Canary Speech

Canary Speech sells the model itself through an API, and there is no application wrapped around it to carry independent value. No portal, no scheduling, no documentation suite, no patient experience layer. A customer buys inference and builds the product around it.

The analytical work is substantial rather than nominal. The engine extracts roughly 2,590 acoustic and linguistic features every 10 milliseconds and returns behavioral and cognitive indicators from 20 to 45 seconds of natural speech, using a custom predictive ensemble trained on acoustic, prosodic and linguistic biomarkers. Reading how something is said rather than what is said is the entire basis of the offering, which is why the company can claim device independence at all.

The API first architecture leaves no doubt about how central the AI is. Where a platform vendor's intelligence has to be separated from its workflow to be seen at all, this vendor has already performed that separation commercially: the intelligence is the unit of sale, priced and documented as such, with audio capture requirements published down to sampling rate and bit depth because the interface is the product.

The corresponding exposure is that every risk in this product concentrates in the models, since nothing else exists to absorb it. A model that misses a case has no workflow layer, no second reviewer and no adjacent product feature standing behind it, which is why the published operating point and the abstention class matter more here than they would elsewhere.

What the system may do and may not do are both published, with a threshold, an override and a route to a person, and one deployment where none of that applies.

The clinical position is documented in a published model card rather than asserted in marketing. Intended use is stated: to inform and augment, explicitly not replace, clinical management of adults at risk of or being treated for depression and anxiety, in ambulatory, telehealth, acute or remote monitoring settings. Human review and clinical correlation are stated as requirements, clinicians maintain oversight, and the card says plainly that clinicians remain responsible for diagnosis per the standard of care. Primary users are named down to the competencies assumed. Out of scope settings and warnings are published as their own sections. Operating conditions are specified to the level of forty seconds of speech, microphone distance and ambient noise, and informed consent for voice recording is written into the procedure.

The threshold is published twice over. Sensitivity and specificity are given at the deployed decision point, so an integrator knows the approximate miss rate before deployment rather than after an incident. And the system can decline to answer: an uncertain class flags low confidence predictions near the decision boundary instead of forcing a binary call, routing ambiguous cases to human judgment rather than manufacturing a confident answer. Abstention carries a commercial cost, since it reduces the proportion of inputs the product appears to resolve, and the company built it in anyway.

Scope language is disciplined elsewhere too: the clinical offering is described as decision support, and consumer outputs are called nondiagnostic indicators rather than assessments.

The open item is the consumer deployment, and it is a real one. Passive analysis embedded in video calls between older adults and their families places a behavioral indicator in front of people with no professional standing, no clinical correlation step and no published routing for a concerning result. Because the model ships inside someone else's application, the vendor also cannot fully control how its output is presented.

Two versioned model cards are published in a public third party registry, and they say what produces the output, what it was trained on and how it is updated.

What produces the output is identified: a custom predictive ensemble model designed to analyze acoustic, prosodic and linguistic vocal biomarkers, extracting roughly 2,590 features every 10 milliseconds, with a patent covering neural network based speech analysis granted in 2024. The cloud environment is named. An explicitly described uncertain class flags predictions near the decision boundary rather than forcing a call.

Training data is characterized rather than gestured at. Input is real world audio from clinical and research participants, typically conversational English, Spanish and Japanese, at 40 seconds or more. Development data is drawn from multi institutional studies approved by institutional review boards, described as encompassing diverse adult populations balanced for age, gender and clinical status where possible, collected from clinics, telehealth studies and academic settings across all fifty states.

Versioning and update practice are published too. The behavioral health card carries a version number and a release date, with release stage given as production. The update process is described: models are released as expanded datasets become available or in response to substantial shifts in data patterns, with comprehensive quality review before production rollout, validity monitored at minimum monthly, and corrections implemented usually within weeks of risk detection.

The operating point sits alongside all of it. Sensitivity of 0.76 and specificity of 0.65 on a remote test set, and sensitivity of 0.68 with specificity of 0.80 on an independent in clinic dataset, published in a peer reviewed paper. Those figures are unflattering and were published anyway, which is what makes them useful: they show how many cases are missed and how many people are flagged unnecessarily.

