acceptodds
Under review as a conference paper at ICLR 2027

PronCoach: Perceive, Reason, and Correct Pronunciation

21 views

Abstract

Pronunciation correction aims to improve speech clarity through realistic pronunciation recognition and targeted pronunciation guidance. Reliable pronunciation coaching faces three challenges: (1) languages represent pronunciation differently; (2) fine-grained pronunciation-correction data and phone-level annotations are scarce, while ambiguity in phone perception leads to unstable expert annotations; and (3) feedback must account for weak forms, tone changes, accent variation, and other context-dependent phenomena. We present PronCoach, a pronunciation coaching framework for Mandarin and English audio without reference transcripts. We decompose coaching into a structured chain-of-thought process covering intended-content recovery, realized-pronunciation inference, contextual comparison, per-phone correction, intervention decisions, and feedback generation. Based on this decomposition, we fine-tune language-specific phonetic-perception models on expert annotations to infer probabilistic candidate realizations from acoustic and canonical-pronunciation evidence. An audio-language model (ALM) then uses the recording, intended content, and candidate evidence to generate aligned diagnoses and coaching feedback. These annotations supervise both complete coaching responses and intermediate tasks. To address the lack of a dedicated pronunciation-correction benchmark, we construct PronCoach-Bench with two complementary tracks. The Diagnostic Track evaluates correction localization, heard-realization agreement, and false corrections, jointly testing phonetic perception and correction decisions; the Clean-Speech Track tests correction restraint by measuring whether models avoid unnecessary feedback. The main PronCoach model achieves a heard-based overall F1 of 0.7528, compared with 0.4139 for the strongest evaluated public ALM baseline. Supplementary English evaluations yield the lowest phoneme error rate among the compared systems on two of three corpora. Furthermore, PronCoach demonstrates emergent cross-lingual correction capabilities, enabling it to perform pronunciation rectification across Mandarin and English representations. PronCoach-Bench will be publicly released soon.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Pronunciation Correction, Speech Recognition, Audio Language Model

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_PqMVOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces PROCOACH, an audio-language model (ALM) framework for Mandarin and English pronunciation coaching. PRONCOACH operates directly on learner audio without a reference transcript to recover intended lexical content, infer realized pronunciation, account for acceptable contextual variations, localize pronunciation corrections, and generate feedback.

The authors described a pipeline to generate the training reasoning data for the ALM, involving intended-content recovery, realized phoneme recognition, grammar diagnosis, contextual comparison, and per-phone correction, in which several teacher ALM and supervised models are used to generate the labels. Then the Qwen3.5-Omni-based ALM was trained using supervised fine-tuning (SFT) and Reinforcement Learning with GRPO. Two variants "Thinking" and "Instruct" are investigated.

Furthermore, a pronunciation coaching benchmark PRONCOACH-BENCH is propsoed, which includes Diagnostic track and Clean-speech track. Empirical results show that the proposed model outperforms several leading ALMs.

Critical strengths
  1. Well-Motivated problem: Moving from transcript-dependent phone recognition/scoring to reference-free, context-aware pronunciation coaching addresses a genuine gap.
  2. The construction of the CoT data looks solid and reasonable: It involves several teacher ALM and strong supervised models to annotate a larger speech pool.
  3. The results look strong. It outperforms several ALMs on the proposed benchmark and phoneme recognition and MDD.
Critical weaknesses
  1. I think many important details are missing, especially in the CoT data construction pipeline. For example, for realized phoneme recognition, it is mentioned that several human annotators have annotated the realized phoneme. Why is the fine-tuned HuPER still needed? How reliable is the phoneme recognizer? What is it PER on non-native speech? What is the size of the training data and how is this model trained? I think many details about this model are missing.
  2. For HuPER fine-tuning, it is mentioned that "We fine-tune language-specific HuPER (Guo et al., 2026) models on expert annotations to extend pronunciation coverage and estimate candidate probabilities. Both combine canonical pronunciation with acoustic evidence from WavLM for Mandarin (Chen et al., 2022) and HuBERT for English (Hsu et al., 2021)." It is still unclear to me how exactly this architecture looks like. How does it compare to the original HuPER? This could an important technical contribution if others want to reproduce this realized phoneme recognizer.
  3. The F1hearF1^{\text{hear}} metrics looks unintuitive to me. It is defined as the harmonic mean of Recall and Acc. Acc is further defined as "predicted–reference realized-pronunciation agreement at correctly localized positions", which is a bit perplexing to me. This seems to be totally ignoring the false positives, i.e. the model could claim many unrealized phonemes and it will not get penalized.
Questions for authors
  1. Per my understanding, the model internalizes the alignment between canonical and realized phonemes in its output. i.e. the output directly writes out the phoneme pairs that are mispronounced, and there is no need to do an explicit alignment between the canonical phonemes and realized phonemes and then compare. In fact, for the claimed transcript-free assessment case, the GT canonical phonemes are unknown. How do you define "correctly localized positions"?
  2. What is the motivation of the RL stage? Wouldn't the SFT model trained on the CoT data do the same job? Maybe a comparison between SFT vs. SFT+RL could be informative here.
5 comments
AuthorsOpenReview ↗

