Callyope

The product and the model are one thing, and the model was built rather than borrowed.

The company trained its own audio and language foundation model on psychiatric data specifically, rather than fine tuning a general purpose speech system, and describes it as covering voice, language, clinical context and behavioral signals in one transdiagnostic model instead of one model per condition. Every function on both sides of the product runs on it: symptom scoring from thirty seconds of speech, chart review generation from complex records, ambient consultation notes, and document drafting. Strip the model out and what remains is a voice recording application and an empty clinician dashboard.

The mechanism claim is backed by primary work rather than asserted in marketing. The founding team is a research team, the technical founder's doctorate was on automatic assessment of cognitive, linguistic and emotional disorders affecting speech in a neurological disease, and the company publishes on the model at speech science venues with a scientific committee named on the site.

Nothing else is sold. There is no services layer, no practice management system, no record system and no non model module underneath.

The boundary is stated clearly and the escalation behind it is undefined, which matters more here than for a product that produces a document.

What is stated is unambiguous. The company says the system is designed to support clinicians rather than replace them, that it handles time consuming administrative work and provides objective data to inform clinical decisions, and that the clinician remains in full control of diagnosis, treatment planning and patient care. Output is a symptom score and a set of drafted documents the clinician reviews. Patients are told they decide when and how their voice is analyzed, through a voice journal entry, a call with a trusted contact, or a brief automated check in, and that recordings happen only with explicit consent. Stating what the patient controls, not only what the clinician controls, is a real scope boundary.

The gap is what happens between the model detecting something and a person seeing it. The product's stated purpose is catching deterioration between appointments in patients who may see a psychiatrist every four to six weeks, and the company describes detecting subtle changes that may signal symptom progression or relapse risk before they are apparent in clinical observation. That is a claim to notice things earlier than a person would. Nothing published states who is notified when a score moves, through what channel, within what time, whether any threshold triggers an alert at all, or what happens if nobody opens the dashboard. In a schizophrenia relapse monitoring product, that is the operating question.

A continuous score presented longitudinally also carries an implied authority a single reading does not, and no confidence interval or uncertainty is published alongside it.

The transparency is in the scientific literature rather than on a marketing page, and the production model is a different artifact from the papers.

Seven publications are listed with authors and years, and the subject matter sits close to the mechanism the product depends on: learning speaker and health related representations from natural language supervision, cross linguistic masked autoencoding for detecting disease from speech, robust fine tuning of speech recognition for disordered speech, quantized approximate signal processing toward homomorphic encryption for audio, in context learning to detect adolescent suicide risk from transcripts, automated speech content analysis for depression, and earlier work on learning strides in convolutional networks. Separate work examines which pretext tasks in speech foundation models transfer to mental health detection and how model layers encode the relevant features. Part of it reports on a public third party dataset alongside the company's own, which makes some of the performance externally checkable.

The approach is named: an audio and language foundation model trained on voice, language, clinical context and behavioral signals, built in house rather than adapted from a general purpose speech system, and described as transdiagnostic by design rather than one model per condition.

The production artifact is what stays undocumented. Weights are not released, no model version, update cadence or model card is published, and the proprietary psychiatric training corpus is uncharacterized. The marketed claim of assessing ten or more symptoms from thirty seconds of speech at over ninety percent accuracy is tied to no published evaluation, and a fifty percent chart review accuracy advantage over general purpose language models is asserted with no comparator or method. So the credible work and the marketed numbers are not connected.

A short chain because the model is built in house, with the parties in it described by category rather than named.

The structural fact matters most. The company trained its own audio and language foundation model on psychiatric data rather than calling a general purpose speech service, so the usual supply chain question, which outside model provider sees the patient's voice, has a different answer here: on the company's account, none. That is a chain that is structurally short rather than one that has been disclosed.

The privacy policy adds partial disclosure by category. Recipients of patient data are listed as the treating professional and their staff, specifically authorized Callyope personnel under confidentiality obligation, the strictly authorized administrators of the certified health data host, persons authorized by law, and partners where data is reused for research. Legal bases and retention periods are given per purpose.

Two qualifications matter. Not one party is named: no model, no hosting provider, no sub processor list, and the research partners appear only as partners. And the public commitment that data is never shared with third parties is narrower in the policy than on the marketing page, where research reuse rests on public interest rather than consent and lists partners among the recipients.

The training corpus is the remaining provenance question. A proprietary psychiatric speech corpus is the company's central asset and is uncharacterized: whose speech it contains, collected where, and on what consent basis is unstated. For recorded speech from psychiatric populations, that is the most pressing gap.