The remaining gap is scope. The published figures and the model cards cover behavioral and cognitive health; the neurological claims carry no numbers.

The model layer is proprietary and partly documented, the cloud is named, and the rest of the chain is not.

The models are built in house rather than assembled: a patent covering neural network based speech analysis was granted in 2024, the published model card describes a custom predictive ensemble designed for acoustic, prosodic and linguistic biomarkers, and one co founder is a speech recognition specialist by background, which makes an internally built speech pipeline plausible. The cloud environment is named in the model card, which the security page had only characterized by the assurance programs its partners hold.

Training provenance is described rather than labeled. Development data is drawn from multi institutional studies approved by institutional review boards, from clinics, telehealth studies and academic settings across all fifty states, with input audio typically English, Spanish and Japanese. Naming the consent basis answers where the training corpus came from.

What is still unnamed is everything between. No sub processor register is published, no third party speech recognition or transcription component is confirmed or excluded, and no third party library position is stated. For a product whose input is recorded human voice, whether any external service ever receives customer audio is the first question a privacy reviewer asks, and it is unanswered.

The hosting provider matters less than the model layer, and there the models are the company's own, which makes the chain structurally short without making it an enumerated one.

Whether audio captured through customer or partner deployments feeds further development remains unstated. The consumer partnership makes that more pressing, since conversations between older adults and their families are an unusually rich source of natural speech.

Honest published performance, vendor authored, with commercial outcome evidence still missing.

The 2026 paper in the Proceedings of Artificial Intelligence in Medicine reports unweighted average recall of 0.70 for depression and 0.68 for anxiety, and a combined assessment reaching sensitivity of 0.76 with specificity of 0.65 on a remote test set. Performance was then examined on an independent in clinic dataset collected by tablet, where sensitivity was 0.68 and specificity 0.80. Testing on a second dataset gathered by a different method in a different setting is a meaningful robustness check rather than a restatement of training performance.

Those are moderate numbers. A specificity of 0.65 means roughly a third of unaffected people are flagged, which matters for a screening tool deployed at volume. Publishing them anyway, at the deployed operating point, rather than asserting equivalence without figures, is disclosure that costs something.

The published model cards add provenance the paper alone does not carry: development data drawn from multi institutional studies approved by institutional review boards, described as balanced for age, gender and clinical status where possible, collected from clinics, telehealth studies and academic settings across all fifty states, with input audio typically English, Spanish and Japanese.

Authorship is the qualification. All five paper authors are company staff, including the chief executive. Peer review is real and vendor authored research is normal in this field, and no independent replication exists.

Two further gaps. Deployment evidence is absent: no customer count, no facility count and no outcome from any named site, so nothing shows that using the technology changes care. And the validated scope is depression and anxiety. Dementia, Alzheimer's disease and multiple sclerosis are all marketed or under study, the last an institutional review board approved study with a health system launched in February 2026 rather than a completed result.

One relationship to hold separately: a named investor is also a health system, so any reference originating there carries a commercial interest.

Real controls described at the right level of specificity, with one category of risk left entirely unaddressed.

The substance is there. Anonymous configurations are offered, API research data is de identified inside the analysis environment, the development lifecycle and continuous infrastructure monitoring are described, and concrete assurance activities are named: control audit against a recognized benchmark, vulnerability scanning, cleared API penetration testing, annual third party audits, TLS 1.2 in transit and AES-256 at rest. Naming penetration testing of the API specifically is apt, since the API is the entire attack surface. A confidential channel is published for anyone to report concerns about the company's handling of data.

Consent is now documented for the clinical path. The published model card writes it into the procedure: the clinician obtains informed consent from the patient regarding the recording of their voice before conducting an assessment. That is a described step rather than a policy sentence.

The unaddressed category is biometric law. A voiceprint is biometric data. Several jurisdictions regulate the collection and retention of biometric identifiers under statutes distinct from health privacy law, with their own consent, notice and retention requirements and, in some cases, a private right of action. A vendor whose product extracts thousands of features from a person's voice sits squarely inside that regime, and nothing published engages it in any form. This is the most consequential legal gap in the record and it is specific to this vendor rather than generic.

The consumer product sharpens it. Passive analysis of an older adult during a family video call raises the question of whose consent was obtained and whether the person being analyzed knows, and the consent step documented for the clinical path has no published counterpart there.

