Paper deep dive
Phoneme-Level Mispronunciation Screening in Polish-Speaking Children with an Explainable Assistant
Milosz Dudek, Daria Hemmerling, Kamil Kwarciak, Maciej Stroinski, Maria Pensko, Mateusz Kowalewski, Leonid Pavlovskyi, Sebastian Jurczak, Anna-Mariia Vitkovska, Zuzanna Miodonska, Natalia Mocko, Michal Krecichwost
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 90%
Last extracted: 7/9/2026, 7:42:10 AM
Summary
The paper introduces a lightweight, phoneme-level mispronunciation screening pipeline for Polish-speaking children to address limited access to speech specialists. The system combines a wav2vec2-based acoustic model with CTC training, LoRA adaptation, and alignment-based error typing to detect sibilant substitutions. It outputs a conservative screening proxy that flags mismatches with 88.7% exact sequence match and an F1 score of 0.67. An explainable, template-grounded caregiver assistant translates token-level alignments into safe, interpretable feedback, explicitly avoiding clinical diagnosis. The approach prioritizes phonetic sensitivity and operational explainability over general transcription quality, with planned clinician-in-the-loop validation.
Entities (8)
Relation Signals (6)
mispronunciation_screening_pipeline → achieves → 88.7%_exact_sequence_match
confidence 98% · the recognizer achieves 88.7 percent exact sequence match
Wav2Vec2 → usedin → mispronunciation_screening_pipeline
confidence 95% · coupling a wav2vec2-based CTC token recognizer with alignment-based error typing
caregiver_assistant → provides → template_grounding
confidence 92% · template-grounded caregiver assistant for screening, not diagnosis
speech_sound_disorders → motivatedevelopmentof → lightweight_screening_tools
confidence 90% · Early identification of speech sound errors in children is often limited by access to specialists, motivating lightweight screening tools
LoRA_adaptation → appliedto → Wav2Vec2
confidence 88% · combine LoRA-based adaptation with partial unfreezing... injected into attention projections
Polish_child_speech_corpus → evaluates → mispronunciation_screening_pipeline
confidence 85% · On a held-out test set of 10 unseen children comprising 559 utterances, the recognizer achieves 88.7 percent exact sequence match
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Early identification of speech sound errors in children is often limited by access to specialists, motivating lightweight screening tools that can operate outside the clinic. We present a screening pipeline for Polish-speaking children focused on sibilant substitutions, coupling a wav2vec2-based CTC token recognizer with alignment-based error typing and a template-grounded caregiver assistant for screening, not diagnosis. On a held-out test set of 10 unseen children comprising 559 utterances, the recognizer achieves 88.7 percent exact sequence match. As a conservative screening proxy, we flag a mismatch when the system emits substitution-evidence bracketed tokens at the target segment, yielding 72.9 percent precision, 61.4 percent recall, F1 = 0.67, and a 2.7 percent false-alarm rate on target-correct items. We describe the assistant's safety boundaries and outline a clinician-in-the-loop validation plan for future deployment.
Tags
Links
- Source: https://arxiv.org/abs/2606.25181v1
- Canonical: https://arxiv.org/abs/2606.25181v1
Trouble viewing inline? Open PDF directly →
Full Text
27,661 characters extracted from source content.
