Not All Speech Is Intent: Adaptive Self-Correcting Inference Layer for Post-ASR False Wake-Up

# Not All Speech Is Intent:Adaptive Self-Correcting Inference Layer for Post-ASR False Wake-Up
Source: [https://arxiv.org/html/2609.12469](https://arxiv.org/html/2609.12469)
## Not All Speech Is Intent: Adaptive Self\-Correcting Inference Layer for Post\-ASR False Wake\-UpThanks:\*Equal contribution\.†Corresponding author\.

Divya Neelagiri\*Anil Yadav\*Affiliation:Samsung Research America, Mountain View, CA, USAAffiliation:\{p\.saraswat, d\.neelagiri, Anil\.y1\}@samsung\.com

###### Abstract

False wake\-up activations remain a persistent challenge in conversational AI\. Speech phonetically similar to a device’s wake word can produce a syntactically valid and semantically coherent ASR transcript that the assistant incorrectly executes\. Most existing systems make a single intent decision in isolation, without a mechanism to learn from recurring errors over time or adapt to individual users through personalized learning\. We introduce theFeedback\-Driven Adaptive Self\-Correcting Inference Layer \(ASCIL\), a complementary post\-ASR correction framework that re\-evaluates wake\-up intent before response generation by fusing acoustic embeddings, linguistic cues, device context, and patterns from past misclassifications\. ASCIL interprets implicit signals, including hesitation, disengagement, and silence, and explicit signals, including cancellation and repetition, as automatically inferred, noisy behavioral indicators of potential misclassification\. These signals drive online pattern updates without manual annotation, whereas the intentional/unintentional reference labels used for offline evaluation are human\-annotated\. It generalizes from prior errors, applies corrective adjustments at inference time, and continuously updates in parallel with natural\-language execution\. Evaluated on a proprietary dataset of 3,667 interactions with human\-annotated intentional/unintentional reference labels spanning 14 acoustic and contextual conditions, ASCIL achieves54\.27% relative error reductionon a session\-disjoint subset constructed from baseline failures, and up to24\.39% relative error reductionat threshold 0\.90 on the issue\-tagged evaluation slice\. These gains are achieved while improving intentional acceptance rates, with a median added latency below 60 ms in the reported benchmark\.

Keywords:False Wake\-Up Detection, False Trigger Mitigation, Device\-Directed Speech Detection, Post\-ASR Intent Classification, User Feedback Learning, Continual Learning, On\-Device Personalization

## 1Introduction

Consider a user watching television when the dialogue contains a phrase phonetically similar to the device’s wake word\. The ASR \(Automatic speech recognition\) system transcribes the output cleanly in a complete valid transcript\. Nothing in the transcript reveals that the activation was unintentional\. The assistant responds\. The user ignores it, or says “stop\.” The same thing happens the next day\. Nothing changes\.

This scenario exposes a structural gap in deployed conversational AI\. While prior work has addressed same\-turn device\-directedness classification and other methods,a closed feedback loop that learns from user reactions after a wake\-up has already been handled remains largely unaddressed\. Once a wake\-up event produces a valid transcript, execution happens\. There is no mechanism to ask whether the activation was truly intended, and no mechanism to remember that this pattern caused an error before\. Two compounding problems make this worse\.

First,*static classifiers*: production systems are trained centrally on population\-level data and cannot adapt to individual users\. A user who has a regional accent, and keeps the television on while cooking experiences a systematically different error distribution than the user who works in a noisy environment, and existing systems offer no mechanism to correct for either pattern individually\.

Second,*no post\-ASR feedback loop*: user reactions to false activations like cancellations, repetitions, disengagement are not involved in current systems improvement or learning\. These signals are a source of noisy behavioral supervision that the system could learn from\. The user may provide an implicit or explicit behavioral signal after a potentially misclassified interaction, and the system can use that signal as noisy supervision for future pattern updates\.