Response to Reviewer PqMV [1/3]

Thank you for your insightful suggestions and comments. We have revised the manuscript to address your concerns, and we provide detailed responses below.

W1a Why fine-tune HuPER?

Because the numbers of recordings they cover differ by an order of magnitude. Expert character-by-character annotations cover 16,811 Mandarin recordings and 9,900 English recordings; the speech pool used to construct the coaching data is much larger. HuPER transfers the probability-distribution-aware perception learned from expert annotations to audio without fine-grained annotations. Using multiple experts for annotation is less efficient and more expensive. The HuPER model can better represent continuous probability distributions rather than discrete vote counts alone. Over continuous probability distributions, the model can better "preserve the distribution." What is passed downstream is not HuPER's single best solution, but the candidate probabilities for each slot. The probabilities for the three Mandarin onset, rime, and tone slots are calibrated (ECE: 0.0094, 0.0076, and 0.0302, respectively), so the "dispersion of votes" itself carries information: when the ALM receives a tone distribution in which each of three candidates has a share of the probability, it knows that the character's tone is inherently ambiguous, rather than seeing a single discrete argmax. This information is lost if the distribution is compressed to an argmax before being passed downstream.

W1b Recognizer reliability

We sampled an agreement test set: 700 recordings covering seven accents, each independently transcribed by two annotators and adjudicated by one reviewer.

Calculated as per-utterance phone edit distance divided by the reference length:

SetValid recordingscorrectorDirect dictionary-form reading
All6800.16700.1871
Untrained stratum1590.19280.2017

The PER between the two annotators is 0.1447. In other words, when the same accented recording is transcribed by two trained people, their transcriptions differ by approximately 14%; the recognizer differs by 19% on recordings it has not been trained on.

The Mandarin component outputs character-level classification over three slots—onset, rime, and tone—rather than an IPA phone sequence. AUROC is 0.970 for onset, 0.952 for rime, and 0.782 for tone; deviation-detection recall is 0.664 / 0.553 / 0.079.

HuPER's tone predictions tend to be conservative. Across all Mandarin annotations, the three annotators disagree at only 4.7% of onset and rime positions, compared with 30.5% for tone.

W1c Training scale and training procedure

Mandarin branch: 16,811 training and 881 test supervision records, corresponding to 244,171 / 12,812 Chinese characters, or 732,513 / 38,436 onset–rime–tone slots. Training uses 8 GPUs, batch 8 per GPU, gradient accumulation 2, and at most 12 epochs. The classification-head learning rate is 6e-4 and the encoder learning rate is 1e-5. Checkpoint selection and early stopping use macro-AUROC, with patience 4; ECE is examined separately for calibration.

English branch: 9,900 / 521 supervision records, 2,256 / 128 distinct recording IDs, and 284,112 / 15,252 canonical phones. HuBERT is frozen; only the interaction layers and edit heads are trained. We use the weights after epoch 7.

AuthorsOpenReview ↗

Response to Reviewer PqMV [2/3]

W2 Fine-tuned HuPER architecture

We use different models for Mandarin and English, rather than the same architecture with different data. The canonical pronunciation supplies "what should be pronounced at this position" as the query, the acoustic side supplies evidence from the recording through cross-attention, and the output heads predict the realized pronunciation or edit operations.

MandarinEnglish
Acoustic encoderWavLM-large, trainedHuBERT-large, frozen
canonical querySum of onset/rime/tone slot embeddings, learned positional encodingPhone embedding, sinusoidal positional encoding
Acoustic interaction4-layer, 8-head TransformerDecoder8 layers, 8 heads, self + cross attention
Output and supervisionThree linear heads, weighted soft CE against expert vote distributionsTwo heads for KEEP/DEL/SUB and insertion, hard CE

Both use d=512, FFN 2048, and dropout 0.1. The Mandarin model additionally uses canonical-dropout 0.3: during training, the query is randomly replaced with a learned MASK, forcing the model to listen to the audio rather than copy the canonical pronunciation. The deviation weight is 2, agreement gamma is 0, and all three slot weights are 1.