Expand or collapse full text
Phoneme-Level Mispronunciation Screening in Polish-Speaking Children with an Explainable Assistant Milosz Dudek ID 1,2,∗ , Daria Hemmerling ID 1,2 , Kamil Kwarciak ID 1,2 , Maciej Stroinski 2 , Maria Pensko 2 , Mateusz Kowalewski 2 , Leonid Pavlovskyi 2 , Sebastian Jurczak 2 , Anna-Mariia Vitkovska 2 , Zuzanna Miodonska ID 3 , Natalia Mocko ID 4 , Michal Krecichwost ID 3 1 AGH University of Krakow, Cracow, Poland 2 SoftServe, Cracow, Poland 3 Department of Biomedical Engineering, Silesian University of Technology, Poland 4 Institute of Linguistics, Faculty of Humanities, University of Silesia in Katowice, Poland miloszdudek@agh.edu.pl, hemmer@agh.edu.pl, kwarciak@agh.edu.pl, mstro@softserveinc.com, mpens@softserveinc.com, mkowale@softserveinc.com, lpavlov@softserveinc.com, sjurc@softserveinc.com, avitk@softserveinc.com, zuzanna.miodonska@polsl.pl, natalia.mocko@us.edu.pl, michal.krecichwost@polsl.pl Abstract Early identification of speech sound errors in children is of- ten limited by access to specialists, motivating lightweight screening tools that can operate outside the clinic. We present a screening pipeline for Polish-speaking children focused on sibilant substitutions, coupling a wav2vec2-based CTC token recognizer with alignment-based error typing and a template- grounded caregiver assistant (screening, not diagnosis). On a held-out test set of 10 unseen children (559 utterances), the rec- ognizer achieves 88.7% exact sequence match. As a conserva- tive screening proxy, we flag a mismatch when the system emits substitution-evidence (bracketed) tokens at the target segment, yielding 72.9% precision, 61.4% recall (F1=0.67) and a 2.7% false-alarm rate (FPR on target-correct items). We describe the assistant’s safety boundaries and outline a clinician-in-the-loop validation plan for future deployment. Index Terms: child speech, mispronunciation detection, speech sound disorders, Polish, phoneme recognition, wav2vec2, ex- plainable feedback 1. Introduction Speech sound disorders (SSDs) are among the most common developmental communication difficulties in early childhood and, when persistent, can affect intelligibility, participation, and later literacy-related outcomes [1]. In many settings, timely access to specialist assessment and therapy is constrained by limited capacity and long waiting times, motivating comple- mentary approaches that support early screening and caregiver- guided practice outside the clinic [2, 3]. Polish is a particularly demanding language for automatic screening because it combines a dense consonantal inventory with frequent clusters. In Polish clinical practice, sibilants are often described as three series: sycz ̨acy (e.g., /s z > ts > dz/), cisz ̨acy (e.g., /Cý > tC > dý/), and szumi ̨acy (e.g., /ù ü > tù > dü/) [4, 5]. Many referrals in Polish speech therapist practice concern er- rors in this space (often discussed under sigmatism or lisping), where substitutions (e.g., /ù/→[s]) and distortions may require different clinical handling [5, 6]. ** indicates the corresponding author. A practical home screening interaction must be short, re- peatable, and focused on high-yield contrasts. We therefore build the screening loop around prompted words and syllables that probe place and manner contrasts relevant to Polish sibi- lants, including szafa, szufelka, sznurek, stra ̇ zak, czapka, dzi- adek, dzwonek, zegar, and minimal syllables such as ca/cia/cza. The goal is not a fully general transcription system, but a sensi- tive and interpretable mechanism to flag likely errors and local- ize them at the token level. Mispronunciation detection and diagnosis (MDD) has pro- gressed from classical Goodness-of-Pronunciation pipelines to- ward neural architectures that aim to detect and localize errors at phoneme or feature levels [7, 8]. However, directly apply- ing off-the-shelf ASR to child screening is problematic [9, 10]. Child speech differs acoustically from adult speech (shorter vo- cal tracts and higher formants), shows higher within-speaker variability, and includes systematic non-canonical realizations that are part of development [9, 11]. Moreover, strong language- model priors in end-to-end ASR can normalize unusual realiza- tions toward likely words, which is helpful for transcription but harmful for minimal-contrast error detection [12]. We therefore propose a caregiver-oriented screening loop that prioritizes phonetic sensitivity and operationally defined explainability over general transcription quality (Figure 1). In this paper, explainable means: (i) the system