We introduce aFeedback\-Driven ASCILthat addresses both problems\. Operating post\-wakeup and post\-ASR, the system fuses multimodal features to revalidate wake\-up intent before response generation\. When an automatically inferred feedback signal indicates a potential misclassification, the system extracts a generalized pattern capturing the acoustic, linguistic, and contextual conditions of that error, stores it compactly on\-device, and applies it at future inference to proactively suppress similar misclassifications\. The base ASR and NLU \(Natural Language Understanding\) models are not modified or retrained\. Our contributions are:

- •A*post\-ASR feedback loop*that is complementary to existing pre\-ASR and same\-turn device\-directedness classifiers: such a classifier can sit as the initial call in our pipeline, with our feedback loop improving its input conditions over time\.
- •A*feedback\-derived supervision mechanism*that interprets implicit signals \(hesitation, disengagement, silence\) and explicit signals \(cancellation, repetition\) as automatically inferred, noisy indicators of potential misclassification\. These signals support online pattern updates without additional manual annotation and are distinct from the human\-annotated reference labels used for offline evaluation\.
- •An*ASCIL*that extracts generalized patterns from prior feedback\-derived error hypotheses, applies them at inference, and updates continuously in parallel with NL execution, with a median added correction\-path latency below 60 ms in the reported benchmark\.
- •Empirical results on a real\-world proprietary dataset showing54\.27% relative error reductionon held\-out failing cases and up to24\.39%on the issue\-tagged evaluation slice at operating thresholds\.

## 2Related Work

Much prior work emphasizes wake\-word detection, false\-trigger mitigation, or same\-turn device\-directedness classification, with comparatively limited attention to feedback\-driven post\-decision adaptation from subsequent user reactions\.[Schönherr et al\. \(2020\)](https://arxiv.org/html/2609.12469#bib.bib1)systematically characterized accidental triggers across smart speakers\.[Chen et al\. \(2021\)](https://arxiv.org/html/2609.12469#bib.bib2)investigated the FakeWake phenomenon, generating fuzzy words that fool wake\-word detectors and proposing detector\-level remedies\.[Wu et al\. \(2018\)](https://arxiv.org/html/2609.12469#bib.bib9)proposed monophone\-based background modeling for two\-stage on\-device detection\.[Agarwal et al\. \(2020\)](https://arxiv.org/html/2609.12469#bib.bib8)explored complementary language modeling to reduce false triggers during decoding, and[Rudovic et al\. \(2023\)](https://arxiv.org/html/2609.12469#bib.bib13)unified multiple invocation types in a multi\-task architecture\.[Lin et al\. \(2025\)](https://arxiv.org/html/2609.12469#bib.bib12)addressed domain mismatch through personalized VAD models\.[Baljekar et al\. \(2014\)](https://arxiv.org/html/2609.12469#bib.bib10)and[Chaudhary and Abrol \(2023\)](https://arxiv.org/html/2609.12469#bib.bib11)addressed keyword spotting efficiency on\-device\.

Personalization for ASR was explored by[McGraw et al\. \(2016\)](https://arxiv.org/html/2609.12469#bib.bib4)using quantized LSTM acoustic models, and SpecAugment[Park et al\. \(2019\)](https://arxiv.org/html/2609.12469#bib.bib3)provided robust augmentation for training\.[Su et al\. \(2016\)](https://arxiv.org/html/2609.12469#bib.bib5)demonstrated on\-line reward learning for dialogue policy;[Liu and Mazumder \(2021\)](https://arxiv.org/html/2609.12469#bib.bib6)surveyed continual learning in dialogue systems\.

A closely related industrial disclosure is[Kanungo et al\. \(2026\)](https://arxiv.org/html/2609.12469#bib.bib7), which describes post\-ASR false wake\-up suppression using textual, acoustic, and contextual evidence together with adaptive decision thresholds\. Our work differs in that ASCIL focuses on post\-interaction behavioral feedback as noisy supervision and stores generalized user\-specific error patterns for future inference, without retraining the base ASR or NLU models\. Other Closest research work is[Mallidi et al\. \(2018\)](https://arxiv.org/html/2609.12469#bib.bib14)which addressed post\-ASR device\-directedness classification, determining whether an utterance was directed at the device using signals available immediately after transcription\. Their approach and ASCIL are complementary\. A same\-turn device\-directedness classifier can provide the initial decision, while ASCIL uses subsequent user behavior to refine future decisions\.

To our knowledge, prior false\-wake and device\-directedness systems have not reported a mechanism that converts automatically inferred post\-interaction feedback into persistent, generalized, on\-device corrective patterns without retraining the base models\. ASCIL differs in three ways: it learns from user reactions after a wake\-up has been executed or suppressed, maintains user\-specific memory of recurring misclassification patterns, and applies those patterns at inference time without updating the base ASR or NLU parameters\.

