Say Less: A Child-Speech Screener Designed to Stop Before Diagnosis

TL;DR for operators

The paper presents a narrowly scoped screening pipeline for detecting common Polish sibilant substitutions in children.¹ It does not attempt to transcribe unrestricted speech, diagnose a speech sound disorder, or replace a speech-language pathologist. That restraint is not an apology. It is the architecture.

The system works by preserving non-canonical pronunciations as explicit tokens, aligning them against a known prompted pronunciation, converting the difference into a structured error description, and allowing a caregiver-facing assistant to speak only through clinician-reviewable templates.

On 559 utterances from 10 children excluded from training, the recognizer produces the exact reference token sequence 88.7% of the time. Its stricter screening rule detects annotated target mismatches with 72.9% precision and 61.4% recall, while falsely flagging 2.7% of target-correct examples.

Operationally, that is a conservative triage profile: relatively few false alarms, but approximately two in five annotated mismatches go unflagged. The system may therefore be useful for prompting repetitions, identifying recurring contrasts, and prioritizing professional review. It is not evidence that the model can clear a child as having no speech difficulty.

The business opportunity lies in a controlled screening workflow, not a general-purpose diagnostic chatbot. The main uncertainties are equally concrete: the evaluation uses a fixed prompt inventory, quiet institutional recording conditions, only 10 held-out children, uncalibrated confidence scores, and no caregiver or clinical-outcome study.

Every screening system must decide what not to erase

Suppose a child is asked to repeat a word containing a difficult speech sound.

A conventional speech-recognition system has a familiar objective: recover the word the speaker probably intended. If the acoustic evidence is unusual but the surrounding pattern strongly suggests a common word, normalization is helpful. The customer wanted a transcript, not a phonetics seminar.

Mispronunciation screening has the opposite requirement.

Here, the unusual acoustic realization is the information. A system that tidies it into the expected word may achieve respectable transcription performance while quietly destroying the evidence needed for screening. It becomes an extremely competent machine for confirming that the child said what the test expected them to say.

The paper’s pipeline is designed around this inversion. Its key contribution is not simply a better child-speech recognizer. It is a sequence of controls intended to prevent clinically relevant variation from disappearing between the microphone and the caregiver message:

1. Represent important substitutions explicitly.
2. Recognize production tokens without an external language model.
3. Align the production against a known target.
4. Translate the difference into a constrained screening category.
5. Suppress detailed feedback when evidence is weak.
6. Escalate persistent patterns rather than diagnosing them.

Each stage limits a different failure mode. Together, they form something closer to an evidence-preservation workflow than an ordinary speech application.

Step one: Make the mistake part of the vocabulary

The system begins with a wav2vec2 speech encoder initialized from a Polish checkpoint. It adds a six-layer Transformer post-encoder and a CTC output layer that predicts short sequences of IPA-like tokens.

The unusual design choice is the token inventory.

Alongside canonical speech tokens, the researchers introduce bracketed tokens representing expert-annotated substitution evidence. A bracketed output does not merely say that the model heard a different canonical phoneme. It says that an intended target was judged to sound closest to a particular substitute.

That distinction matters because many speech deviations are not clean jumps from one dictionary phoneme to another. They may occupy acoustically ambiguous territory between the expected realization and a recognizable substitute. The bracketed labels encode the annotators’ interpretation of that evidence.

This gives the model three broad possibilities at a target location:

• emit the expected canonical token;
• emit an explicit bracketed substitution marker;
• emit some other non-canonical result whose meaning is less certain.

Only the second category automatically triggers the paper’s mismatch-screening rule. Other discrepancies can lead to a request for another recording or professional review rather than a confident explanation.

The conservatism is deliberate. The authors are not treating every edit-distance discrepancy as a childhood speech disorder. An acoustic model making an odd prediction and a child producing a clinically relevant pattern are not the same event, however much a product dashboard might enjoy merging them.