produces an au- ditable token-level alignment and named error type, (i) care- giver feedback is generated from a fixed, reviewable tem- plate inventory, and (i) the system has explicit suppres- sion/refusal rules under uncertainty. The acoustic module is a self-supervised speech encoder fine-tuned for Polish token recognition [13, 14, 15]. To make common sibilant substitu- tions explicit, we extend the token inventory with a small set of bracketed IPA tokens that represent expert-judged substitu- tion outcomes (e.g., “closest to [s]”). On top of recognition, we perform alignment-based screening: recognized production tokens are aligned to the prompted canonical sequence to ex- tract an interpretable diagnostic vector, which is then translated into caregiver-friendly feedback by an assistant constrained by templates [16]. Importantly, we do not claim clinical diagnosis. We evalu- ate recognition and a conservative screening proxy on held-out children, and we present the assistant as a deployable design arXiv:2606.25181v1 [eess.AS] 23 Jun 2026 constrained by clinician-authored templates; full clinical and caregiver validation is part of our planned next step. Our contributions are: • a substitution-aware token recognizer for Polish child speech built on wav2vec2 with a compact 6-layer Transformer post- encoder, • a screening-oriented evaluation on held-out children that re- ports both recognition quality and conservative detection of target sibilant mismatches, • a template-grounded assistant design that translates token- level screenings into caregiver-friendly feedback with ex- plicit safety boundaries and a clinician-in-the-loop validation plan. 2. Data We use a proprietary corpus of prompted Polish child speech licensed for research [6]. Each utterance is associated with (i) the orthographic prompt, (i) a canonical token sequence, and (i) a token-level transcription of the child’s production. Aggregate, privacy-preserving dataset description: • Participants: 201 children (107F/94M), age 4–8 years (mean 76.2 months, SD 8.6), native Polish speakers recruited from schools/kindergartens in Poland. • Recording conditions:On-site SLT examinations us- ing a dedicated multimodal device:15-channel audio (44.1 kHz/16-bit) and dual-camera oral-region video (30 fps). Quiet educational environments. • Prompt inventory: 51 words and 12 logotomes covering all 12 Polish sibilants, elicited via picture naming and repetition. Total: 12,830 word/logotome segments and 53,951 phoneme segments. • Annotation protocol: Manual segmentation into words and phonemes with IPA labels. For common sibilant substitu- tions, speech and language therapy experts provided brack- eted substitution evidence (Section 3.1). 3. Method 3.1. Token representation and substitution markers We represent pronunciations as sequences of space-delimited IPA-like tokens for CTC training and decoding. Tokens are pri- marily phoneme-like segmental units; importantly, Polish af- fricates are represented as single tied symbols (e.g., > ts, > dz) in line with Polish phonology, even though they are written with digraphs in orthography (e.g., c, dz, cz, d ̇ z). Tokeniza- tion choices are driven by CTC stability on very short prompts: in a small subset of items we keep a fused consonant–vowel label used in the corpus annotations to reduce boundary jitter, hence we refer to tokens rather than strictly phonemes and re- port token-sequence length statistics in Section 4. To retain evidence of clinically frequent sibilant substitu- tions, we augment the label set with a small number of brack- eted IPA tokens (e.g., [s], [ù], [C], [ > ts], [ > tù]). These markers encode an expert-assigned “closest-match” label for a substi- tuted target and are treated as separate classes during training and inference. Bracketed tokens are used conservatively at screening time: we treat a bracketed token at the target position as explicit evidence of a likely deviation, while non-bracketed but non- canonical outputs are handled as uncertain (e.g., request a re- peat recording or defer to speech-language expert review). This choice reduces false alarms in caregiver-facing use, and aligns with the assistant safety rules. The scheme covers 12 common outcomes, grouped by the three Polish sibilant series (ortho- graphic forms in italics): (s, z, c, dz)→ [s], [z], [ > ts], [ > dz]; ( ́ s, ́z, ́c, ́ d ́z)→ [C], [ý], [ > tC], [ > dý]; (sz, ̇ z/rz, cz, d ̇ z)→ [ù], [ü], [ > tù], [ > dü]. 