## 3System

### 3\.1Overview

Figure[1](https://arxiv.org/html/2609.12469#S3.F1)illustrates the pipeline\. At each interaction steptt, the system receives a multimodal feature vectorXt∈ℝdX\_\{t\}\\in\\mathbb\{R\}^\{d\}consolidating four signal categories: acoustic descriptors \(pitch, SNR, speech rate, HNR, barge\-in events, pre\-inverse\-text\-normalization \(pre\-ITN\) ASR text\), linguistic cues from ASR and NLU \(token and sentence embeddings, POS distributions, named entity presence, syntactic completeness, topic relevance\), temporal and device context \(session history, foreground application, wake modality, device motion\), and adaptive patterns from prior misclassifications\. A base classifier mapsXtX\_\{t\}to a binary intent predictionY^t∈\{0,1\}\\hat\{Y\}\_\{t\}\\in\\\{0,1\\\}, where11denotes an unintentional activation\. LetYtY\_\{t\}denote the human\-annotated intentional/unintentional reference label, used only for offline evaluation, and letFtF\_\{t\}denote an automatically inferred behavioral feedback signal, used only for online pattern updates\. ASCIL then refines the prediction:

Y^t′=C⁡\(Xt,Pu\(t\)\)\\hat\{Y\}^\{\\prime\}\_\{t\}=C\\\!\\left\(X\_\{t\},\\,P\_\{u\}^\{\(t\)\}\\right\)wherePu\(t\)P\_\{u\}^\{\(t\)\}is the on\-device user\-specific pattern memory at timett, andCCis the correction function\.

![Refer to caption](https://arxiv.org/html/2609.12469v1/figures/structered_figure.png)Figure 1:The adaptive correction pipeline\. Post\-ASR multimodal features feed a base post\-ASR wake\-up intent classifier\. User feedback signals trigger pattern extraction\. Stored patterns are injected as context at future inference, proactively correcting misclassifications without modifying base ASR or NLU models\.
### 3\.2Post\-ASR Intent Classification

The base classifier operates in two modes\. For standard activations, it predicts intentional vs\. unintentional fromXtX\_\{t\}\. When the input contains a user feedback signal like user cancellation, a repeated command, or sustained silence following an unexpected response, the classifier does not produce an intent decision\. Instead, it identifies the misclassification type from the prior turn feature vector:

- •False negative\(unintentional activation executed\): user hesitates, disengages, cancels, or ignores the system response in case of follow ups\.
- •False positive\(intentional activation suppressed\): user repeats the command, rephrases, or explicitly corrects\.

This feedback identification step is the critical architectural distinction\. Rather than discarding user reactions, the system treats them as structured supervision with two error types:Et∈\{FP,FN\}E\_\{t\}\\in\\\{\\mathrm\{FP\},\\mathrm\{FN\}\\\}\.

We distinguish two labeling processes\. The intentional/unintentional activation labels and the 14 acoustic and contextual condition tags used for offline evaluation \(§[4](https://arxiv.org/html/2609.12469#S4)\) are human\-annotated; they provide the reference ground truth for CAR, UICR, FIR, FAR, andPerrorP\_\{\\text\{error\}\}\. The feedback signals used for adaptive pattern induction are not manually annotated rather they are extracted automatically from the user’s next\-turn behavior in multi\-turn production logs, for example a same or paraphrased re\-invocation within a short window \(false\-positive signal\), or a cancellation, disengagement, or explicit rejection such as “stop” or “I didn’t ask you” \(false\-negative signal\)\. Because these behavioral signals can be ambiguous, we treat them as noisy supervision behavior labels rather than definitive ground\-truth labels\.

### 3\.3ASCIL

When an automatically inferred feedback signal indicates a potential misclassification, the system extracts a generalized pattern from the feature vectorXtX\_\{t\}of the previous turn and stores it in a structured schema keyed by misclassification type\. Forsuppressed\_intentionalerrors \(false positives\), patterns recordbackground\_noisetype,word\_count,speech\_energy, andspeech\_rate\_wpm\. Formissed\_unintentionalerrors \(false negatives\), patterns additionally capturepreITNTextcategory \(e\.g\., simple request vs\. command\) andasr\_transcriptclarity which indicates a cross\-model transcript agreement signal between Voxtral Mini 3B’s transcript of the audio and the base ASR transcript\. Agreement indicates a reliably\-transcribed utterance, while disagreement signals possible base\-ASR rounding to a phonetically nearer valid phrase, a common false\-negative condition\. Each pattern includes a human\-readable description injected as context to the LLM at inference\. Pattern memory updates as:

