightened suggestibility, with therapeutic applications in a range of mental and somatic conditions. While prior research has emphasized the linguistic content of hypnotic suggestions (ie, what is said), the paralinguistic and acoustic characteristics of the hypnotist's speech (ie, how it is said) remain underexplored. This study investigated condition-associated paralinguistic characteristics in recordings of 32 hypnotherapists across hypnosis and control conditions, during both reading and spontaneous tasks. We first applied classical acoustic analysis using Praat to examine common temporal, energy, frequency, and spectral parameters of connected speech on the entire recordings. The findings identify a paralinguistic profile within this sample characterized by a lower proportion of voicing, reduced intensity, and decreased spectral energy at low frequencies in the hypnosis condition. We then employed deep learning techniques, training classifiers on both handcrafted acoustic features and mel spectrograms. Within the present dataset and for therapists represented during training, the models reliably distinguished 12-second speech segments from hypnosis and control conditions, with fundamental frequency and formant-related features (F0 and F1-F3) emerging as key predictors. Spectrogram-based models achieved higher performance, suggesting that additional acoustic cues contribute to classification under the present recording protocol.