Why ordinary ASR is poorly matched to this task

General ASR systems benefit from linguistic context. They are rewarded for mapping uncertain acoustics to plausible words. In open transcription, that is a feature.

The paper instead uses greedy CTC decoding without an external language model. The prompts are already known, so there is little business value in having the system guess which word was intended. The valuable question is whether the production at a particular location agrees with the expected speech target.

This reframes the model from:

What word did the child probably mean?

to:

What token-level evidence did the child produce at the location we intentionally tested?

That narrower question makes the system less impressive in a product demonstration and more appropriate for the actual screening problem. A recurring theme in applied AI, apparently still requiring repetition, is that usefulness often begins when the model is prevented from being clever in the wrong direction.

Step two: Align the evidence to the prompted target

Once the recognizer emits a production sequence, the system aligns it with the prompt’s canonical token sequence using minimum-edit alignment.

The alignment identifies matches, substitutions, insertions, and deletions. For substitutions involving the 12 Polish sibilant and affricate targets covered by the system, a deterministic mapping assigns an interpretable error type. These include:

• place-of-articulation shifts within the sibilant space;
• voicing alternations;
• manner mismatches, such as frication versus affrication.

Other mismatches receive a generic label rather than a fine-grained explanation. This is another scope boundary embedded in the mechanism: the system describes distinctions for which it has an explicit linguistic mapping and becomes less specific outside them.

The output is a compact screening vector containing:

prompt → target → realized token → error type → position → confidence

This structure is more operationally important than it may first appear.

A bare model score is difficult to audit and difficult to translate into a safe caregiver interaction. A structured vector can be logged, reviewed, aggregated across prompts, compared against deterministic rules, and routed differently depending on confidence or error type.

It also separates the recognition system from the communication layer. The model produces evidence. A downstream policy decides whether that evidence is sufficient to support a message.

That separation is a useful design pattern beyond speech screening. In sensitive workflows, allowing the generative interface to interpret raw model output freely is not explainability. It is delegation with good branding.

Step three: Let the assistant explain only what the evidence permits

The caregiver assistant does not receive unrestricted authority to reason about the child’s condition. It fills a fixed inventory of clinician-authored templates using the structured screening vector.

A caregiver-facing response can include:

• the prompted word;
• a match or likely-mismatch indicator;
• the approximate location of the contrast;
• a plain-language description of the sounds involved;
• a brief practice cue;
• a suggestion to consult a speech-language professional if the pattern persists.

The underlying IPA sequence, alignment, error type, and confidence remain available for professional audit but are hidden from the default caregiver view.

The assistant also has explicit refusal and suppression rules. When confidence is low or the evidence is inconsistent, it requests another recording instead of generating detailed advice. It is prohibited from diagnosing, making medical claims, guaranteeing improvement, or converting one uncertain production into a broad developmental conclusion.

This is not a secondary interface feature. It is the final control in the mechanism.

Without it, an accurate token recognizer could still become an unsafe product. A language model could turn a localized acoustic discrepancy into an authoritative-sounding clinical narrative. Template grounding limits that transformation.

The resulting assistant is intentionally less expressive than a general chatbot. It can localize a contrast, suggest a repetition, and recommend escalation. It cannot improvise a diagnosis because it had a particularly inspired afternoon.

Recognition performance establishes the alignment backbone

The held-out test set contains 559 utterances from 10 children who were not included in training or validation. The prompts themselves are shared across the splits, reflecting the intended fixed-screening protocol.

The reference sequences are short: an average of 3.66 tokens, with a median of four. Exact sequence match is therefore a strict utterance-level measure. A single wrong token makes the entire sequence non-exact.

The exact-match result is useful because the downstream screening logic depends on locating the target token correctly. If the recognizer frequently inserted, deleted, or displaced tokens, a reported error position could reflect alignment instability rather than the child’s production.