3.2. Acoustic model We formulate token recognition as CTC-based sequence label- ing [17]. The model consists of: (i) a wav2vec2 encoder ini- tialized from a Polish checkpoint: jonatasgrosman/wav2vec2- large-xlsr-53-polish, (i) a 6-layer Transformer post-encoder that refines temporal context, and (i) a linear projection to the token vocabulary trained with the CTC objective. We add the post-encoder to stabilize CTC emissions in sibilant-heavy seg- ments and to provide task-specific contextualization. In the con- trolled ablation reported in Table 2, removing the post-encoder reduces test exact sequence match from 88.7% to 84.5% (-4.2 p) under otherwise identical training and decoding settings. 3.3. Parameter-efficient adaptation To adapt the model to child speech and the extended token in- ventory, we combine LoRA-based adaptation [18] with partial unfreezing. Specifically, LoRA modules are injected into atten- tion projections and feed-forward sublayers with rank r=32, α=64, and dropout 0.1, and we additionally unfreeze the last 6 encoder layers.This yields 33.3% trainable parameters (119.7M out of 359.6M). We describe this as efficient relative to full fine-tuning rather than fully lightweight. 3.4. Alignment-based screening and confidence Given a prompted canonical sequencey ref and a recognized production sequence ˆ y, we compute a minimum-edit (Leven- shtein) alignment over tokens. We interpret alignment opera- tions as: • Match: token-level agreement. • Substitution: candidate mispronunciation event (or bracket evidence). • Insertion/Deletion: candidate epenthesis/omission (reported conservatively). For substitutions involving the sibilant/affricate focus set (Sec- tion 3.1), we map the target–realized pair to an interpretable cat- egory such as (i) a place-of-articulation shift within the sibilant space, (i) a voicing alternation, or (i) a manner-of-articulation mismatch (affrication vs. frication). Substitutions outside this focus set (e.g., non-sibilant tokens such as /f/ or /v/) are reported as generic mismatches without fine-grained typing, keeping the screening logic conservative. This yields a com- pact diagnostic vector: prompt, target, realized, type, position, confidence. To calculate confidence let p t (k) denote the frame-level posterior probability for token k at frame t under the CTC model. For an aligned target position i with realized token ˆy i , we define a confidence score as the mean posterior of ˆy i over the set of frames assigned to it by greedy CTC collapse: c i = 1 |T i | X t∈T i p t (ˆy i ), c i ∈ [0, 1]. Here, T i denotes the set of non-blank CTC frames assigned to the aligned realized token ˆy i during greedy decoding, after repeated-label collapse and blank removal. The assistant uses Input Child Audio (Prompted) Polish wav2vec 2.0 Init 6-Layer Transformer Post-Encoder CTC Head (Extended Vocab) Acoustic Recognition Backbone Recognized Sequence ˆ y (w/ Subst. Tags) /s z [ts] .../ Levenshtein Alignment & Error Typing Canonical Prompt y ref Diagnostic Vector [Target, Realized, Type, Conf] Screening Logic Explainable Assistant (Template-Grounded) Caregiver Report Safe guidance & Cues Logits LoRA Adaptation + Substitution Tokens Figure 1: Overview of the screening loop. A wav2vec2-based recognizer outputs a production token sequence, which is aligned to the prompted canonical sequence to derive an interpretable screening vector. A caregiver-facing assistant maps this vector to template- grounded feedback with conservative escalation rules. c i only as a suppression signal (e.g., request a repeat recording when c i < τ ); we do not claim calibrated probabilities. For each prompt we define a small, manually curated map- ping p 7→ F(p) that specifies the index/indices of the primary sibilant/affricate target(s) within the canonical token string. For prompts with multiple focus targets, we mark the utterance as a mismatch if any focus target is bracket-labeled in the reference (and analogously for predictions). In our prompt design, most items isolate one primary sibi- lant target, making alignments effectively unambiguous. We therefore use uniform edit costs and perform linguistic error typing after alignment via a deterministic target–realized map- ping. 4. Experimental setup We train on 10,508 utterances and validate on 1,170 utterances, with a speaker-disjoint held-out test set containing 559 utter- ances from 10 unseen children. All splits share the same fixed prompt inventory (words and syllables) by design, reflecting the intended screening setting where a child is asked to repeat a pre- defined set of diagnostic prompts. Consequently, our evaluation primarily measures generalization to unseen speakers rather than