Speech markers of psychological change following a psychedelic 5-MeO-DMT retreat

Kuc and colleagues note that psychedelics, including the short-acting compound 5-MeO-DMT, are being studied for their potential effects on cognition, emotion, self-awareness, and mental health. The authors argue that language and voice are especially promising because they can express psychological change directly and indirectly, yet no previous longitudinal studies had tracked naturally produced speech and acoustic features before and after psychedelic use in everyday settings. They therefore frame the paper as addressing a gap in ecological, time-resolved evidence about how psychedelic experiences may be reflected in language and vocal behaviour. The study set out to develop and test a framework for analysing baseline speech, pre-to-post psychedelic change, and the extent to which speech markers might predict subjective experience and later psychological outcomes. In particular, the authors aimed to examine whether vocabulary, emotional expression, and voice quality changed around a single 5-MeO-DMT session, whether these changes unfolded over time, and whether pre-session speech features related to preparedness, acute subjective effects, emotional breakthrough, and post-experience well-being. The paper presents this as the first longitudinal investigation of both textual and acoustic speech markers around a psychedelic retreat, using daily voice journals collected through a chatbot-based ecological momentary assessment approach.

The researchers conducted a naturalistic study at the Tandava Retreat Centre with healthy adult participants recruited globally through the F.I.V.E. platform, social media, and word of mouth. The final cohort comprised 29 participants, most of whom were White/Caucasian and non-religious, and all had prior experience with 5-MeO-DMT to varying degrees. Participants attended one of six three-day retreats. Day 1 involved orientation and briefing, Day 2 involved the 5-MeO-DMT session and primary data collection, and Day 3 focused on integration through facilitator-guided discussion. On the dosing day, participants received a single 12 mg dose of synthetic 5-MeO-DMT by vaporisation under supervised, standardised conditions. The compound was inhaled in one breath while participants were in a semi-supine position on a padded recliner. The retreat setting was described as controlled but naturalistic rather than laboratory-based. Speech data were collected via RetreatBot, a custom Telegram chatbot designed to prompt daily voice journals from seven days before dosing to seven days after, with analysis extended to a 28-day window centred on the session because engagement continued beyond the planned period. The chatbot used a rule-based decision-tree structure rather than an adaptive conversational AI. Voice entries were transcribed automatically and checked manually, with sensitive information redacted. The researchers analysed transcribed text using LIWC-22 bag-of-words categories and a RoBERTa model trained on GoEmotions to classify sentence-level emotional content. Acoustic features were extracted with openSMILE/eGeMAPS, yielding 88 voice parameters including pitch, jitter, and shimmer. For analysis, the authors used linear mixed-effects models to compare pre- and post-session language features, paired t-tests for acoustic means, and day-by-day as well as weekly temporal analyses. They also examined whether pre-dosing language predicted five psychometric outcomes: psychedelic preparedness, ego dissolution, oceanic boundlessness, emotional breakthrough, and well-being. Pearson correlations were corrected for multiple comparisons using the Benjamini-Hochberg false discovery rate procedure, and significant features were entered into regression models. Ridge regression with leave-one-out cross-validation was used to test whether multivariate prediction generalised. In addition, principal component analysis was applied to pre-dosing emotion scores to derive broader emotional profiles. Sensitivity analyses tested whether findings were explained by participants’ prior 5-MeO-DMT exposure. The extracted text also notes that the study had ethical approval, participants gave informed consent, and the researchers used a range of safeguards including manual transcription checks, anonymisation, and false discovery rate correction.

