2026 ECNP Congress

PS03-3102

Affective state-specific speech biomarkers in bipolar disorder: signal-loss in depression and signal-gain in mania

Gerard Anmella — P. Segura-Retana, M. De Prisco, R. Volta, C. Valenzuela-Pascual, M. Korniyenko, C. Escala, V. Oliva, G. Fico, J. Raduà, E. Vieta, D. Hidalgo-Mazzei, C. Escolano-Peinado · Hospital Clínic of Barcelona- Bipolar and Depressive Disorders Unit- IDIBAPS- CIBERSAM- University of Barcelona- Barcelona- Spain., Hospital Clínic of Barcelona- Bipolar and Depressive Disorders Unit- IDIBAPS- CIBERSAM- University of Barcelona- Barcelona- Spain, Barcelona, Spain

2026 ECNP Congress

Background

Speech offers a non-invasive window into affective state in bipolar disorder, integrating two complementary domains. Paralinguistic features capture how speech is produced — response latency, speech rate, vocal energy, pitch, voice quality, spectral balance, and prosodic dynamics — and have been linked to psychomotor retardation in depressive episodes and pressured speech in manic states [1,2,3]. Linguistic features capture what is said — semantic content, and affective valence, arousal, and dominance. Whether both domains track episode severity through shared or pole-specific signatures remains unclear.

Aims

To explore which linguistic and paralinguistic speech features track depressive and manic episode severity in bipolar disorder, and to compare the physiological direction of these associations across both poles.

Methods

We analysed 85 semi-structured psychiatric interviews from 61 patients with bipolar disorder (mean age 46.7 years, SD 15.3; 57.6% women) recruited within the CALIBER study at Hospital Clínic of Barcelona [4]. Sessions were obtained during depressive episodes (n=41; mean HDRS 17.1, SD 4.4), manic episodes (n=20; mean YMRS 24.0, SD 8.5), and euthymia (n=24, with HDRS and YMRS both <8 by inclusion). After voice-activity-detection filtering, 88 features were extracted: paralinguistic descriptors via openSMILE eGeMAPS [5] (pitch, loudness, voice quality, formants, spectral balance, pause timing) and linguistic descriptors from automated transcripts using spaCy with NRC-VAD and NRC-Emotion lexicons (response latency, speech rate, semantic coherence, valence, arousal, dominance). Each feature was tested for monotonic association with HDRS and YMRS using Spearman rank correlation, treating ρ as an effect-size measure in an exploratory framework.

Results

For depressive severity (HDRS), paralinguistic correlates were dominated by reductions in vocal energy and prosodic dynamics: loudness mean falling slope (ρ = −0.30), loudness variability (ρ = −0.27), and F0 variability (ρ = −0.20). Positive associations with mean unvoiced-segment length (ρ = +0.17) indicated longer silences. Linguistic correlates were weaker but convergent: slower speech rate (ρ = −0.13) and shorter mean answer duration (ρ = −0.14). The depressive signature emerged as a coherent loss of vocal signal compatible with psychomotor retardation.

For manic severity (YMRS), paralinguistic correlates were dominated by spectral bias and phonatory instability: mean Hammarberg index (ρ = +0.45), mean alpha ratio (ρ = −0.45), variability of the voiced alpha ratio (ρ = +0.30), shimmer mean (ρ = +0.27), and shorter mean voiced-segment length (ρ = −0.30). Linguistic correlates pointed to elevated affective and interpersonal drive: text-derived dominance (ρ = +0.18) and faster speech rate (ρ = +0.15), with positive trends for valence and arousal. The manic signature emerged as a spectral gain of signal with phonatory instability and assertive affective expression.

Conclusions

Linguistic and paralinguistic speech features show distinct exploratory associations with depressive and manic episode severity in bipolar disorder. The two poles follow opposite physiological signatures — depression marked by loss of vocal energy and flattened prosody, mania by spectral tension, phonatory instability, and assertive affective expression — supporting state-specific rather than shared speech biomarkers. Replication in larger and longitudinal cohorts is needed before these markers can be considered for clinical monitoring.

References

  1. Faurholt-Jepsen, M., Busk, J., Frost, M., Vinberg, M., Christensen, E.M., Winther, O., Bardram, J.E., Kessing, L.V., 2016. Voice analysis as an objective state marker in bipolar disorder. Transl. Psychiatry 6, e856.
  2. Low, D.M., Bentley, K.H., Ghosh, S.S., 2020. Automated assessment of psychiatric disorders using speech: A systematic review. Laryngoscope Investig. Otolaryngol. 5 (1), 96–116.
  3. Maran, P.L., Braquehais, M.D., Vlaic, A., Alonzo-Castillo, M.T., Vendrell-Serres, J., Ramos-Quiroga, J.A., Rodríguez-Urrutia, A., 2025. Performance of automatic speech analysis in detecting depression: Systematic review and meta-analysis. JMIR Ment. Health 12, e67802.
  4. Anmella, G., De Prisco, M., Joyce, J.B., Valenzuela-Pascual, C., Mas-Musons, A., Oliva, V., Fico, G., Chatzisofroniou, G., Mishra, S., Al-Soleiti, M., Corponi, F., Giménez-Palomo, A., Montejo, L., González-Campos, M., Popovic, D., Pacchiarotti, I., Valentí, M., Cavero, M., Colomer, L., Grande, I., Benabarre, A., Llach, C.-D., Raduà, J., McInnis, M., Hidalgo-Mazzei, D., Frye, M.A., Murru, A., Vieta, E., 2023. Automated speech analysis in bipolar disorder: The CALIBER study protocol and preliminary results. Front. Psychiatry 14, 1101423.
  5. Eyben, F., Wöllmer, M., Schuller, B., 2010. openSMILE — the Munich versatile and fast open-source audio feature extractor. In: Proceedings of the 18th ACM International Conference on Multimedia (MM’10), ACM, Firenze, pp. 1459–1462.