The company's own research on computing over encrypted audio is relevant here and is research rather than deployed practice.

Real published research answering a different question from the one the marketing asks a buyer to believe, and a clinical trials page that is still a template.

The research is named and attributable. Seven papers are listed with authors and years, spanning cross linguistic masked autoencoding for detecting Parkinson's disease from speech, learning speaker and health related representations from natural language supervision, robust fine tuning for disordered speech, in context learning to detect adolescent suicide risk from speech transcripts, and automated speech content analysis for depression with large language models. One is published in the Journal of Medical Internet Research and linked from the site, and parts of the work report on a public third party dataset alongside the company's own, so some of the performance record can be checked against a standard the company did not set.

What it does not do is answer the product claim. The marketed headline is that thirty seconds of speech assesses ten or more mental health symptoms at over ninety percent accuracy. No published evaluation is tied to that number, the production model is a different artifact from the research models, weights are not released and the proprietary psychiatric training corpus is uncharacterized. A separate claim that chart reviews are fifty percent more accurate than general purpose language models carries no method, comparator detail or denominator either. The published work establishes that speech signal carries symptom information; it does not establish that this product measures what it says it measures, and it does not establish that acting on a score changes what happens to a patient.

The clinical program is real in design, but its trials are not listed and no outcome results are published. Eight trials are claimed as ongoing with leading institutions, and a named research and development partnership with GHU Paris is announced. The clinical trials page the site directs a reader to is unfilled template content: eight rows reading Study title 8, Institution [Centre], Condition [Condition], Patients [N]. A registered trial of voice based biomarkers for predicting schizophrenia relapse enrolls an estimated 200 participants and completes in October 2026, so no completed outcome study exists yet.

Retention and training use are set out in detail, and deidentification is not addressed at all.

Retention is published as a schedule rather than a posture, which is rare. The privacy policy sets out a table by purpose: account and platform data held for the contractual relationship with five years of intermediate archiving, data used to improve the product held five years from collection, device manufacturer records held for the statutory period, and research reuse governed by the French data protection authority's MR-004 reference methodology at two years after the last publication of results. Each of those is a figure a customer could hold the company to.

Training use is stated plainly on the public pages: patient data is never used to train the models without explicit consent and never shared with third parties, and the policy carries that through with health data processed for product improvement on the consent basis under Article 9(2)(a).

Two qualifications belong beside that sentence rather than after it. Research reuse is a separate purpose resting on public interest under Article 44.3 of the French data protection act rather than on individual consent, so the never without consent commitment covers model improvement and not every secondary use. And the recipients listed for that reuse include partners, unnamed, alongside the company's own staff and the certified host's administrators, so never shared with third parties is narrower in the policy than on the marketing page.

Deidentification is the gap. No anonymization or pseudonymization practice is described anywhere. For voice that matters a great deal, because speech carries speaker identity independently of what is said and cannot be deidentified the way a record field can.

The company does conduct primary research on the problem, presenting work on quantized approximate signal processing as a route toward homomorphic encryption for audio, which addresses the risk at its root rather than wrapping it in policy. That is research, not a published practice, and it does not close the gap.

This is a French company operating under the General Data Protection Regulation. It states full compliance and the use of hosting certified under the French Hebergeur de Donnees de Sante regime for all health data, which is an audited health data specific certification rather than a self declaration.

The supporting document is substantive. The privacy policy names the controller for each purpose, distinguishing the treating professional or partner care structure as controller for care delivery from Callyope as controller for the account and the product, with Callyope acting as processor under Article 28 for records the clinician uploads. It sets out the legal basis and retention period for every purpose in a table, lists the categories of recipient, states the rights available and the route to the French supervisory authority, restricts the application to adults except in specific studies with parental consent, and publishes a data protection officer with both a postal and an email address.

The gap is the agreement itself. No data processing agreement terms are published, no processing register is available, and the flow down obligations on the certified host and on research partners are described by category rather than by term, so a care structure cannot read what it would be signing before contracting.

One distinction matters: the company states it uses certified hosting, which places that certification with the hosting provider. Using a certified host and holding certification yourself are different assurances, and the second covers the vendor's own handling of the data once it arrives.

An audited certification rather than an assertion, named without the scope that decides what it covers.

The company states that it is certified to ISO 27001, uses hosting certified under the French health data hosting regime for all health data, and encrypts data in transit and at rest. An ISO 27001 certification is an external audit of how the organization operates security, and it is stated as the company's own rather than attached to its hosting provider.

