A Persian AI-Based CBT Chatbot and Voice Analysis System for Managing Perceived Stress

This study is a three-arm, parallel-group pilot randomized controlled trial designed to evaluate the feasibility, acceptability, psychological safety, and preliminary effectiveness of a Persian-language artificial intelligence (AI)-based cognitive behavioral therapy (CBT) chatbot for the management of perceived stress. A separate research component will use voice recordings collected during the study to develop and preliminarily evaluate an AI model for stress detection. The voice-based model is independent of the randomized psychological intervention and will not be used to guide or modify the intervention during this trial. The study consists of two complementary research components. The first component is a pilot randomized controlled trial evaluating an AI-based CBT chatbot. The second component involves the collection and analysis of voice data for the development of a machine-learning model capable of detecting or estimating stress from acoustic and paralinguistic characteristics of speech.

Randomized Controlled Trial Component

Following completion of the study entry procedures and baseline assessments, eligible participants will be randomly allocated in a 1:1:1 ratio to one of three parallel study conditions: an AI-based CBT chatbot intervention, an active mental health content condition, or a waitlist control condition. The intervention period will last 14 days. Randomization will be performed using variable block randomization, with stratification based on baseline stress level and sex. The allocation sequence will be generated independently or through a secure allocation procedure. Due to the nature of the psychological interventions, participants cannot be blinded to the type of support they receive. Where feasible, the data analyst responsible for the primary analysis will remain unaware of group allocation until completion of the main analysis. Participants assigned to the experimental condition will interact with a Persian-language AI-based chatbot designed to deliver structured psychological support based on cognitive behavioral therapy principles. The intervention will be delivered through structured sessions during the 14-day study period. The chatbot intervention is designed to incorporate core CBT-oriented components including psychoeducation about stress, identification of stress-related thoughts and behavioral patterns, cognitive restructuring, development of adaptive coping strategies, behavioral exercises, and stress-management techniques. The system is intended to provide interactive and contextually relevant conversations while maintaining the predefined scope of the intervention. A large language model will support the conversational functionality of the chatbot. The language model will operate within a predefined prompting, intervention, and safety framework rather than providing unrestricted mental health advice. The system will be designed to maintain conversations within the intended CBT and stress-management scope and to reduce the likelihood of inappropriate, unsafe, or clinically unsupported responses. Before participant use, candidate language models and intervention components will undergo preliminary evaluation using predefined mental health scenarios. These evaluations will include CBT-related scenarios, psychologically sensitive situations, crisis-related scenarios, and adversarial or Red-Team testing intended to identify unsafe, inappropriate, or out-of-scope model behavior. Mental health professionals will contribute to the evaluation of therapeutic appropriateness, adherence to CBT principles, empathy, safety, and response quality. Participants assigned to the active comparator condition will receive non-adaptive mental health and stress-management content. This content may include psychoeducational information and structured material related to stress management, relaxation, breathing techniques, mindfulness, coping strategies, and self-care. Unlike the experimental condition, the active comparator will not provide individualized, adaptive, conversational CBT interactions generated in response to participant messages. Participants assigned to the waitlist control condition will not receive an active psychological intervention during the 14-day study period. They will complete the study assessments according to the study schedule and will remain subject to the same safety procedures applicable to the study population.

Psychological Safety and Human Oversight

Psychological safety is an important component of the study design. The chatbot will be explicitly presented to participants as an artificial intelligence system and not as a human therapist, psychiatrist, psychologist, or substitute for professional mental health care. Participants will be informed that AI-generated responses may have limitations and should not be interpreted as independent medical or psychiatric diagnosis or treatment. The system will incorporate predefined safety guardrails intended to identify potentially high-risk interactions, including indications of acute psychological crisis or possible self-harm or suicidal risk. When a potentially high-risk situation is identified, the automated intervention will not be relied upon as the sole response mechanism. The study safety protocol will provide procedures for escalation, human review where appropriate, and referral to suitable professional or emergency mental health services. A human-in-the-loop framework will therefore be incorporated into the safety architecture of the chatbot component. Mental health professionals will participate in the development and review of CBT protocols, assessment of chatbot response quality, review of potentially unsafe interactions, evaluation of therapeutic fidelity, and review of relevant safety events. Interaction data generated during the intervention may also be used to characterize participant engagement and to evaluate how the system performs within the intended CBT framework. These data may include the number and timing of interactions, session duration, participant engagement with exercises, and anonymized conversational information required for research evaluation.