Retention is unstated for audio and for derived features. The training position is partly answered, since development data is described as coming from institutional review board approved studies, and nothing states whether audio captured through customer or partner deployments feeds further development.

A compliance claim stated plainly, architecture that reduces the exposure, and no contracting position behind either.

The architecture matters because it is a design decision rather than a policy sentence. The company states that fully anonymous solutions are offered and that research data collected through the API is de identified in the analysis environment. For an API embedded in someone else's product, that arrangement can mean the vendor never holds identity at all, which lowers the stakes of the compliance question rather than answering it more loudly.

What is asserted is HIPAA compliance, described as full and listed on the published model card alongside three certifications. HIPAA has no certifying body, so this remains the company's own assessment of itself, and the certifications that do carry external assessment are security and artificial intelligence management standards rather than a health privacy attestation.

Business associate agreement handling is not addressed. No template, no execution requirement, no statement of which entity contracts, and no description of how the agreement chain works when the API sits inside a partner's product and the patient relationship belongs to that partner.

That last point is the distinctive gap here. In an embedded model the flow of responsibility runs vendor to integrator to provider, and nothing published describes who signs what at each hop, or whether Canary contracts directly with covered entities at all. The consumer deployment sits outside the chain entirely, since a family video calling product is not a covered entity and the person being analyzed is not its patient.

The company now holds certifications of its own, named with their versions, and the artifacts behind them are not published.

What Canary holds, stated on its published model cards: HITRUST e1 certification, confirming foundational cybersecurity controls for healthcare, and ISO/IEC 27001:2022 and ISO/IEC 42001:2023 certifications, the second being the management system standard for artificial intelligence. Naming the revision year of each standard dates the claim precisely. A governance, risk and compliance program is described with enterprise wide oversight, periodic risk assessments, intrusion detection, server log review and continuous monitoring. Canary states its own activities specifically rather than by badge: control audit against a recognized benchmark, vulnerability scanning, cleared API penetration testing, annual third party audits, TLS 1.2 in transit and AES-256 at rest. Penetration testing of the API matters most, since the API is the entire product surface.

The security page enumerates assurance programs spanning SOC 1 and ISAE 3402, SOC 2, SOC 3, FISMA, DIACAP, FedRAMP, and ISO 9001, 27001, 27017 and 27018, introduced by a sentence stating these are programs the company's cloud partners adhere to. That attribution is accurate and correctly worded. The effect is not: anyone skimming a security page sees ten certifications and forms an impression, and inherited infrastructure assurance is not assurance covering the vendor's own controls, staff, development practice and handling of customer data. The company's own three certifications are real and are listed elsewhere, not on that page.

The weakness is retrievability. No certificate, report or audit period is published, no assessor is named, no scope statement appears, and there is no trust center and no vulnerability disclosure policy, though a confidential reporting channel for concerns about data handling and artificial intelligence impact is published.

Careful scope language, a published intended use statement, and no stated device determination behind either.

The care is evident and documented rather than implied. The published model card sets out an intended use: to inform and augment, explicitly not replace, clinical management of adults at risk of or being treated for depression and anxiety, operating as a clinical decision support tool, with clinicians remaining responsible for diagnosis per the standard of care. Out of scope settings and warnings are published as their own sections, the target population is bounded at 18 and over, and consumer outputs are described as nondiagnostic indicators. No registration is presented as a credential, no designation is described as a clearance, and no approval language appears.

One marketing sentence pulls against all of that. The home page states that the ambient product objectively detects cognitive and behavioral conditions ahead of traditional clinical screening and before observable symptoms. Detecting conditions before symptoms are observable is a stronger claim than decision support, and the home page carries both. That is tension rather than contradiction, because a position is published and published clearly, but it is the sentence a regulator or counsel would read first.

The determination itself is missing. No device classification, no rationale, no clearance, and no published statement that the products sit outside device regulation. The model card lists regulatory approval as HIPAA compliance and three certifications, none of which is a device status.

