Your Voice, Your Phone, Your AI Therapist: Who's Listening?

A 30-Second Voice Clip May Flag Depression or Anxiety

You can hear when a friend is coming down with something even before they say 'I feel terrible' — turns out a machine can too.

Think about how much your voice changes when you're exhausted or anxious: pace, wobble, flatness, the tiny quivers between syllables. A team published results from a deep learning system — built on a modified version of Whisper, the speech recognition model from OpenAI — trained to pick those signals up automatically. Feed it 30 seconds of your voice, and it estimates your score on two standard mental health questionnaires: the PHQ-9 for depression (a nine-question clinical checklist) and the GAD-7 for anxiety (its seven-question equivalent). The model was trained on recordings from about 24,000 people and tested on roughly 5,300 more, with a careful split so no one's voice appeared in both groups. The result: 71% sensitivity and specificity simultaneously, meaning it correctly flagged about seven in ten true cases while also correctly clearing about seven in ten people who were fine. Crucially, the system is content-agnostic — it doesn't care what you say, only how you sound. Adding what you said on top (lexical features) pushed performance higher still in real-world settings. Here's the catch: 71% is meaningfully better than a coin flip, but it is nowhere near the threshold you'd need to make a clinical decision. One in three cases would still be wrong. And the recordings came from a single proprietary platform, so we don't yet know how the model performs on a hospital ward, a phone line, or a different microphone. This is a promising biomarker, not a diagnostic tool.

AI Misses Anxiety and PTSD When You Seem Like You're Coping

Imagine describing every symptom of anxiety to a doctor, and they say you're fine because you still made it to work.

A team built a benchmark from 555 real clinical interviews — the kind where a trained clinician sits with someone and works through a structured diagnostic checklist. They then ran five large language models — including LLaMA 3, DeepSeek, GPT-4o Mini, GPT-4.1 Mini, and GPT-5 Mini — through those same interviews as text, asking each model to screen for depression, anxiety, PTSD, and any mental health disorder. The accuracy range was sobering: 0.49 to 0.86 depending on model and condition. The best-performing models were GPT-4.1 Mini and GPT-5 Mini. But here's the more important finding: the researchers dug into the cases the models missed. For anxiety and PTSD especially, the model transcripts often contained the exact symptoms that should have triggered a positive result. The model still said 'no diagnosis.' Why? Because alongside the symptoms, the person also mentioned they were still functioning, had supportive friends, or said they were 'managing.' Think of it like a smoke detector that doesn't go off because the kitchen window is open — the smoke is there, but the context muted the alarm. That protective-context language systematically pushed models toward false negatives. The catch is that this analysis of model reasoning was conducted only on GPT-4.1 Mini outputs, so we don't know if other models fail the same way for the same reasons. And the overall MCC scores — a measure of how much better than random chance the models are — topped out at 0.38. That is modest. These tools are not ready to screen unsupervised.

An AI Agent Watches Cancer Survivors' Phones to Predict Emotional Crises

Your phone probably noticed you were having a rough week before you told anyone — this team built a system to act on that.

Cancer survivors experience depression and anxiety at elevated rates, but they tend not to fill in mental health diaries at exactly the moments when they most need help — a problem the researchers call the 'diary paradox.' The PULSE team, working with 50 cancer survivors, tried a different approach: let an AI agent observe passive smartphone data — step counts, screen-on time, sleep patterns, app usage — and reason its way to a prediction about whether someone wants emotional support right now. The system used a technique called ReAct, where a large language model is given a set of tools it can call autonomously, a bit like giving an intern access to a spreadsheet and letting them run their own queries rather than handing them a fixed report. This agentic approach — where the AI investigates before concluding — outperformed a simpler pipeline that just fed pre-extracted features directly to the model. For predicting 'does this person want help regulating their emotions right now,' the agentic system reached 74.3% balanced accuracy. Prior machine learning approaches on similar problems had landed at 52–60%. The catch is the study size: 50 people is a small number for a system intended to be deployed at scale, and all participants were cancer survivors from a single longitudinal study. Whether this transfers to other populations — people with depression, anxiety, or no medical diagnosis — is completely untested. And the system completes its investigation in about 45 seconds and five tool calls, which is fast, but raises obvious questions about battery, privacy, and what happens when it gets it wrong.

All three papers are circling the same idea: what if we could detect a mental health crisis without waiting for someone to report it? Voice pitch, phone behaviour, clinical interview text — each is a different window into the same problem. And each paper runs into the same wall from a different angle. The voice model is accurate enough to be interesting but not accurate enough to act on alone. The LLM screener reads every symptom correctly and still misses the diagnosis because the person seems fine on the surface. The phone-based agent works in a small study but hasn't been tested outside one very specific group. What these results collectively tell you is this: the signal is real. Mental state leaves detectable traces in how we sound, move, and behave. But we are still at the stage of proving the signal exists, not at the stage of building a reliable instrument around it. The harder engineering problem — and the harder ethical one — is still ahead.

The voice biomarker team explicitly flagged a clinical validation study as the logical next step; watch for a follow-up that tests this model in a care setting rather than a consumer platform. On the LLM screening side, the key open question is whether the 'protective context' bias the team found in GPT-4.1 Mini shows up identically in other models — that analysis hasn't been done yet. Honestly, the thing I most want to see is any of these three systems tested in a prospective trial where the output actually shapes a clinical decision. That experiment hasn't been run.

• Wikipedia: PHQ-9 — the depression questionnaire used as the target in the voice biomarker paper
• Wikipedia: Structured Clinical Interview for DSM (SCID) — the gold-standard tool used to build the LLM screening benchmark
• Wikipedia: Whisper (speech recognition system) — the model architecture the voice biomarker team adapted