Speech, measured the same way twice.

Research

An instrument is only worth what its numbers can be trusted to say. Ours have not been checked yet — and that is what research cooperation is for.

Speech is measured three ways today. A clinician scores the session live, which is reliable but needs an examiner and a schedule. A recording is scored by hand, which is what most laboratories do — slow, and the score depends on who did it. Or software scores a tightly scripted task and returns a number without the record behind it.

SonaXR is designed to run the session itself: the same voice, the same prompts, the same pacing, for every participant. Scoring is deterministic — compared against a fixed reference, or against a key a linguist set in advance. Nothing is scored by hand. And the record comes out: every item, every response, every timing, in open formats, so that a run six months later can be compared with this one.

Fields where researchers already measure speech

These are research directions, not products. None is a validated protocol; each would be a study designed with a laboratory, under its ethics approval, with content authored to its own constructs.

Research use only. SonaXR is a research instrument in development. It is not intended to diagnose, treat, cure, monitor or prevent any disease or condition, and it has not been cleared or approved by the US Food and Drug Administration or any other regulator.

Any use of participant data would be governed by a written agreement with the institution, the approval of its ethics or review board, and the participants' consent, and would be limited to what those permit.

The scientific basis for each design decision is set out in the Foundations brief — one peer-reviewed construct per implementation choice, including where the evidence is mixed.

Language readiness for defense and government programs

Proficiency is perishable between formal evaluation cycles, and the measurement in between is usually nothing at all.

This is the same instrument, pointed at a language learner rather than a research participant. The session is designed to be administered identically every time, without an examiner and without a schedule, so that a reading taken in March can be compared with a reading taken in September. It is intended as measurement between formal evaluations — never as a substitute for one.

Format

A short, seated session on a commercial standalone headset, without an examiner.

Record

Every item, every response, every timing, in open formats — the same record as the research work.

Reference scale

Results are described against the published Interagency Language Roundtable skill-level descriptions as a reference only. SonaXR does not produce an official rating, and no correspondence to one has been established.

Deployment

Use on government networks, devices or facilities depends on each organization's own security, privacy and device approvals, which we would work through with the program.

We would welcome a conversation with language programs, training commands and their research partners about what between-evaluation measurement should look like and how it should be tested.

Where the instrument could go next, once a partner defines it

The fields above exist because researchers described them. Another happens the same way: a partner defines the construct, and the content is authored to it.

Where existing approaches stop.

Spoken proficiency has been measured by machine for decades. Knowing precisely where each established approach stops is the basis of what we are building.

	A human rates it	A machine scores it
A person delivers it	The certified interviewThe gold standard. Twenty to thirty minutes, one trained tester, functional tasks, a rating on a recognized scale.	—
A computer delivers it	Avatar-delivered interviewPrompts on demand, in many languages. Keeps the functional tasks — and pays for it, because recorded sessions still go to certified human raters.	Automated scripted-task scoringNo human rater, results in minutes, in use for decades. Works by keeping tasks tightly controlled — which is also why it measures something other than functional language use.
A computer delivers it inside a situation	—	What SonaXR is buildingFunctional speech in a situated scenario, scored without a rater, on a record designed to carry no name or participant identifier.

The gap, in one sentence. Automated scoring is usually bought by giving up functional tasks. Functional tasks are usually kept by paying humans to rate them.

Why a headset, when existing tests need only a phone

Not immersion for its own sake. Repeating a sentence into a handset measures something real, but it is not the same construct as handling a situation in the language. Functional tasks need a situation to be functional in, and virtual reality is how that situation is constructed identically for every participant, every session, at every site. The headset holds the conditions constant. It is a means, not the product.

What makes a task functional rather than scripted

A scripted task fixes the correct answer in advance — repeat this sentence, read this passage. A functional task fixes the goal and leaves the wording open: describe what happened, narrate this sequence, handle this situation. Those are the tasks recognized proficiency scales are written around, and they are hard to score automatically for exactly the reason they are worth scoring. SonaXR narrows that space without scripting it — the situation constrains what a reasonable answer looks like, and scoring runs against a frozen reference and answer sets a linguist approved in advance, rather than against free transcription of accented speech.

The instrument, in parts

Status: pre-seed prototype. What follows is running code demonstrated on a commercial headset, not a shipped product.

Delivery

Virtual reality on a commercial standalone headset. Meta Quest 3 is the current development target; other headsets have not been tested.

Speech

Phoneme-level pronunciation scoring through a commercial cloud speech service (currently Microsoft Azure). Spoken audio is sent to that service for scoring. Adding a language is mainly content work.

Scoring

Deterministic — compared against a fixed reference, or against a key a linguist set in advance. Nothing scored by hand.

Coaching

A test that helps you is not a measurement. Measurement sessions are designed to run without coaching or feedback.

Record

Every item, every response, every timing, in open formats, exportable into an organization's own analysis tools.

Languages

The pipeline is designed to be language-agnostic. A demonstrated English scenario runs on the same architecture that scores European Portuguese.

Designed with review boards in mind

This is the first question a review board asks, so it is answered here rather than in an appendix.

• Separated identity. Analytics records are keyed to a session, not to a person, and are designed to carry no name, account or participant identifier.
• Audio. Spoken audio goes to the speech-scoring service described above, under that service's data-processing terms. SonaXR does not write raw audio to its analytics database.
• Deletion. Deleting study data is designed to require two independent authorizations.
• Residency. The analytics database is hosted in the United States. Institutions with other residency requirements should raise them early.

Two of the provisional patent applications cover this architecture. A detailed data description for ethics and review-board submissions is available on request. See also our privacy notice.

Two people

One engineer, one operator.