Performance nevertheless varies across the 10 held-out children. Per-child exact match ranges from 72.6% to 100%. The cluster bootstrap confidence interval, which resamples children rather than treating every utterance as independent, places exact match between 83.8% and 93.2%.

That wider interval is the more relevant uncertainty estimate for deployment reasoning. The product would be used on new children, not on additional statistically independent recordings from the same conveniently familiar speakers.

The post-encoder helps, but the table does not validate every design choice

The paper includes two relevant variants alongside the proposed system.

Removing the six-layer post-encoder reduces exact sequence match from 88.7% to 84.5%, token accuracy from 95.0% to 90.2%, and screening F1 from 0.67 to 0.62. This is a controlled ablation: it supports the claim that task-specific temporal contextualization improves recognition of the short, sibilant-heavy sequences.

Replacing the wav2vec2 encoder with WavLM-Base while retaining the remaining setup produces 78.6% exact match, 86.6% token accuracy, and a screening F1 of 0.54. This is primarily an encoder comparison, not a general verdict that wav2vec2 dominates WavLM across child-speech tasks.

One qualification is worth making explicit: the displayed comparison table does not include a version trained without bracketed substitution tokens. The paper motivates bracket supervision strongly, and those tokens are central to the screening rule, but the table directly isolates the post-encoder and encoder selection—not the incremental effect of bracket labels themselves.

That does not invalidate the mechanism. It simply distinguishes an intuitively important design element from a quantitatively isolated one. Academic tables, like corporate roadmaps, occasionally claim a slightly broader jurisdiction than their rows actually provide.

Screening performance is the result that matters—and it is less flattering

Recognition accuracy answers whether the model reproduced the annotated sequence. Screening performance answers whether the system identified the target events it was built to flag.

For the primary sibilant targets, the held-out confusion matrix is:

This produces:

• 72.9% precision: roughly three out of four flags correspond to annotated mismatches.
• 61.4% recall: the system detects a little over three out of five annotated mismatches.
• 0.67 F1: the combined precision-recall score.
• 2.7% false-alarm rate: relatively few target-correct items are flagged.

Among true-positive flags, the system predicts the same bracketed substitution class as the reference in 85.7% of cases. When it detects a mismatch correctly, it can therefore usually preserve the specific error interpretation needed for localized feedback.

The low false-alarm rate is operationally attractive. Caregiver-facing tools can lose trust quickly if normal productions repeatedly trigger warnings. False positives can also create unnecessary anxiety and consume specialist capacity through avoidable referrals.

The recall figure is the necessary correction.

Of 57 annotated mismatch cases, 22 are not flagged. The screening rule therefore misses 38.6% of the target events represented in the held-out set. A negative result cannot be treated as clearance, and the system cannot safely support messages such as “no speech issue detected” without qualification.

This is the central misconception the paper must resist: 88.7% exact sequence match is not equivalent to 88.7% clinical screening accuracy.

The former supports a reliable alignment backbone. The latter does not exist here as a validated diagnostic metric. The paper’s actual mismatch detector is deliberately narrower and substantially less sensitive.

That profile may still be rational. In an unsupervised caregiver workflow, the cost of frequent false warnings may be higher than the cost of asking families to repeat the protocol or seek professional evaluation when broader concerns persist. But this is a product-policy choice, not a mathematical law. Different deployment settings will assign different costs to false positives and false negatives.

The remaining errors reveal a label problem, not just a model problem

Among the 63 test utterances without an exact sequence match, 58 contain exactly one sibilant-unit substitution. In other words, 92.1% of the non-exact cases are localized to a single sibilant decision rather than reflecting broad sequence failure.

Of those 58 localized substitutions, 35 are canonical-versus-bracket confusions.

This error analysis is not a second headline result. Its purpose is explanatory: it identifies where the recognizer still struggles and what kind of model revision may help.