open-vocabulary generalization to unseen lexical items. Audio is converted to mono (channel averaging if needed) and resampled to 16 kHz to match the input interface of the SSL encoders used in this work. Recordings are single-word/syllable productions (typically around 1 s). Because prompts are short (words/syllables), token se- quences are also short by design: on the test set, reference strings have mean token length 3.66 (median 4, min 1, max 7). Reporting this distribution helps interpret exact-match metrics for short screening prompts. We train with mixed precision and early stopping (patience 50; max 200 epochs). Training stops for wav2vec2 model with post-encoder layers and bracket tokens, with the best check- point at epoch 54 selected by validation performance (val WER 5.51%, val CER 3.29%). To contextualize the encoder choice, we additionally train a WavLM-Base baseline by replacing only the SSL encoder while keeping the token vocabulary, post- encoder, LoRA configuration, training schedule, and greedy de- coding unchanged (Table 2). This baseline is reported for com- parison and is not used in the proposed screening pipeline. For decoding, we use greedy CTC decoding (argmax over logits) without an external language model in order to avoid lexical priors that could mask minimal-contrast errors [12]. Table 1: Main recognition results on held-out test children (559 utterances). MetricValue Exact sequence match88.7% (496/559) Token accuracy95.0% WER (token string)5.95% CER (token string)4.09% Table 2: Baselines and ablations (held-out test set). ModelExact (%)Token Acc. (%)Screening F1 wav2vec2 + post-enc + bracket tokens 88.795.00.67 wav2vec2 + bracket tokens 84.590.20.62 WavLM + post-enc + bracket tokens 78.686.60.54 5. Results Table 1 summarizes headline recognition performance. Table 2 provides a compact set of baselines/ablations to contextualize the design choices (post-encoder and bracket supervision). We emphasize exact sequence match because a 1:1 match enables stable downstream alignment and error typing without addi- tional heuristics. Validation metrics were comparable to test (val WER 5.51% vs. test WER 5.95%), suggesting limited over- fitting under the current prompt set. Across the 10 test children (prompted with a fixed inven- tory of 64 prompts/items; some recordings may be missing), exact match ranges from 72.6% to 100% (mean 89.5%± 9.4%). Utterance-level bootstrap confidence intervals are: exact match 95% CI [86.1, 91.2] and token accuracy 95% CI [93.8, 96.3]. Because utterances are clustered by child, we additionally re- port a cluster bootstrap by child (resampling children with re- placement and keeping all their utterances).With only 10 test children, these CIs are wider and better reflect speaker- generalization uncertainty: Exact match 95% CI [83.8, 93.2], token accuracy 95% CI [92.7, 97.0]. This quantifies uncertainty given the small held-out cohort. Table 3: Screening proxy confusion matrix on the held-out test set (focus targets only). ˆy = 1 (flag) ˆy = 0 (no flag) y = 1 (ref mismatch)TP = 35FN = 22 y = 0 (ref correct)FP = 13TN = 489 5.1. Screening proxy evaluation (target sibilant mismatch detection) Recognition quality alone does not guarantee screening utility. We therefore evaluate a minimal screening decision focused on the primary sibilant target in each prompt. In our corpus, devi- ations at the target segment are annotated using bracketed ev- idence tokens, so we are using the rule that if there is a flag mismatch we have recognized token at target position that is bracketed. On the held-out test set (the resulting confusion ma- trix can be seen in Table 3 where the child has pronounced particular word), this rule yields 72.9% precision and 61.4% re- call (F1=0.67) for detecting target mismatches, with a low false- alarm rate of 2.7% on target-correct items. When a mismatch is flagged (true positives), the predicted bracket class matches the reference bracket label in 85.7% of cases, supporting inter- pretable error typing. 5.2. Residual error analysis on sibilant units To focus on clinically relevant contrasts, we analyze errors at a sibilant unit level by extracting the 12 symbols (canoni- cal and bracketed) from each token string and aligning these unit sequences with a minimum-edit algorithm. Among the 63 non-exact utterances, 58 (92.1%) contain exactly one unit-level substitution, indicating that most recognition errors are local- ized to a single sibilant decision. Of these, 35/58 (60.3%) are canonical-bracket confusions (one side bracketed), consistent with the role of bracket tokens as graded substitution evidence. 