Float Tanks, Voice Wobbles, and an AI That Misses Women

Tiny wobbles in your voice may signal depression, anxiety, or ADHD

Your voice wobbles in ways you can't hear — and those wobbles may be tracking your mental health in real time.

Think of a guitar string that is slightly out of tune. It still plays a note, but something in the sound is off — there's a wavering quality an experienced ear can detect. Your voice works the same way. Two acoustic properties called jitter (tiny, irregular fluctuations in pitch from syllable to syllable) and shimmer (similarly small fluctuations in loudness) are basically your vocal cords' version of that out-of-tune string. When your mental health is under strain, those fluctuations change in measurable ways. A research team systematically tested whether these and related speech features — including the emotional tone of your word choices and the grammatical complexity of your sentences — correlate with validated symptom scores across depression, anxiety, and ADHD. The finding that matters: the relationships held up not just in controlled lab datasets but in real-world clinical recordings. That is actually harder than it sounds. Most studies work only in clean, controlled conditions. Why does this matter to you? Because a voice-based screening tool would cost almost nothing to run. No blood draw, no imaging machine, no specialist visit — just a short voice recording on a phone. The catch: correlation is not a diagnosis. The study shows these features track symptom severity statistically, across groups. It does not mean your individual voice reading would be reliable enough to tell you whether you are depressed. The gap between 'this feature correlates at a population level' and 'this tool works for you, on a Tuesday, after bad sleep' is enormous. We are not there yet.

AI reads therapy transcripts for depression — but misses women more than men

The AI got the diagnosis right 86% of the time — and wrong 51% of the time — depending entirely on which condition you asked about.

Imagine a spell-checker that was trained mostly on one dialect of English. It catches the errors it has seen before and confidently misses the ones it hasn't. That is roughly what a team of researchers found when they tested five large language models — including GPT-4.1 Mini, GPT-5 Mini, LLaMA 3, and DeepSeek — on a benchmark of 555 real patient interviews, cross-referenced against structured diagnostic interviews used by actual clinicians. The models were asked to screen for four conditions: depression, anxiety, PTSD, and any mental health disorder. Results varied wildly: accuracy ranged from 0.49 (barely better than flipping a coin) to 0.86 depending on the model and the condition. A statistic called MCC — which measures whether a model is genuinely discriminating or just getting lucky — ranged from 0.16 to 0.38 across the board. Those are modest numbers, not impressive ones. The most troubling finding: depression was classified more accurately in male participants than in female participants. When the researchers dug into false negatives — cases the AI missed — they found a pattern: the models were discounting symptom descriptions when surrounded by language about coping well, having support, or maintaining daily function. Women's descriptions in this dataset more often included that kind of protective-context language. The AI read 'I have support and I'm managing' and concluded 'probably fine,' even when clear symptoms were present. The catch: demographic analyses here are descriptive, not statistically confirmed. This is a signal worth watching, not a settled finding. But it is exactly the kind of bias that, if baked into a deployed screening tool, could widen existing gaps in mental health diagnosis.

Floating in a sensory-deprivation tank creates a psychedelic-like experience in anxious people

A dark, silent tank of salt water can produce something that sounds a lot like a psychedelic trip — without any drug involved.

Picture turning down every dial in your sensory environment at once: no light, no sound, no feeling of gravity because you are floating in body-temperature water so dense with salt that you cannot sink. This is Floatation-REST — Reduced Environmental Stimulation Therapy — and it has been studied for anxiety and depression for decades. What researchers publishing this secondary analysis found is more specific: people with anxiety and depression who used float tanks reported a distinct altered state of consciousness they are now calling 'aquahenosis.' The word is borrowed from the feeling of merging with water. The state involves three components that will sound familiar to anyone who has read about psychedelic research: Oceanic Boundlessness (a sense of unity and transcendence), Disembodiment (feeling disconnected from your physical body), and Experience of Unity. The researchers compared this to a zero-gravity chair control condition and found the float tank produced significantly more of these experiences. Participants also reported heightened awareness of their own heartbeat and breathing — something clinicians call interoception — compared to the control. One detail worth noting: effects were strongest in participants who chose to float for longer and on a flexible schedule rather than a prescribed one. Self-direction seemed to matter. The catch: this is a secondary analysis of a small feasibility trial — not a definitive study. We do not know whether the altered state itself is what produces any therapeutic benefit, or whether it is just an interesting side effect of floating. The mechanism, and whether it actually helps, remains genuinely open.

Look at these three stories as a set and a pattern emerges. We are building mental health detection and treatment tools from every direction at once — your voice, your brain scan, your therapy transcript, a tank of salt water — and every single one of them is revealing the same uncomfortable truth: the signal is real, but we are not yet good enough at reading it reliably for individuals. The voice paper says: yes, speech predicts symptoms — statistically, in groups, under the right conditions. The LLM paper says: yes, AI can screen for depression — except when it can't, and the people it fails most may be the people least well-served by current psychiatry already. The float tank paper says: yes, a non-drug intervention can produce psychedelic-like states in people with anxiety — but we don't know if that state is doing anything useful. This is what a field in serious development looks like. Not solved. Not hopeless. Working.

The LLM bias finding — that AI misses women's depression symptoms more than men's — is the one I'll be tracking most closely. If that pattern replicates in larger, statistically powered studies, it becomes a regulatory question, not just a research one. On the float tank side, watch for larger controlled trials out of the Laureate Institute for Brain Research in Tulsa, which has been the main engine of Floatation-REST research and has ongoing work in treatment-resistant depression. The open question I'd want answered first: does the altered state predict clinical improvement, or are they just correlated bystanders?

• Wikipedia: Flotation therapy — background on the technique and its history
• Wikipedia: Jitter and shimmer in voice acoustics
• Wikipedia: Matthews correlation coefficient — what that number actually means