The certificate is named without a number, an issuing body, an issue or renewal date, or a declared scope. Scope is where ISO 27001 certifications differ most: a certificate covering a corporate function is not the same as one covering the production platform that processes patient speech, and nothing published says which this is. The health data hosting certification sits with the hosting provider, not with Callyope, and covers the environment rather than the company's own handling once data arrives.

No SOC 2 report, penetration testing statement, trust center or vulnerability disclosure policy is published, and no attestation document is reachable without asking.

The data at issue raises the stakes on scope. The product carries recorded speech from patients with schizophrenia, bipolar disorder and psychosis, captured on personal phones and transmitted to vendor infrastructure.

A device position exists under the regime that governs this company, and the class it declares raises a question.

The company states that its symptom assessment tool is classified as a Class I medical device and CE marked, describes itself in its own privacy policy as a medical device manufacturer with the record keeping obligations that carries, and is registered with the Nanterre commercial court under number 913 646 840. That is a stated regulatory position under the European medical devices regulation, which is the regime that applies to a French company selling in Europe. No United States market entry exists and no FDA involvement applies.

Three problems remain.

The class is self declared. Class I is the only class under the European regulation that a manufacturer certifies itself, with no notified body reviewing the file. So CE marked here means no external assessor has examined this product's clinical evaluation.

The classification invites a question. Software intended to provide information used to take decisions for diagnostic or therapeutic purposes is generally placed above Class I under the regulation's classification rule for software, and higher again where the decision could lead to serious deterioration of health. This product produces continuous psychiatric symptom scores intended to inform clinical decisions, including monitoring for relapse. No written classification rationale is published, and nothing explains why the software rule does not place it higher.

The device is named inconsistently. The clinical trials page identifies the Class I device as CallyoVoice, a voice symptom monitoring tool. The site footer and the privacy policy identify it as CallyoCare, a mental health symptom assessment tool. Published material does not say which product carries the marking or what its stated intended use is, and intended use is what a device certification actually covers.

The company has separately stated an intention to freeze a version of its model and submit it to health authorities for certification, which is the correct sequence.

A claim of equal performance across languages is published, and no per language result is.

The company's own answer to whether the model works in different languages reads: "Yes. Our model is trained on multiple languages and designed to generalize across linguistic contexts. This multilingual capability ensures consistent performance regardless of patient language, which is critical for diverse clinical populations." No numbers, method, evaluation set or per language breakdown is published to support the word consistent.

The company's own prospective work runs the other way. Its registered trial of voice based biomarkers for predicting schizophrenia relapse recruits French speakers only, and excludes anyone with a condition that impairs French. So the study designed to establish that the product works is monolingual, and excludes the speech impairments most likely to expose the failure mode.

The mechanism is specific rather than generic, and it is why this matters more here than for a transcription tool. Speech models carry documented accuracy variation by accent, dialect, sex, age and first language. In psychiatric populations the signal is further confounded: antipsychotic and sedative medication alters speech rate, prosody and articulation, which are the exact features the model reads, and so do the conditions being measured. A falling symptom score can reflect deteriorating mental state, a medication change, or an accent the model handles less well, and nothing published lets a clinician separate those. The output is a clinical symptom score rather than a transcript a human immediately checks, so there is no natural point at which the error surfaces.

The company publishes research adjacent to the question, including cross linguistic modeling for speech based disease detection, robust fine tuning for disordered speech, and multilingual depression detection. Those study research models on other conditions and other datasets. None reports performance for the deployed model, and the production model is a different artifact from the research models by the company's own account.

No bias testing methodology, model card, calibration data, drift monitoring or external audit of model behavior is published.

Clinician review stands between the score and any action, and nothing stands behind the score itself.

What exists is a mechanism rather than a commitment. The company states that the system supports clinicians rather than replacing them and that the clinician remains in full control of diagnosis, treatment planning and care, so a symptom score reaches a patient only through a person who can disagree with it. Data subject rights under the European regulation are published with a named data protection officer and a route to the French supervisory authority, which is a real channel for a patient who objects to processing, though it addresses the handling of data rather than the correctness of an output.

What is absent is anything answerable. No service level agreement, accuracy warranty, indemnity or remediation obligation is published, and no route is described for a patient harmed by a wrong score.

The marketed figure does not fill the gap. Assessing ten or more symptoms from thirty seconds of speech at over ninety percent accuracy is falsifiable in principle and has nothing behind it: no method, no denominator, no reference standard and no consequence if it fails. The company's published research is rigorous and reports against a public benchmark, but it evaluates research models on research questions, and the company itself does not connect that work to the production claim. Credible science elsewhere in the company is not a commitment about this product's output.

