Interspeech 2025 official challenge

Introduction

Suicide is a significant global health problem, especially for young people. The SpeechWellness Challenge aims to detect suicide risk in teenagers through speech analysis. Using both spontaneous and reading speech data from 600 adolescents, including those identified as at risk, participants aim to develop models to identify digital biomarkers associated with suicide tendencies. The challenge encourages diverse approaches, including signal processing, speech and emotion recognition, and large language models. By using speech data, the challenge aspires to advance the early detection of mental health issues and provide scalable, accessible tools for intervention in adolescent suicide prevention.

The Challenge

The 1st SpeechWellness challenge consists of a suicide risk detection task, where participants are required to produce a model to predict the label (has suicide risk or not) for a subject. The training and testing dataset comprises speech recordings from 600 Chinese teenagers aged 10-18 years, with 50% identified as having suicide risk based on psychological scales. All subjects (speakers) have been anonymized before making this dataset available to the participants, which provides a unique opportunity to apply and refine advanced speech technologies for public health while preserving the privacy of all subjects.

Participants are encouraged to leverage a variety of methodologies, including but not limited to signal processing, self-supervised learning foundation models, speech recognition, and large language models. The goal is to push the boundaries of current technology to create effective tools for early detection and intervention.

The performance will be evaluated by accuracy:

Apart from accuracy, the participants are also encouraged to report other metrics (e.g., F1, AUROC).

How to Participate

To register for the SpeechWellness Challenge, please email speechwellnesschallenge@gmail.com with your contact information, affiliation, and agree to the terms and conditions. All participants must register by January 15, 2025. A confirmation email with further instructions on accessing the dataset will be sent to registered participants.

Participants may withdraw from the challenge at any time by notifying the organizers. Upon withdrawal, they must delete all copies of the dataset and related materials.

Data Access and Usage

Registered participants will be provided access to the training and testing datasets. They must sign a data usage agreement (see below), agreeing to maintain confidentiality and follow data protection protocols. The dataset is to be used solely for challenge purposes, with strict prohibitions on redistribution or any other use. Participants are responsible for securing the data from unauthorized access.

The dataset contains sensitive information from minors aged 10-18, making it essential for all participants to strictly adhere to ethical guidelines and comply with relevant legal regulations. Even if the data has been anonymized, participants MUST refrain from attempting to identify or contact any individuals whose data is included in the dataset. It is crucial to understand that anonymization does not eliminate the ethical responsibility to protect the privacy and dignity of the subjects. Any findings must be reported with the utmost care to ensure that no harm, stigma, or unintended consequences arise from the research. Additionally, participants are expected to prioritize the well-being of the individuals represented in the data by avoiding conclusions or interpretations that may inadvertently lead to their identification or negatively impact their communities. Failure to uphold these standards could result in legal or disciplinary action and could compromise the integrity of the research.

Data License Agreement

Please sign the license and return by email to speechwellnesschallenge@gmail.com 

Important Dates

The following dates are in accordance with Interspeech 2025:

Participants must submit a paper detailing their methods, experiments, and results to the Interspeech 2025 main conference (select subject area “14.15 The 1st SpeechWellness Challenge [Challenge]”), following the Interspeech 2025 submission guidelines. Papers should adhere to the Interspeech 2025 formatting instructions, and represent original, unpublished work.

The submission will undergo the official peer review process for Interspeech 2025. The evaluation will consider both quantitative performance metrics and qualitative aspects of the work, including methodological innovation, reproducibility, presentation, and ethical considerations.

🔥Submission of Test Results

Test set and baseline release

We are excited to announce the release of the challenge baselines:

• Baseline: a system uses eGeMAPS feature set with SVM classifier, which achieves an accuracy of 0.53 on dev set and 0.51 on test set. It serves as the foundational benchmark, provides a reference for all participants to build upon and improve their models.

• Baseline-bonus: a stronger system uses wav2vec 2.0 + BERT models with neural classifier, which achieves an accuracy of 0.56 on dev set and 0.61 on test set*. It serves as an additional challenge, designed for teams aiming to exceed a higher standard and earn a bonus distinction.

For more details about the baseline implementation, please refer to the GitHub repository and the challenge paper.

* FYI: For the original speech before timbre anonymisation, the accuracy is 0.58 on the dev set and 0.60 on the test set for the .

Leaderboard launched

The official challenge leaderboard is now live! Track your team’s progress and compare your performance with others:

Access the Leaderboard

Instructions for submitting test results

1. Save your predictions in a JSON file formatted as a dictionary with the structure: {id: prediction}. See  sample.json for example (0: no-risk, 1: at-risk).

2. Rename the JSON File as “{TeamName}_{ID}.json”

   “TeamName”: Team name (as it should appear on the leaderboard).

   “ID”: the submission attempt number (e.g., 1, 2, 3).

3. Email your JSON file to: speechwellnesschallenge@gmail.com.

N.B. Each team is allowed up to three submissions for the test results. Ensure that the “ID” field matches your current attempt number.

Submit paper to Interspeech 2025 main conference

Participants should submit a paper detailing their methods, experiments, and results to the Interspeech 2025 main conference (select subject area “14.15 The 1st SpeechWellness Challenge [Challenge]”), following the official Interspeech 2025 submission guidelines. The submission will undergo the official peer review process for Interspeech 2025. The evaluation will consider both quantitative metrics and qualitative aspects, including methodological innovation, reproducibility, presentation, and ethical considerations.

Due to the scientific and ethical considerations of this study, we kindly request that all participants include the following limitation statement in the papers submitted for the challenge:

“The findings presented in this study are based on the scoring framework of the MINI-KID scale, which assesses current suicide risk as at risk or no risk. This classification reflects subjects’ immediate responses to the MINI-KID assessment and should not be interpreted as a prediction of future suicidal behaviour. Although the MINI-KID suicide module is widely recognised as a gold standard for assessing current suicide risk among adolescents, it has limitations. It relies heavily on self-reported data, which may lead to underreporting or misinterpretation of symptoms, and its fixed set of items may not fully capture the complex and dynamic nature of suicidal ideation and behaviour. Accordingly, the results reported herein are strictly confined to the context of this assessment.”

Additionally, we would like to emphasise that the discussion section of your submission should avoid overgeneralising the findings beyond the specific scope of the MINI-KID framework, and please ensure that the limitations of the assessment are clearly acknowledged.

If you use our dataset, please cite the following paper:

Wen Wu, Ziyun Cui, Chang Lei, Yinan Duan, Diyang Qu, Ji Wu, Bowen Zhou, Runsen Chen, and Chao Zhang. "The 1st SpeechWellness Challenge: Detecting Suicide Risk Among Adolescents." arXiv preprint arXiv:2501.06474, 2025.

FAQ

1. Will the test set be provided later with real "label" information?

To ensure fairness in the competition, we will not release the ground truth labels for the test set. Instead, participants can submit their predictions for the test set and we will evaluate the results, as described in “Submission of Test Results” section above. We encourage participants to report results where the model is trained on the train set and evaluated on the dev set, for comparison between different works.

2. Will the top-performing teams be invited to write a paper, or will all the participants be able to submit the paper?

All participants are welcome to submit the paper. The submission will undergo the official peer review process for Interspeech 2025. The evaluation will consider both quantitative metrics and qualitative aspects, including methodological innovation, reproducibility, presentation, and ethical considerations. Top-performing in the leaderboard is neither a sufficient nor a necessary condition for paper acceptance.

Contact

For any inquiries related to the challenge, please contact, please contact speechwellnesschallenge@gmail.com

Organizers