6. Explainable assistant design The recognition and alignment modules provide token-level screening evidence, but caregiver value depends on present- ing it in a safe and interpretable form. The assistant con- sumes the screening vector [target, realized, type, position, confidence] and generates a short caregiver report grounded in clinician-authored templates. Caregiver-facing view. For each prompt, the caregiver sees the word in orthography, a match/likely-mismatch flag (with a re- peat suggestion under low confidence), a plain-language local- ization (e.g., “at the start”), the minimal contrast in orthography (e.g., s vs. sz), and an optional brief practice cue. IPA strings are hidden by default; clinicians can audit the underlying align- ment, token outputs, error type, and confidence. Template grounding and safety. Messages are produced by filling a fixed, reviewable template inventory keyed by type. Templates include conservative escalation rules and refusal con- straints (no diagnosis, no medical claims, no guarantees). Under low confidence (c i < τ ) or inconsistent evidence, the assistant requests re-recording instead of giving detailed advice [16]. Report aggregation. Across prompts, the system can aggregate flags into a compact summary (top recurring contrasts + short practice plan), while keeping the interaction within a screening scope. Table 4: End-to-end example of the screening loop (illustrative test utterance). Promptsznurek Target (IPA)/ ù n u r E k / Recognized (IPA) [s] n u r E k Screening entry target=ù, realized=[s], type=place shift, pos=1, conf=0.91 Caregiver message “At the start of sznurek, it sounded closer to s than sz. Try repeating slowly and aim- ing for a slightly more retracted tongue position than for s. If this pattern per- sists over days/weeks, consult a speech- language pathologist.” 7. Discussion and limitations The results suggest that SSL-based token recognition can pro- vide a practical backbone for caregiver-oriented screening in Polish. By designing the pipeline around prompted contrasts and alignment to the canonical target, we avoid reliance on word-level language-model priors that may mask minimal- contrast errors [12]. The high exact-match rate supports stable downstream alignment; nevertheless, screening performance depends on which deviations are detected versus missed, and on controlling false alarms. Bracketed evidence labels reflect an expert interpretation of a non-canonical realization. Our error analysis shows that many residual errors involve canonical-bracket confusions. This moti- vates future exploration of alternative formulations (e.g., a two- head model that predicts a base phone and a separate deviation flag, or label smoothing between canonical and bracketed coun- terparts) to better reflect the graded nature of these contrasts. Because the intended use case relies on a fixed diagnostic prompt list, we evaluate speaker-disjoint generalization on the same prompt inventory across splits; generalization to unseen lexical items is left for future work. The held-out evaluation covers 10 unseen children and 559 utterances; broader evaluation across ages, severities and recording conditions is needed before deployment. We report per-child variability and bootstrap confidence intervals, and we plan to add age-stratified analyses. While the assistant is template-grounded by design, we do not yet present user studies with caregivers or clinicians. Evalu- ating safety, clarity and risk of misinterpretation (screening vs. diagnosis) is essential future work. 8. Conclusion We presented a phoneme-level mispronunciation screening loop for Polish-speaking children based on wav2vec2 token recog- nition, alignment-based screening, and a template-grounded caregiver-facing assistant. The recognizer achieves 88.7% exact sequence match on unseen children, supporting stable token-to- token alignment. We additionally report a conservative screen- ing proxy evaluation based on bracketed substitution evidence, yielding low false-alarm rates and interpretable error typing. Future work will expand coverage beyond substitutions, vali- date diagnostic utility against clinician outcomes, and evaluate assistant feedback quality and safety in caregiver-facing studies. 9. Acknowledgments This work was supported by the National Centre for Research and Development (NCBR), Poland, under research project No. 0179/L-15/2024, entitled “Speech therapy computer system for recording and analyzing multimodal articulation data using the 4D speaker model and deep learning (SpeechCAD).” This study was approved by the Bioethics Committee for Scientific Re- search at the Jerzy Kukuczka University of Physical Education in Katowice, Poland (Decision No. 3/2021). Written parental consent and child assent were obtained for all participants. 