Compared with the original HuPER (in which the DRRC module is responsible for "recognizing the realized pronunciation from audio and comparing it with the canonical pronunciation to identify deviations"), we made four changes: for Mandarin, we added an output structure with onset, rime, and tone slots and changed the loss from single-label supervision to fitting expert vote distributions, which is a prerequisite for retaining candidate probabilities; for English, we predict edit operations instead of directly emitting a phone sequence.

We plan to open-source the relevant realized-pronunciation model code after acceptance to help the open-source community reproduce the realized-pronunciation recognizer in this area.

W3 F1_hear

F1_hear is the harmonic mean of recall and ACC_hear, while ACC_hear is calculated only at correctly localized positions. Therefore, the number of additional corrections the model makes at positions that should not be corrected does not affect ACC_hear. We aim to evaluate these aspects separately, using the clean track to assess the model's ability to avoid unnecessary corrections. If the model corrects phones that were not actually produced, it is penalized on the clean track, as we report for Acc_c.

We will include precision in the revised manuscript: a hit is counted only when all three conditions are met—the correction is localized to a reference correction position, target is a valid canonical pronunciation at that position, and the first heard belongs to the set of valid realized pronunciations at that position.

$ P=\frac{TP}{N_p},\qquad R=\frac{TP}{N_g},\qquad F_1=\frac{2,TP}{N_g+N_p} $

For the main model in Mandarin: Np 1,634, TP 776, FP 858, and FN 542, giving P 0.4749, R 0.5888, and F1 0.5257. English F1 is 0.3277. On the Clean track, the results are 36/81 for Mandarin and 55/97 for English. Several internal configurations under the same definition are: Discrete, Mandarin 0.2531 (Clean 77/81), and Scale, 0.6786 (Clean 28/81). Discrete has the highest precision and the lowest recall, while Scale is high on both, consistent with the effect of data upscaling. A future direction may be to explore how to improve the model's ability to avoid unnecessary corrections while maintaining high recall.

AuthorsOpenReview ↗

Response to Reviewer PqMV [3/3]

Q1 Text alignment

The main model receives learner audio and a coaching request, recovers the content itself, and directly writes "which character was mispronounced, what it should sound like, and how it was actually pronounced." There is no explicit alignment between canonical and realized sequences at inference.

Reference-free describes the information supplied to the evaluated system, not the absence of references for the offline evaluator. Main writes its intended text and pronunciation diagnoses directly. Data construction uses explicit phonetic alignment to prepare supervision; deployment produces the answer; offline scoring localizes the stated corrections against a fixed reference transcript. These are three different procedures.

The final PronCoach-Bench includes reference text, but this text is not supplied at model inference. This also tests the large model's basic semantic reasoning and contextual understanding. Phone alignment is performed only after matching the text it generates.

The submitted localization procedure matches normalized text units and target pronunciation, with a surface fallback and deterministic handling of repeated units. Corrections that cannot be placed are recorded separately. Reference text and correction labels are not supplied to Main.

There is a point that is easy to confuse and that we did not previously explain clearly: reference-free constrains the input the user supplies to the model, not the evaluator. The model has no access to the reference transcript. The evaluator has reference text and compares the manually transcribed text with the model-predicted text to associate corrections with individual reference positions. During scoring, the text unit in a model report is located in the reference transcript, and a frozen monotonic search then makes a one-to-one assignment, ensuring that reference positions correspond uniquely to the text understood by the large model.

Q2 Motivation for the RL stage and comparison between SFT and SFT+RL

The initial motivation for RL training was to address two issues. The first was misaligned corrections. During SFT-only training, we found the following problem in the model trained on instruct data: it could produce almost correct results on the separate realized-pronunciation and target-pronunciation tasks, yet exhibited severe missed corrections or excessive corrections in the coach task—it instead made corrections at positions where the realized and target pronunciations were the same, or missed corrections at positions where the preceding two separate tasks produced different pronunciations. We attributed this to a conflict, for the model, between realized and target pronunciations in the coach task. We therefore hoped to address this through a reward that compares the two paths and rewards recall and canonical. However, during subsequent parallel development, we found that CoT data could come close to resolving this issue effectively, and that the capabilities learned from CoT data could transfer fairly well to instruct tasks. We therefore did not further design rewards to enhance this capability through RL. The second issue was json-format errors. We found occasional json parse errors in the coach_json task, so we added a format parse reward at the RL stage. These were our two initial motivations for the RL stage.