Pu\(t\)=Update⁡\(Pu\(t−1\),Xt,Y^t,Ft\)P\_\{u\}^\{\(t\)\}=\\mathrm\{Update\}\\\!\\left\(P\_\{u\}^\{\(t\-1\)\},\\,X\_\{t\},\\,\\hat\{Y\}\_\{t\},\\,F\_\{t\}\\right\)whereY^t\\hat\{Y\}\_\{t\}is the model’s prediction for the current activation andFtF\_\{t\}is an automatically inferred behavioral feedback signal observed after the interaction\. The feedback signal, together with the prior prediction, determines the feedback\-derived error hypothesis used as the pattern key\. The human\-annotated reference labelYtY\_\{t\}is used only for offline evaluation and is not available to the online pattern\-update process\. The model merges overlapping conditions and discards rare anomalies to prevent unbounded growth, minimizing cumulative misclassification loss\. No raw audio or verbatim transcripts are retained; only abstract feature signatures are stored for pattern matching and prompt construction\.

### 3\.4Deployment

The correction layer runs in parallel with standard NLU processing\. If an activation is classified as unintentional, the NLU output is silently suppressed; the user experiences no delay\. All pattern storage and updates occur on\-device with no data transmission\. At deployment time, the online pattern\-update process requires no additional manual annotation, no offline retraining, and no modification to base ASR or NLU models, making it a drop\-in layer for any voice assistant stack\. Offline evaluation in this work uses human\-annotated reference labels and condition tags, but these labels are not accessed by the deployed feedback\-update mechanism\. The correction layer adds a median latency below 60 ms on the target hardware in our benchmark with no user\-perceived latency\.

## 4Experiments

### 4\.1Dataset

We evaluate on a proprietary dataset of 3,667 interactions from a commercial voice assistant with human\-annotated intentional/unintentional reference labels: 2,358 intentional activations and 1,309 unintentional activations\. These reference labels are used only to compute offline evaluation metrics and are distinct from the feedback signals used for adaptive pattern induction, which are inferred automatically from subsequent user behavior\. Intentional instances carry multi\-label annotations across 14 overlapping acoustic and contextual conditions: background noise \(stationary, car, non\-stationary\), background speech \(single person, small group, large group\), foreground speech, pronunciation issues, audio cut\-offs, bad audio, multiple speakers, TTS speech, abandoned utterances, and a clean no\-issues baseline\. Table[1](https://arxiv.org/html/2609.12469#S4.T1)summarizes category distribution\.

Table 1:Per\-category occurrence counts\. Utterances are multi\-label; counts exceed total instances\.CategoryCountUnintentional wake\-up1,309Background noise – non\-stat\.1,593Background noise – car483Background noise – stationary528Background speech – single525Background speech – small grp159Background speech – large grp6Pronunciation issue476Audio cut\-off373Foreground speech87Bad audio100Multiple speakers3TTS speech5Abandoned utterance8No issues \(clean\)100
### 4\.2Setup

The baseline is a prompt\-only classifier built on Qwen3\-14B\-Instruct, producing an intent label and confidence score from structured multimodal prompts without parameter updates, enriched with annotations from Voxtral Mini 3B, an audio\-understanding model\. It serves as a direct ablation of the pattern memory component with same multimodal features and LLM classifier as ASCIL, but without pattern storage or feedback\-driven correction isolating the adaptive layer’s contribution\.

The proposed system augments this baseline with ASCIL\. We use two distinct threshold criteria\. For pattern induction, we select baseline misclassifications with confidence≥0\.80\\geq 0\.80and split them evenly into a feedback\-builder subset \(N=180\) for pattern induction and a held\-out subset \(N=180\) for measuring generalization\. This split is session\-disjoint\. The two subsets share no interaction sessions\. We do not enforce a user\-disjoint split, since the data spans multi\-turn production interactions with substantial per\-user variability, and a strict user\-disjoint split would leave many users too sparse to evaluate; session\-level disjointness preserves evaluation power while avoiding interaction\-level overlap between the subsets\. Consequently, the held\-out subset measures transfer across sessions rather than generalization to unseen users\. The human\-annotated reference labels are used to identify baseline errors and to compute offline metrics for the builder and held\-out subsets\. They are not available to ASCIL during online feedback\-driven pattern induction or inference\. The 0\.80 threshold here serves only as a selection criterion for high\-confidence errors worth learning from and it is not a reporting threshold\. For evaluation, we report all metrics at operating thresholds of 0\.85, 0\.90, and 0\.95\. Activations with predicted\-unintentional confidence below the threshold are treated as intentional, those at or above as unintentional\. Patterns capture SNR ranges, pitch statistics, background noise type, stop\-word ratios, POS distributions, and ASR transcript type, with natural language text descriptions to guide the LLM at inference\.