10. Generative AI Use Disclosure Generative AI tools were used to assist with language editing, polishing, and local wording. They were not used to produce a significant part of the manuscript. All authors reviewed, edited, and verified the final manuscript and remain responsible for its content. 11. References [1] R. Waring and R. Knight, “How should children with speech sound disorders be classified? a review and critical evaluation of current classification systems,” International Journal of Language & Communication Disorders, vol. 48, no. 1, p. 25–40, 2013. [2] N. McGill, S. McLeod, K. Crowe, C. Wang, and S. C. Hopf, “Waiting lists and prioritization of children for services: Speech- language pathologists’ perspectives,” Journal of Communication Disorders, vol. 91, p. 106099, 2021. [3] E. Sugden, N. Munro, C. M. Trivette, E. Baker, and A. L. Williams, “Parents’ experiences of completing home practice for speech sound disorders,” Journal of Early Intervention, vol. 41, no. 2, p. 159–181, 2019. [4] M. Zygis and J. Padgett, “A perceptual study of polish fricatives, and its implications for historical sound change,” Journal of Pho- netics, vol. 38, no. 2, p. 207–226, 2010. [5] A. Sage, Z. Miodonska, M. Krecichwost, and P. Badura, “Hy- bridization of acoustic and visual features of polish sibilants produced by children for computer speech diagnosis,” Sensors, vol. 24, no. 16, p. 5360, 2024. [6] M. Krecichwost, Z. Miodonska, A. Sage, J. Trzaskalik, E. Kwas- niok, and P. Badura, “Pavsig: Polish multichannel audio-visual child speech dataset with double-expert sigmatism diagnosis,” Scientific Data, vol. 12, no. 1612, 2025. [7] M. Shahin, J. Epps, and B. Ahmed, “Phonological level wav2vec2-based mispronunciation detection and diagnosis method,” Speech Communication, vol. 173, no. 103249, 2025. [8] M. Lounis, B. Dendani, and H. Bahi, “Mispronunciation detection and diagnosis using deep neural networks: a systematic review,” Multimedia Tools and Applications, vol. 83, p. 62 793–62 827, 2024. [9] S. Dudy, S. Bedrick, M. Asgari, and A. Kain, “Automatic analy- sis of pronunciations for children with speech sound disorders,” Computer Speech & Language, vol. 50, p. 62–84, 2018. [10] T. Ahn, Y. Hong, Y. Im, D. H. Kim, D. Kang, J. W. Jeong, J. W. Kim, M. J. Kim, A.-R. Cho, H. Nam, and D.-H. Jang, “Auto- matic speech recognition (asr) for the diagnosis of pronunciation of speech sound disorders in korean children,” Clinical Linguis- tics & Phonetics, vol. 39, no. 10, p. 913–926, 2025. [11] T. Patel and O. Scharenborg, “Improving end-to-end models for children’s speech recognition,” Applied Sciences, vol. 14, no. 6, p. 2353, 2024. [12] A. Radford, J. W. Kim, T. Xu, G. Brockman, C. McLeavey, and I. Sutskever, “Robust speech recognition via large- scale weak supervision,” 2022. [Online]. Available:https: //arxiv.org/abs/2212.04356 [13] A. Baevski, H. Zhou, A. Mohamed, and M. Auli, “wav2vec 2.0:A framework for self-supervised learning of speech representations,” 2020. [Online]. Available: https://arxiv.org/abs/ 2006.11477 [14] W.-N. Hsu, B. Bolte, Y.-H. H. Tsai, K. Lakhotia, R. Salakhut- dinov, and A. Mohamed, “Hubert: Self-supervised speech rep- resentation learning by masked prediction of hidden units,” IEEE/ACM Transactions on Audio, Speech, and Language Pro- cessing, vol. 29, p. 3451–3460, 2021. [15] L. Peng, Y. Gao, R. Bao, Y. Li, and J. Zhang, “End-to-end mispro- nunciation detection and diagnosis using transfer learning,” Ap- plied Sciences, vol. 13, no. 11, p. 6793, 2023. [16] N. Balaji Shankar, K. Zhang, A. Mai, M. Shi, A. Long, J. Wash- ington, R. Morris, and A. Alwan, “Leveraging asr and llms for automated scoring and feedback in children’s spoken language assessments,” in 10th Workshop on Speech and Language Tech- nology in Education, 2025, p. 1–5. [17] A. Graves, S. Fernández, F. Gomez, and J. Schmidhuber, “Con- nectionist temporal classification: Labelling unsegmented se- quence data with recurrent neural networks,” in Proceedings of the 23rd International Conference on Machine Learning (ICML), 2006, p. 369–376. [18] E. J. Hu, Y. Shen, P. Wallis, Z. Allen-Zhu, Y. Li, S. Wang, and W. Chen, “Lora: Low-rank adaptation of large language models,” in International Conference on Learning Representations (ICLR), 2022. [Online]. Available: https://arxiv.org/abs/2106.09685