The study collected 288 unique voice journals from 29 participants, comprising 4127 sentences and just over 506 minutes of audio. On average, each participant contributed 10 voice journals, although engagement varied. Most participants provided both pre- and post-ceremony recordings, but sample sizes differed somewhat across analyses because some questionnaires or recordings were missing. Following 5-MeO-DMT, several clear language shifts were observed. Cognitive language increased, driven by more use of cognitive process, insight, tentative, and certitude words. Social language decreased, particularly words related to communication and social referents. In the linguistic category, adverb use increased and conversational markers declined. Time-related language shifted towards the past and away from the future. Emotion classification with the RoBERTa model showed reduced excitement, nervousness, and fear, alongside increases in admiration, relief, joy, and realisation. In other words, the post-session speech sounded less anticipatory or anxious and more reflective or meaning-oriented. Acoustic analyses found significant post-dosing increases in average local jitter and shimmer, alongside a decrease in normalised jitter variability. When the authors ranked effects by Cohen’s d, the largest changes were the decrease in excitement, the decrease in jitter variability, and the increase in past-focused language. Across the full set of features, 13 increased and 9 decreased after dosing. The day-by-day analyses showed a crossover pattern: social language was initially above baseline and then dropped after dosing, while cognitive language rose above baseline and stayed elevated into the second post-dosing week. Vocal deviation analyses showed that pitch, jitter, and shimmer followed distinct trajectories around the retreat. Pitch deviations were more common before dosing and became less frequent afterwards. Jitter and shimmer showed more pronounced post-session changes, with high deviations peaking shortly after the session and low deviations becoming less common in the post-period. The authors report that these temporal patterns were broadly consistent in individual-level analyses, although shimmer was more variable and did not always reach corrected significance. For prediction, pre-dosing language was most clearly related to psychedelic preparedness. Positive tone, work-related language, and lifestyle references were each correlated with higher preparedness scores, but only positive tone remained an independent predictor in the joint regression model. In the principal component analysis of pre-session emotions, PC1 reflected positive emotional engagement and higher scores on this component predicted higher preparedness and better well-being. PC2 reflected negative emotional states but did not predict outcomes. PC3 reflected reflective integration and predicted stronger emotional breakthrough. The ridge regression models using all LIWC features did not generalise well to held-out participants, suggesting overfitting. The sensitivity analyses indicated that the main pre-post effects were not explained by lifetime 5-MeO-DMT use. No baseline speech features were significantly associated with prior lifetime use after correction, the direction of the main effects was preserved in both lower- and higher-experience groups, and the effects remained after including lifetime use as a covariate. Only one interaction, past-focused language, was moderated by prior experience, with a larger post-session increase among more experienced users.

Kuc and colleagues interpret the findings as evidence that naturalistic speech and voice can capture psychologically meaningful change around a psychedelic retreat. They argue that the post-session rise in cognitive language and fall in social language suggest a shift away from externally oriented discourse towards introspection, self-examination, and meaning-making during the two weeks after dosing. The authors present the observed increase in past-focused language and the reduction in future-oriented expression as consistent with a move from anticipation to retrospective processing. They also interpret the emotional-language findings as showing that speech did not simply become more positive overall, but changed in the specific kinds of emotions being expressed. Anticipatory emotions such as excitement, nervousness, and fear declined, whereas joy, admiration, and realisation increased. The authors link this to theories that psychedelics may promote insight and perspective change, and they note that these patterns may reflect an active integration process rather than a straightforward mood elevation. Similarly, the increase in jitter and shimmer is discussed as potentially reflecting a transient, psychologically malleable state rather than distress, especially since anxiety was not elevated in the sample. Relative to earlier research, the authors state that their work extends prior language-based psychedelic studies by moving beyond acute effects, retrospective trip reports, or single-time-point therapy transcripts. They argue that the study adds two major advances: longitudinal, ecologically valid voice journalling and simultaneous analysis of lexical, emotional, and acoustic markers. They also suggest that these multimodal methods may be useful for tracking preparation and integration processes in future psychedelic research. In discussing prediction, the authors say that positively valenced pre-dosing language appears to index readiness for the experience and may be associated with better outcomes. They further interpret the PCA findings as indicating that a broader pattern of positive emotional engagement predicts both preparedness and well-being, while reflective emotional complexity may facilitate emotional breakthrough. They suggest that some experiential outcomes are not well captured by simple word categories, which is why transformer-based emotion modelling added value. The main limitations they acknowledge are the small, self-selecting sample; the participants’ extensive prior psychedelic experience; the fact that the sample was made up of healthy volunteers rather than clinical patients; and the retreat context, which may have shaped responses independently of the drug. They also note the open-label design, absence of a control group, possible limitations of dictionary-based and model-based language tools, and the lack of follow-up assessments to establish durability. The authors stress that the findings are hypothesis-generating and require replication in larger, blinded, controlled studies. They also note that journalling itself could have influenced the observed changes. For future work, the authors propose larger studies with active controls or placebo arms, comparisons with non-psychedelic retreats or other intense experiences, multimodal links to neuroimaging and physiological measures, and longer follow-up to determine persistence. They suggest that automated speech monitoring could eventually help guide preparation, support integration, and personalise psychedelic-assisted care.

The authors conclude that everyday language and vocal patterns are sensitive indicators of psychological change surrounding a 5-MeO-DMT retreat, although controlled studies are still needed to separate drug effects from contextual factors. They state that the observed move towards introspective, cognitively oriented speech, together with changes in voice quality, points to substantial shifts in self-related processing after a single retreat experience. They also conclude that longitudinal voice analysis may become a useful tool for mapping preparation and integration, and for providing future real-time support in psychedelic-assisted therapy or retreat settings.