### 4\.3Metrics

We report five task\-specific metrics:CAR\(Correct Acceptance Rate, i\.e\., the proportion of intentional activations correctly accepted,↑\\uparrow\);UICR\(Unintentional Interception Rate, the proportion of unintentional activations correctly intercepted,↑\\uparrow\);FIR\(False Interception Rate, intentional activations incorrectly suppressed,↓\\downarrow\);FAR\(False Acceptance Rate, unintentional activations incorrectly accepted,↓\\downarrow\); andPerrorP\_\{\\text\{error\}\}\(the overall proportion of incorrect decisions,↓\\downarrow\)\. All five metrics are computed against the human\-annotated reference labels\. The system prioritizes precision over recall, suppressing false activations while minimizing the rejection of genuine requests\.

## 5Results

Table[2](https://arxiv.org/html/2609.12469#S5.T2)presents the complete evaluation across all conditions and operating thresholds\.

Table 2:Complete evaluation results at operating confidence thresholds \(0\.85, 0\.90, 0\.95\)\. CAR: Intentional Correct Acceptance Rate \(%\); UICR: Unintentional Intercept Rate \(%\); FIR: False Interception Rate \(%\); FAR: False Acceptance Rate \(%\);PerrorP\_\{\\text\{error\}\}: Total Error Rate \(%\), computed as\(FP\+FN\)/N\(\\text\{FP\}\{\+\}\\text\{FN\}\)/N\.Δ​Perror\\Delta P\_\{\\text\{error\}\}: relative error reduction \(%\),\(Perrorbase−Perrorpattern\)/Perrorbase×100\(P\_\{\\text\{error\}\}^\{\\text\{base\}\}\-P\_\{\\text\{error\}\}^\{\\text\{pattern\}\}\)/P\_\{\\text\{error\}\}^\{\\text\{base\}\}\\times 100; positive denotes improvement\. Per\-slice eval counts: No\-Issues \(Nint=100N\_\{\\text\{int\}\}\{=\}100,Nunint=1,309N\_\{\\text\{unint\}\}\{=\}1\{,\}309, totalN=1,409N\{=\}1\{,\}409\); Issue\-Tagged \(Nint=2,258N\_\{\\text\{int\}\}\{=\}2\{,\}258,Nunint=1,309N\_\{\\text\{unint\}\}\{=\}1\{,\}309, totalN=3,567N\{=\}3\{,\}567\)\. The No\-Issues slice contains the 100 intentional utterances tagged as no\-issues \(clean audio\); the Issue\-Tagged slice contains the remaining 2,258 intentional utterances \(each carrying at least one acoustic or contextual condition tag\)\. The 1,309 unintentional samples are shared across both slices, so UICR and FAR are identical by construction between the No\-Issues and Issue\-Tagged conditions at a given threshold; only CAR and FIR vary\. Total intentional accounting: 100 \(No\-Issues\) \+ 2,258 \(Issue\-Tagged\) = 2,358\. The Held\-Out set \(N=180\) is session\-disjoint from the feedback\-builder set \(N=180\) used for pattern induction and is the only split intended to measure transfer beyond the feedback\-builder interactions; it is session\-disjoint rather than user\-disjoint\. The Issue\-Tagged slice includes builder\-derived cases and therefore measures overall system behavior rather than clean held\-out generalization\.ConditionSystemThresh\.CAR↑\\uparrowUICR↑\\uparrowFIR↓\\downarrowFAR↓\\downarrowPerrorP\_\{\\text\{error\}\}↓\\downarrowΔ​Perror\\Delta P\_\{\\text\{error\}\}No\-Issues\(clean Intentional audio\)Baseline \(only model\)0\.8594\.0055\.316\.0044\.6941\.94—model with ASCIL layer0\.8598\.0043\.162\.0056\.8452\.95−\-26\.25Baseline \(only model\)0\.9097\.0025\.443\.0074\.5669\.48—model with ASCIL layer0\.9098\.0040\.722\.0059\.2855\.22\+20\.52Baseline \(only model\)0\.95100\.002\.830\.0097\.1790\.28—model with ASCIL layer0\.95100\.0017\.950\.0082\.0576\.22\+15\.57Issue\-Tagged\(multi\-condition intentionalactivations resemblingunintentional activations\)Baseline \(only model\)0\.8579\.0555\.3120\.9544\.6929\.66—model with ASCIL layer0\.8593\.1843\.166\.8256\.8425\.18\+15\.10Baseline \(only model\)0\.9089\.1525\.4410\.8574\.5634\.23—model with ASCIL layer0\.9093\.4940\.726\.5159\.2825\.88\+24\.39Baseline \(only model\)0\.9599\.782\.830\.2297\.1735\.80—model with ASCIL layer0\.9597\.2117\.952\.7982\.0531\.88\+10\.94Held\-Out\(all failing cases\)Baseline \(only model\)0\.90consistent failure on all samples \(Perror=100%P\_\{\\text\{error\}\}\{=\}100\\%by construction\)—model with ASCIL layer0\.9070\.3239\.8829\.6860\.1245\.73\+54\.27