The question is live at the edges of the marketed range rather than at its center. Screening for depression and anxiety with a clinician deciding is comfortably within decision support. Identifying dementia or Alzheimer's disease from voice moves closer to detecting disease, and the multiple sclerosis study, approved by an institutional review board and aimed at identifying individuals with a neurological condition from vocal features, is investigating a capability that would be difficult to characterize as anything but diagnostic if it succeeded and shipped. The consumer deployment tests the wellness boundary, since the same underlying models carry clinical claims elsewhere.

An externally certified artificial intelligence management system with a named fairness process, and no fairness result published.

The process is real and audited. The company holds ISO/IEC 42001:2023 certification, the management system standard for artificial intelligence, and its published model cards describe what that system does. Fairness is assessed through structured audits using named metrics, demographic parity and equal opportunity, applied to general and locally sourced data, conducted biannually and at every major retraining cycle, with ad hoc reviews triggered by reported incidents. Validity is monitored at minimum monthly. Identified disparities trigger detailed analysis and prioritized remediation, with corrections usually implemented within weeks. Bias mitigation is described concretely: collection across a spread of locations, ages and genders, and acoustic data augmentation simulating changes in voice type, microphone type and acoustic environment. A confidential reporting channel is published for anyone to raise concerns about the impact of the company's artificial intelligence systems.

That is a governance framework with named process behind it, considerably more than a principles page.

What is absent is the output. No subgroup performance is published by sex, age, race, accent or primary language, no calibration analysis appears, and no result of any biannual fairness audit is disclosed. A named metric applied on a stated cadence with no figure released shows that the work happens and nothing about what it found.

The risk remains specific to this modality and well documented. Speech models carry differential error by sex, age, accent and dialect. The language claim is where it bites hardest: training input is described as typically English, Spanish and Japanese, while the approach is marketed as language agnostic on the basis that it reads acoustic form rather than lexical content. That is a reasonable expectation and it is not a result, and availability is stated as the United States with select international markets for pilot. The consumer deployment adds age, since older adults' speech is altered by hearing loss, dentition, medication and comorbidity, all plausible confounds for the features being read.

A published error rate at the deployed operating point, and a reporting channel the affected person can use, against no contractual allocation at all.

The error rate is the substance. A peer reviewed paper reports sensitivity of 0.76 and specificity of 0.65 on a remote test set, and sensitivity of 0.68 with specificity of 0.80 on an independent in clinic dataset collected by tablet, with unweighted average recall of 0.70 for depression and 0.68 for anxiety. Those are figures at the threshold the product actually deploys at, with the method and the datasets stated, and the company cites them on its own estate. A sensitivity of 0.76 means roughly one case in four is missed, and the company put that number in the literature rather than leaving a buyer to discover it after an incident. A miss rate that is known can be planned around, staffed for, insured against and written into a contract. A miss rate that is unpublished cannot.

Two mechanisms sit alongside it. The uncertain class flags low confidence predictions near the decision boundary instead of forcing a binary call, which limits confidently wrong outputs at exactly the point where they are most likely. And a confidential, anonymous channel is published for anyone to report concerns about the company's handling of data or the impact of its artificial intelligence systems, reviewed by a compliance team. That is a route an affected person can exercise against the vendor directly rather than through their provider.

What remains absent is the contract. No service level agreement, accuracy warranty, performance guarantee, indemnity or remediation commitment is published, and terms governing the API were not reachable without contact.

The embedded model complicates allocation beyond the usual. When the analysis ships inside a partner's product, responsibility for a missed indication divides between a model with a published error rate, an integrator who chose how to present it, and a clinician or family member who acted or did not. The consumer deployment puts a behavioral signal in front of family members with no professional standing, and nothing describes how that division is handled contractually.

Three major record systems named, with the mechanism behind the compatibility unstated.

The published model card lists Epic, Cerner and Oracle Health, and Meditech as compatible record systems, which covers the great majority of the American acute market. That is a named integration position and it is published in a third party registry rather than only in marketing copy.

The architecture supports it. This product is built to be embedded: an API first design with documented integration capability and public developer documentation, audio capture requirements specified down to sampling rate, bit depth and channel count so an integrator knows what the interface needs. One named embedding runs in production outside the clinic, a consumer caregiving partnership placing the analysis inside another company's video calling product. Telehealth organizations, health systems and clinical call centers are described as an existing partnership network. A vendor whose commercial model depends on other people integrating it has stronger incentives here than one selling a portal.