Integration is claimed and no system is named.

The company states that it integrates directly with major record systems to eliminate double data entry, that clinical information flows automatically, and that its outputs, summaries, reports and notes, can be pushed back into the record in the right format, with secure data exchange offered for systems without a direct integration. Not one system is named, no interoperability standard is cited, no partnership with a hospital information system supplier is announced, and nothing is published that a buyer could verify.

The product does handle clinical data in both directions. It ingests uploaded medical records for chart review and case summarization, takes smartphone sensor signals covering sleep and activity, and emits structured documents including consultation notes, referral letters and hospitalization reports, the last of these marked as coming soon. Delivery is a patient smartphone application plus a clinician web interface.

The unresolved question is whether the automatic flow described in the answer above exists for any named system today, or whether documents are in practice produced for a clinician to place into the record by hand. That manual step is where adoption of this kind of product usually fails.

The integration burden in European hospital systems differs from the American one, so an absence of named connections here is less telling than it would be for a United States vendor. It is still an absence, and the claim is now specific enough to be tested.

A clear and well governed answer to where the data rests, with the path it takes to get there left open.

Health data is processed vendor side and hosted under the French health data hosting certification regime, which fixes it inside a European jurisdiction and inside an audited hosting framework rather than leaving residency to a contract term. The privacy policy names the certified host's administrators as a distinct category of recipient with access limited to their functions. That is the strongest posture available to a vendor that must centralize, and centralizing is unavoidable here because the model consumes speech captured on the patient's own phone.

The data does move, and the movement is undescribed. Patients transmit voice recordings from personal devices to vendor infrastructure, which is a materially different exposure from software that runs inside a hospital and never sends anything out. No hosting provider is named, no region within France or the European Union is specified, no tenancy or isolation model is described, and no architecture is published. Nothing states what the mobile application holds on the handset before it uploads, whether it is encrypted at rest there, or what happens to a queued recording if the device is lost.

No customer managed key option, single tenant option or on premises path is offered or discussed.

Nothing is published. No price, no pricing mechanism, no contracting model and no stated basis of charge appears anywhere, and every commercial path terminates at a fifteen minute demo booking or a contact form. Site navigation carries no pricing entry.

No reimbursement pathway is identified in any market, and for this product the gap carries weight. Its value proposition is continuous symptom monitoring between appointments, a service most health systems have no existing code to pay for, so a buyer cannot tell whether the intended payer is the care structure, the hospital, a research budget or an insurer. The product is delivered through partner care structures under agreement, which implies institutional contracting, and nothing describes how that contract is priced: per clinician, per monitored patient, per site or as a research collaboration.

The visible commercial facts are funding rather than pricing: a seed round of roughly 2.2 million euros co led by 360 Capital and Bpifrance Digital Venture with No Label Ventures participating, plus selection for Google's artificial intelligence for health program and an Amazon Web Services cohort. That is a small round against a product carrying both a research program and eight claimed clinical trials, which is worth a buyer's attention on supplier continuity grounds for a monitoring service a patient would depend on continuously.

Deliberately narrow on specialty, and the narrowness is stated clearly with nothing validating it yet.

Specialty is psychiatry. The company names depression, anxiety, insomnia, fatigue, cognitive decline, bipolar disorder, schizophrenia and psychosis, and frames its own scope as the psychiatric, cognitive and motor symptoms of brain disorders more broadly. The approach is described as transdiagnostic by design, one model across the spectrum rather than one per condition, on the argument that comorbidity is the clinical norm.

Setting is outpatient psychiatric care between appointments, plus hospital use through a named research and development partnership with GHU Paris. The problem targeted is real and specific: a patient with schizophrenia may see a psychiatrist only every four to six weeks, and deterioration in that interval goes unobserved.

What is missing is validation behind any of it. Eight clinical trials are claimed as ongoing, the page listing them is unfilled template content, and the one registered trial that can be identified is a 200 participant schizophrenia relapse study completing in October 2026. So the coverage claim is a statement of intent supported by studies in progress rather than by results.

Language constrains this before anything else does, and the record cuts two ways. The company states the model is trained on multiple languages and performs consistently regardless of patient language; its registered trial recruits French speakers only and excludes anyone whose condition impairs French. For a speech model, coverage means the languages and accents it was tested in, so expansion is a revalidation problem per language rather than a translation problem, and nothing published says which languages have been tested.

No pediatric position is published; the application is stated to be for adults except within specific studies carrying parental consent.