The confusion suggests that the boundary between a canonical realization and bracketed substitution evidence may be graded. Yet the CTC classifier is asked to choose between them as separate discrete classes. The annotation may encode an expert judgment along a continuum that the model must treat as a hard categorical boundary.

The authors propose alternatives such as predicting the base phone and deviation flag through separate output heads, or smoothing labels between canonical and bracketed counterparts.

That direction makes sense. A two-part representation could distinguish:

1. which underlying speech category is acoustically closest; and
2. whether the realization provides evidence of a deviation from the target.

For product teams, the broader lesson is that explainable labels are not automatically easy labels. Adding human-interpretable classes can improve downstream communication while making the learning problem more ambiguous. Explainability has an annotation bill. Someone eventually receives it.

The practical product is a protocol, not an API endpoint

The system is best interpreted as a controlled screening protocol with an AI component.

A child receives a predefined set of prompts chosen to test specific contrasts. Recordings are processed by a recognizer trained for those prompts and token categories. The output is aligned to known targets. Only explicitly supported substitutions are typed in detail. Uncertain cases trigger repetition or deferral. Caregiver communication is restricted to approved templates.

That operating model is very different from exposing a general speech model and asking it to assess arbitrary recordings.

The strongest near-term business pathway is therefore workflow extension.

A school, telepractice provider, clinic network, or caregiver-support platform could use a similar protocol to collect structured speech samples before professional review. The system might highlight recurring contrasts, identify low-quality recordings, and organize evidence for a clinician.

Its value would come from reducing triage friction:

• fewer recordings requiring immediate manual inspection;
• more consistent elicitation of relevant speech contrasts;
• structured summaries rather than unorganized audio;
• earlier identification of cases worth professional attention;
• repeatable home practice without pretending that practice is treatment.

The paper does not measure any of those economic outcomes. They are business inferences from the workflow, not reported findings.

A valid deployment study would need to test whether the tool actually changes referral quality, clinician review time, caregiver adherence, or access delays. Model metrics are inputs to that evaluation, not substitutes for it.

Low false alarms may be commercially valuable, but sensitivity sets the ceiling

The screening rule’s 2.7% false-alarm rate could matter more commercially than a small improvement in token accuracy.

In caregiver-facing health-adjacent products, every false warning has downstream cost. It can reduce trust, create anxiety, encourage unnecessary consultation, and make clinicians skeptical of the entire intake channel. A conservative threshold can protect the workflow from alert fatigue before alert fatigue has acquired a pediatric accent.

Yet conservative detection also limits the product’s role.

With 61.4% recall, the system is better positioned to say:

“This production contains evidence worth repeating or reviewing.”

than:

“This child has passed the screen.”

The first statement routes positive evidence. The second requires dependable exclusion of problems, which the reported results do not establish.

This distinction should shape interface design, customer claims, and performance dashboards. Precision and false-alarm rates support the reliability of positive flags. Recall defines how much unseen risk remains after a negative result.

A commercially responsible product would make this asymmetry visible rather than hiding it beneath a single composite score.

“Parameter-efficient” does not mean small enough to disappear

The model uses LoRA modules together with unfreezing the last six layers of its speech encoder. Approximately 119.7 million of 359.6 million parameters remain trainable, or 33.3%.

That is efficient relative to full fine-tuning. It is not a tiny adaptation regime.

The distinction matters for cost interpretation. The paper demonstrates that the system can be adapted without updating every parameter, but it does not show on-device inference, real-time smartphone performance, edge deployment, or a low-cost production serving architecture.

The added six-layer post-encoder also contributes measurable accuracy while adding computation. Since the intended prompts are short, serving costs may still be manageable. The paper simply does not benchmark them.

For operators, the relevant takeaway is that the architecture is technically plausible for a bounded service—not that it has already solved deployment economics.

The current evidence stops before the home

The full dataset contains speech from 201 Polish-speaking children aged four to eight, recorded during on-site speech-language examinations. The recordings were collected in quiet educational environments using a dedicated multimodal device with 15 audio channels and two oral-region cameras.