### 5\.1Session\-disjoint held\-out subset

The held\-out subset comprises interactions misclassified by the baseline across both false\-positive and false\-negative types\. So baselinePerrorP\_\{\\text\{error\}\}is 100% on this subset by construction \(UICR = 0% on all unintentional cases\)\. The pattern layer recovers meaningful performance, achieving 70\.32% CAR and 39\.88% UICR at threshold 0\.90 and reducingPerrorP\_\{\\text\{error\}\}to 45\.73%; this corresponds to a54\.27% relative error reductionagainst the failure baseline\. Because this subset is constructed entirely from baseline failures, the figure reports conditional recovery from those failures rather than an absolute error reduction over the full dataset\. This suggests that the learned patterns transfer to session\-disjoint, structurally similar misclassification cases\. Because the split is session\-disjoint but not user\-disjoint, this result measures transfer across sessions for users who may occur in both subsets, rather than generalization to unseen users\.

### 5\.2Intenional activations with multi labeled Issues\-tagged evaluation slice\.

On the intentional with issue\-tagged evaluation slice \(2,258 intentional utterances carrying at least one issue condition tag, plus the shared 1,309 unintentional utterances\), ASCIL reducesPerrorP\_\{\\text\{error\}\}at all reported thresholds, with the largest relative reduction of24\.39%at threshold 0\.90\. This slice includes builder\-derived cases and is therefore a system\-level diagnostic rather than a contamination\-free held\-out test\. The issue\-tagged intentional activations combine multiple acoustic and contextual conditions, making them closely resemble unintentional activations\. These challenging cases involve simultaneous multi\-factor conditions like background noise, foreground speech, and pronunciation variation occurring together\. At threshold 0\.90, ASCIL improves both intentional acceptance and unintentional interception: CAR rises from 89\.15% to 93\.49%, while UICR improves from 25\.44% to 40\.72%\. Correspondingly, FIR decreases from 10\.85% to 6\.51% and FAR decreases from 74\.56% to 59\.28%\. These simultaneous improvements are threshold\-dependent\. At other thresholds, ASCIL exhibits different acceptance–interception trade\-offs\.

### 5\.3No\-issues dataset\.

On clean audio, the ASCIL produces mixed results at threshold 0\.85 \(−26\.25%\-26\.25\\%Δ​Perror\\Delta P\_\{\\text\{error\}\}\)\. At this lower threshold, the pattern layer shifts decisions toward acceptance\. FIR improves \(6\.00%→\\rightarrow2\.00%\) and CAR rises \(94\.00%→\\rightarrow98\.00%\), but FAR worsens \(44\.69%→\\rightarrow56\.84%\) and UICR falls \(55\.31%→\\rightarrow43\.16%\)\. Because the unintentional class dominates the No\-Issues evaluation slice \(1,309 vs\. 100 samples\), the FAR increase drives the overall rise inPerrorP\_\{\\text\{error\}\}\. This behavior reflects a limitation at the 0\.85 threshold and motivates further evaluation of operating points at or above 0\.90\. At operating thresholds of 0\.90 and 0\.95, the ASCIL recovers strongly, reducingPerrorP\_\{\\text\{error\}\}by 20\.52% and 15\.57% respectively\. At threshold 0\.90, CAR improves from 97\.00% to 98\.00% and UICR rises substantially from 25\.44% to 40\.72%\. At threshold 0\.95, UICR improves from 2\.83% to 17\.95% while CAR remains at 100%\. This precision\-first behavior of filtering edge\-case false positives without degrading strong true positives is particularly valuable in clean\-audio conditions where false suppression of genuine requests directly erodes user trust\. Because the unintentional samples are shared across the No\-Issues and Issue\-Tagged slices, UICR and FAR are identical by construction between the two conditions\. The CAR and FIR differences are where the discriminative signal lies\. Issue\-Tagged intentional utterances carry acoustic and contextual conditions \(background noise, background speech, pronunciation issues, cutoffs\) that acoustically resemble unintentional wake\-ups, making them harder to classify correctly\. The CAR and FIR improvements observed on Issue\-Tagged therefore reflect improved discrimination on this harder population, not a change in unintentional detection\.