Separate Voice Data and AI Model Development Component

In addition to the randomized intervention, participants will provide voice recordings for a separate AI model-development component. Voice recordings will be collected at predefined study time points, including baseline and the end of the study period. The voice recordings are collected solely as research data for the development and preliminary evaluation of an AI-based stress-detection model. Voice analysis is not part of the CBT intervention delivered to participants. Specifically, voice-derived information will not be used during this trial to assign participants to study groups, determine eligibility for a particular intervention, select CBT content, personalize chatbot conversations, modify the intensity or duration of the intervention, generate chatbot responses, provide feedback to participants, or make clinical decisions. The chatbot will therefore operate independently of the voice-analysis model throughout the randomized trial. No prediction generated by the voice model will be communicated to the chatbot as an input for intervention delivery. Similarly, participants will not receive clinical recommendations or psychological treatment decisions based on the output of the experimental voice model. The purpose of collecting the voice recordings is to create a research dataset that can be used to train and preliminarily validate a machine-learning model capable of identifying patterns in speech that may be associated with stress. Before model development, voice recordings will undergo quality-control procedures. Technical characteristics of the recordings, such as signal-to-noise ratio, clipping, silence ratio, and available recording-device metadata, may be evaluated to identify recordings with insufficient technical quality. Recordings that do not meet predefined quality requirements may be excluded from model development or recollected when appropriate. Following quality control and preprocessing, acoustic and paralinguistic characteristics will be extracted from the voice recordings. These may include Mel-Frequency Cepstral Coefficients (MFCCs), fundamental frequency or pitch, signal energy, jitter, shimmer, spectral characteristics, and other relevant speech-derived variables. These features will be transformed into numerical inputs suitable for machine-learning analysis. The extracted voice features, together with the study reference measures of stress, will be used to train and preliminarily evaluate the stress-detection model. Because this is a pilot study with a limited dataset, the objective of the voice component is not to establish a clinically validated diagnostic system. Instead, it is intended to assess the feasibility of developing a stress-detection model from Persian-language voice data, explore potentially informative acoustic features, and generate preliminary estimates of model performance that can guide larger future validation studies. Model development and validation procedures will be structured to reduce inappropriate information leakage between training and evaluation data. Where applicable, validation will be performed at the participant level so that voice data from the same individual are not inappropriately distributed across training and validation datasets. The model will be evaluated using predefined discrimination and calibration measures appropriate for the intended stress-detection task. Exploratory analyses may also investigate whether model performance differs across relevant demographic subgroups and whether particular acoustic characteristics contribute more strongly to stress prediction. These analyses are exploratory and are intended to support future model refinement and assessment of algorithmic fairness. Importantly, the model developed from these recordings will remain an investigational research model during this study. It will not be used as a standalone diagnostic tool, will not replace standardized psychological assessment, and will not replace evaluation by a qualified mental health professional.

Pilot Nature of the Study

The study is designed as a pilot investigation rather than a definitive efficacy trial. The randomized component is intended to generate preliminary evidence regarding the practical implementation of the AI-based CBT intervention, participant engagement and acceptability, psychological safety, and potential effects on perceived stress. Information generated in this pilot study will be used to refine the intervention protocol, safety procedures, study workflow, AI system, and methodological assumptions required for a subsequent larger-scale randomized controlled trial. Similarly, the voice component is intended to provide an initial dataset and proof-of-concept evaluation for future development of an AI-based stress-detection system. Further development would require larger and more diverse datasets, broader representation of speakers and recording conditions, and independent external validation before any future clinical application could be considered.

Data and Privacy Considerations

Study data will include questionnaire-based psychological assessments, study participation and engagement information, chatbot interaction data, and separately collected voice recordings and derived acoustic features. Particular attention will be given to the protection of voice data because raw voice recordings may contain identifiable characteristics of the speaker. Access to sensitive study data will therefore be restricted to authorized research personnel, and appropriate data-security, access-control, de-identification, and storage procedures will be applied according to the study protocol and ethics requirements. Voice recordings and voice-derived data will be handled separately from their intended use in the randomized psychological intervention. The collection of voice data is for research and AI model development purposes and does not alter the intervention that participants receive during the trial.

Overall, the study is intended to independently investigate two complementary questions: whether a Persian-language AI-based CBT chatbot is feasible, acceptable, safe, and potentially useful for perceived stress management, and whether voice recordings collected from the study population can support the preliminary development of an AI model for stress detection. The results of the voice-analysis component will not influence the delivery or evaluation of the randomized intervention during the current trial.

Inclusion criteria

• Age between 18 and 45 years
• Mild or moderate perceived stress based on the Perceived Stress Scale-10 (PSS-10) screening assessment
• Ability to use a smartphone
• Access to the internet during the study period
• Ability and willingness to provide electronic informed consent
• Willingness to participate in study assessments and complete the required study procedures

Exclusion criteria

• Severe perceived stress at the initial screening assessment
• Active suicidal ideation or risk of self-harm
• Acute psychological crisis requiring immediate professional intervention
• Active psychosis
• Active mania
• Severe substance use disorder
• Psychiatric hospitalization in recent months
• Initiation of a new psychiatric medication or change in the dose of a psychiatric medication within the previous four weeks
• Concurrent participation in intensive psychotherapy or another intensive psychological intervention
• Any mental health condition requiring immediate specialist evaluation or treatment