The model itself receives mono audio resampled to 16 kHz, but the original collection environment remains important. It is substantially cleaner and more standardized than an ordinary kitchen, classroom, or budget smartphone recording.

The held-out evaluation contains 10 unseen children. This is enough to demonstrate speaker-disjoint testing, but not enough to characterize performance across age bands, severity levels, regional pronunciation patterns, device types, or household noise conditions.

Several other boundaries materially constrain interpretation:

The prompts are fixed

Training, validation, and test splits use the same prompt inventory. The system generalizes to new speakers producing known words and syllables. It has not demonstrated generalization to unseen lexical items or spontaneous speech.

For the proposed protocol, this is not necessarily a flaw. Fixed prompts are part of the screening design. It does mean the result should not be marketed as general Polish child-speech understanding.

The target is narrow

The work focuses on common substitutions involving Polish sibilants and affricates. Insertions, deletions, distortions outside the represented label scheme, and broader speech or language difficulties are not comprehensively covered.

Confidence is a suppression signal

The confidence value is calculated from mean CTC posterior probabilities over frames associated with the decoded token. The authors explicitly do not claim that it is calibrated.

A score of 0.9 should therefore not be presented to a caregiver as a 90% probability that a child made a particular error. It is an internal signal for controlling whether the assistant should respond in detail.

Clinical validity has not been established

The bracketed reference labels come from speech and language therapy experts, but the screening outputs have not yet been validated against clinical diagnoses, treatment decisions, or longitudinal outcomes.

The system detects its operational proxy: explicit bracket evidence at a target location. Whether that proxy produces clinically useful screening decisions remains a future research question.

The assistant has not been tested with its users

No caregiver or clinician user study is reported. Template grounding reduces the space of possible messages, but it does not automatically make those messages understandable, emotionally appropriate, or behaviorally safe.

A technically conservative message can still be misunderstood. Humans remain regrettably unwilling to behave like validation sets.

What an operator should validate next

A sensible pilot should evaluate the whole decision chain rather than celebrating the acoustic model in isolation.

The most important questions are:

• Does performance survive ordinary phones, room noise, distance from the microphone, and incomplete cooperation?
• Do repeat-recording requests improve evidence quality without causing families to abandon the session?
• How often do clinicians agree with the system’s localized error type?
• Does the tool reduce professional review time or merely add another dashboard?
• Do caregivers understand that a non-flag is not a diagnosis or clearance?
• Does template feedback change practice behavior appropriately?
• Which recall level is acceptable for the intended screening and escalation policy?
• How do results vary across ages, speech profiles, and degrees of deviation?

The commercial gate should be based on workflow outcomes: referral quality, clinician workload, completion rates, safety events, and user comprehension.

A few extra points of exact sequence match would be welcome. They are not the business case.

The valuable intelligence is the decision to remain bounded

This paper does not demonstrate automated diagnosis. It demonstrates the technical backbone of a conservative screening loop for one linguistically specific class of speech contrasts.

The recognizer is strong enough to support stable token alignment under the fixed protocol. The post-encoder contributes materially. Explicit substitution tokens enable interpretable flags. The false-alarm rate is low. The assistant design places useful constraints between acoustic evidence and caregiver language.

The system also misses a substantial share of annotated mismatches, operates on known prompts, relies on controlled data, and has not been validated with caregivers or against clinical outcomes.

Those facts are not awkward footnotes to an otherwise finished product. They define the product.

The most transferable contribution is the sequence of boundaries: preserve the anomaly, align it to a known target, type only supported errors, suppress uncertain advice, and escalate rather than diagnose.

In sensitive AI systems, intelligence is often framed as the ability to produce an answer. Here, the more commercially credible capability is knowing when the evidence supports only three words:

Please record again.

Cognaptus: Automate the Present, Incubate the Future.