### 5\.4Interpretation\.

The results reveal a consistent pattern across conditions\. The adaptive layer is most effective at the higher operating thresholds \(0\.90–0\.95\) where the model is operating at high confidence\. In this framework, the ASCIL adds targeted corrections for the specific acoustic and contextual signatures it has learned, without disrupting the base model’s already\-reliable judgments\. From a user\-experience perspective, this translates to fewer disruptive unintended responses and fewer suppressed genuine commands\. These are the two failure modes that most directly erode trust in voice interfaces\.

## 6Conclusion

We introduced a feedback\-driven ASCIL that closes a specific gap left open by prior post\-ASR approaches\. A closed loop that learns from user reactions*after*a wake\-up has already been handled by treating implicit and explicit user feedback as noisy behavioral supervision, extracting generalized on\-device patterns, and applying corrections at inference without retraining\. This framework achieves meaningful, consistent improvements on real\-world multi\-condition audio\. The layer autonomously decides when to extract patterns from feedback signals and when to apply them at inference, without external orchestration\. The current architecture keeps added latency low by running in parallel to NLU and suppressing the response rather than the request, generalizing beyond false wake\-up suppression to any conversational AI setting where systematic misclassifications are exposed by user feedback\. We present these results as a preliminary systems study rather than definitive evidence of personalized generalization or production\-level efficacy\.

## Limitations

ASCIL’s effectiveness depends on sufficient user interaction history, leaving cold\-start adaptation for new users as a known challenge\. Feedback signals are ambiguous like silence for example may indicate disengagement, satisfaction, or a natural pause\. The current study does not independently validate the precision and recall of the feedback heuristics\. The evaluation uses a proprietary dataset with session\-disjoint rather than user\-disjoint splits, so generalization across unseen users, languages, device categories, and interaction styles remains a part of future work\.

## Ethics Statement

Interactions were collected from users who consented to data usage for product improvement under the product’s terms of service\. Samples were de\-identified, with personally identifiable information removed from content prior to annotation\. Users are referenced only by pseudonymous stable identifiers that support session\-level grouping without identity linkage\. Annotators accessed data only through a controlled labeling interface, and data retention follows standard company policy\. Demographics are not tracked at the interaction level for the evaluation dataset used in this work\.