What is not published is how the compatibility works. No interface standard such as FHIR or HL7 version 2 is named, no direction of flow is stated, no marketplace or partner listing appears for any of the three systems, and no named site is identified as running one in production. Compatible is a claim about what is possible rather than evidence of what is live.

The consequential question for a screening product is where the result lands. A behavioral or cognitive indicator reaching a clinician inside the chart as a discrete, trendable result is a different clinical object from one displayed in a partner application, and nothing describes which occurs. For a product whose value rests on detecting change earlier than scheduled screening, longitudinal storage in the record is close to essential.

The hosting provider is named, and everything about where it runs is not.

The published model card names the cloud environment the configuration files sit in, which resolves a question the security page had only gestured at by listing the assurance programs its cloud partners hold. That list, spanning the SOC family, FedRAMP, FISMA and several ISO standards, shows that the underlying infrastructure sits at the tier where government and regulated workloads run.

Everything specific below that is missing. No region is stated, no residency commitment is made, no tenancy model is described, no segregation approach is given for an API serving many integrators, and no customer controlled or private deployment option is mentioned.

Residency matters more here than the omission alone suggests, and the model card sharpens it rather than settling it. Global availability is stated as the United States with select international markets open for pilot and integration, while training input covers English, Spanish and Japanese and an investor base includes a Japanese corporate. European and other regimes impose location and transfer requirements on health and biometric data that a voice analysis service operating internationally cannot avoid, and nothing published addresses where audio is processed or where derived features are stored for a non United States deployment.

A related question follows from the architecture, and it comes first for an integrator. Whether audio leaves the integrator's environment at all, or whether feature extraction can occur locally with only derived values transmitted, materially changes the residency analysis. The published audio capture requirements describe what the interface needs and not where the analysis happens.

Cost is absent from every published surface. No pricing page, unit of charge, range, tiering, implementation fee position, minimum commitment or trial terms appears anywhere, and every route terminates in a demo request or a contact form.

The one commercial adjective published is that the API is cost effective, offered without a figure, a comparison or a basis. That is a claim a buyer cannot act on.

An API first product makes the omission more surprising rather than less. Usage priced infrastructure is the category most able to publish a rate, because the unit is obvious and inspectable: a call, a minute of audio, a completed assessment. The company publishes developer documentation, audio capture specifications down to sampling rate and bit depth, and a public model card with a version number, and does not publish what any of it costs. An integration partner has to model cost of goods before committing engineering effort to embed a third party model, and nothing here lets them.

Segment spread compounds it. Healthcare providers, payers, clinical call centers, contact centers, clinical trials and a consumer caregiving embed cannot plausibly share one rate, and nothing describes how they differ. A pharmaceutical sponsor running a trial and a family video calling product are not the same commercial conversation.

Broad reach for a company of this size, with a validated core considerably narrower than the marketed range.

The breadth is real and structurally enabled. Because the product is an API rather than an application, it reaches settings an application could not. The company now sells by segment: healthcare providers, payers, clinical call centers, contact centers and clinical trials, with condition lines for behavioral, cognitive, neurological, safety and wellness. February 2026 added consumer caregiving, embedding passive analysis into video calls between older adults and their families, its first deployment outside clinical and research environments. Covering the clinic, the trial, the call center and the living room from one model is unusual.

Condition coverage is where the claim outruns the evidence. Depression and anxiety carry published performance and a published model card, with a second card for cognitive health. Dementia and Alzheimer's disease are longstanding marketing claims with no published figures. Multiple sclerosis is an active study rather than a product. The segment list reads as if all of them are equally supported, and they are not.

Language coverage is partly answered, and the answer narrows the claim. Input audio is described as typically conversational English, Spanish and Japanese, and global availability is stated as the United States, with select international markets open for pilot and integration. That is a much more bounded position than language agnostic, which remains the marketing framing, and no per language performance is published for any of the three.

Age coverage is bounded at one end and open at the other. The behavioral model card states a target population of 18 and over. The consumer product targets older adults, where speech is affected by hearing loss, dentition, medication and comorbidity, and nothing separates those effects from the signal being measured.