## References

- Agarwalet al\.\(2020\)R\. Agarwal, X\. Niu, P\. Dighe, S\. Vishnubhotla, S\. Badaskar, and D\. NaikComplementary language model and parallel bi\-lrnn for false trigger mitigation\.arXiv preprint arXiv:2008\.08113\.Cited by:[§2](https://arxiv.org/html/2609.12469#S2.p1.1)\.
- Baljekaret al\.\(2014\)P\. Baljekar, J\. F\. Lehman, and R\. SinghOnline word\-spotting in continuous speech with recurrent neural networks\.In2014 IEEE Spoken Language Technology Workshop \(SLT\),pp\. 536–541\.Cited by:[§2](https://arxiv.org/html/2609.12469#S2.p1.1)\.
- Chaudhary and Abrol \(2023\)A\. Chaudhary and V\. AbrolTowards on\-device keyword spotting using low\-footprint quaternion neural models\.In2023 IEEE Workshop on Applications of Signal Processing to Audio and Acoustics \(WASPAA\),pp\. 1–5\.Cited by:[§2](https://arxiv.org/html/2609.12469#S2.p1.1)\.
- Chenet al\.\(2021\)Y\. Chen, Y\. Bai, R\. Mitev, K\. Wang, A\. Sadeghi, and W\. XuFakewake: understanding and mitigating fake wake\-up words of voice assistants\.InProceedings of the 2021 ACM SIGSAC conference on computer and communications security,pp\. 1861–1883\.Cited by:[§2](https://arxiv.org/html/2609.12469#S2.p1.1)\.
- Kanungoet al\.\(2026\)T\. Kanungo, P\. Saraswat, and S\. M\. WALSHSystem and method for post\-asr false wake\-up suppression\.Google Patents\.Note:US Patent 12,536,993Cited by:[§2](https://arxiv.org/html/2609.12469#S2.p3.1)\.
- Linet al\.\(2025\)Y\. Lin, J\. Chen, W\. Li, L\. Xiao, and C\. WengRobust personal voice activity detection for mitigating domain mismatch and false acceptance scenarios\.InProc\. Interspeech 2025,pp\. 5783–5787\.Cited by:[§2](https://arxiv.org/html/2609.12469#S2.p1.1)\.
- Liu and Mazumder \(2021\)B\. Liu and S\. MazumderLifelong and continual learning dialogue systems: learning during conversation\.InProceedings of the AAAI conference on artificial intelligence,Vol\.35,pp\. 15058–15063\.Cited by:[§2](https://arxiv.org/html/2609.12469#S2.p2.1)\.
- Mallidiet al\.\(2018\)S\. H\. Mallidi, R\. Maas, K\. Goehner, A\. Rastrow, S\. Matsoukas, and B\. HoffmeisterDevice\-directed utterance detection\.arXiv preprint arXiv:1808\.02504\.Cited by:[§2](https://arxiv.org/html/2609.12469#S2.p3.1)\.
- McGrawet al\.\(2016\)I\. McGraw, R\. Prabhavalkar, R\. Alvarez, M\. G\. Arenas, K\. Rao, D\. Rybach, O\. Alsharif, H\. Sak, A\. Gruenstein, F\. Beaufays,et al\.Personalized speech recognition on mobile devices\.In2016 IEEE International Conference on Acoustics, Speech and Signal Processing \(ICASSP\),pp\. 5955–5959\.Cited by:[§2](https://arxiv.org/html/2609.12469#S2.p2.1)\.
- Parket al\.\(2019\)D\. S\. Park, W\. Chan, Y\. Zhang, C\. Chiu, B\. Zoph, E\. D\. Cubuk, and Q\. V\. LeSpecaugment: a simple data augmentation method for automatic speech recognition\.arXiv preprint arXiv:1904\.08779\.Cited by:[§2](https://arxiv.org/html/2609.12469#S2.p2.1)\.
- Rudovicet al\.\(2023\)O\. Rudovic, W\. Chang, V\. Garg, P\. Dighe, P\. Simha, J\. Berkowitz, A\. H\. Abdelaziz, S\. Kajarekar, E\. Marchi, and S\. AdyaLess is more: a unified architecture for device\-directed speech detection with multiple invocation types\.InICASSP 2023\-2023 IEEE International Conference on Acoustics, Speech and Signal Processing \(ICASSP\),pp\. 1–5\.Cited by:[§2](https://arxiv.org/html/2609.12469#S2.p1.1)\.
- Schönherret al\.\(2020\)L\. Schönherr, M\. Golla, T\. Eisenhofer, J\. Wiele, D\. Kolossa, and T\. HolzUnacceptable, where is my privacy? exploring accidental triggers of smart speakers\.arXiv preprint arXiv:2008\.00508\.Cited by:[§2](https://arxiv.org/html/2609.12469#S2.p1.1)\.
- Suet al\.\(2016\)P\. Su, M\. Gasic, N\. Mrkšić, L\. M\. R\. Barahona, S\. Ultes, D\. Vandyke, T\. Wen, and S\. YoungOn\-line active reward learning for policy optimisation in spoken dialogue systems\.InProceedings of the 54th Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),pp\. 2431–2441\.Cited by:[§2](https://arxiv.org/html/2609.12469#S2.p2.1)\.
- Wuet al\.\(2018\)M\. Wu, S\. Panchapagesan, M\. Sun, J\. Gu, R\. Thomas, S\. N\. P\. Vitaladevuni, B\. Hoffmeister, and A\. MandalMonophone\-based background modeling for two\-stage on\-device wake word detection\.In2018 IEEE International Conference on Acoustics, Speech and Signal Processing \(ICASSP\),pp\. 5494–5498\.Cited by:[§2](https://arxiv.org/html/2609.12469#S2.p1.1)\.