We added a comparison experiment to demonstrate the role of RL. As can be seen, with instruct-only data without RL, the metrics of correction correctness, P_loc and ACC_hear, do not decrease much, but R and FC decrease substantially. The model makes many unnecessary corrections, substantially worsening its performance.

ModelLanguageR ↑P_loc ↑F1_loc ↑ACC_hear ↑F1_hear ↑FC ↓ACC_c ↑Zero correctionsOverall ↑
PronCoachMandarin0.69800.59130.64020.95000.80480.25230.345736/810.7142
English0.49670.40540.44640.80000.61290.07850.567055/97
w/o GRPOMandarin0.54470.56950.55680.91870.68390.58420.04944/810.5865
English0.43050.23640.30520.38460.40630.14420.402139/97
Reviewer_VKjdOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents PronCoach, an audio-language model framework for reference-free pronunciation coaching in Mandarin and English. Instead of just outputting phoneme sequences or numeric assessment scores, the system aims to recover the intended transcript from learner audio, identify both grammar and pronunciation issues while accounting for acceptable contextual variants (e.g., weak forms, tone sandhi), and generate coaching explanations with practice plans. To build training data, the authors combine expert phonetic annotations, fine-tuned HuPER phonetic perception models (which provide candidate pronunciation probabilities), rule-based filters, and a teacher model (Gemini 3.1 Pro Preview) to synthesize 5-stage reasoning traces ("Thinking") and direct feedback ("Instruct"). They fine-tune Qwen3.5-Omni on 8 subtasks via SFT, followed by GRPO using an F1.2F_{1.2}-based diagnosis reward. They also introduce PronCoach-Bench (510 diagnostic recordings and 178 clean-speech recordings) and show that PronCoach outperforms general-purpose ALMs and open-source MDD baselines.

Critical strengths
  • The practical motivation is clear. Standard MDD systems typically require a ground-truth reference text and flag every acoustic deviation from dictionary canonical forms, which is unhelpful when learners make lexical/grammar mistakes or use natural connected-speech reductions. Formulating the task as reference-free, context-aware coaching is a practical direction for computer-assisted pronunciation training (CAPT).
  • Building a bilingual dataset that pairs audio with multi-expert phonetic annotations, intended vs. realized alignments, and structured coaching explanations takes substantial engineering and annotation effort. Including a Clean-Speech track in the benchmark to test whether models over-correct standard speech is also a good design choice.
  • On English phoneme recognition (Table 2), extracting the realized phoneme sequences from PronCoach's coaching outputs achieves competitive PER on L2-ARCTIC and noticeably lower PER than prior open-source phonetic models on SpeechOcean762 and EpaDB.
Critical weaknesses
  • The work reads primarily as a domain-specific data construction and rule-engineering pipeline for CAPT rather than a methodological contribution suited for ICLR. Much of the system relies on stitching together existing components (Gemini 3.1 Pro Preview, pypinyin/espeak-ng, WavLM/HuBERT-based HuPER, SFT, and standard GRPO) alongside heavily hand-crafted linguistic heuristics, such as hardcoded word exclusion lists for Mandarin particles/prepositions and English function words, specific 4-letter short-vowel exemptions, and fixed thresholds like requiring the leading expert weight to be below 70 and the second weight at least 20 (Appendices A.2 and C.1).
  • The headline metric F1_all (hear) in Table 1 is formulated in a way that masks severe over-correction in Mandarin. According to Eq. 8 in Appendix C.5.2, F1_hear is the harmonic mean of recall (R) and conditional realized-pronunciation agreement on hits (ACC_hear), completely omitting localization precision (P_loc). Looking at Table 8 in Appendix D, PronCoach's Mandarin localization precision P_loc is only 0.1930 (meaning ~80% of its localized error flags are false positives), and its Mandarin Clean-Speech accuracy (ACC_c) is only 0.1728 (it falsely flags errors on over 82% of standard broadcasting Mandarin clips). Because F1_all (hear) ignores P_loc, aggressively over-predicting errors boosts recall (0.8187) and inflates the overall headline score to 0.7528.
  • The ablation in Table 3 actually contradicts the paper's central claim regarding probabilistic phonetic targets in Mandarin. In Mandarin (the only language where expert annotations are kept as probability distributions rather than adjudicated single sequences), replacing probabilistic targets with Discrete targets improves localization F1_loc from 0.3124 to 0.6348, reduces false-correction density (FC) from 0.3501 to 0.0234, and boosts Clean-Speech accuracy (ACC_c) from 0.1728 to 0.5309. The Main model only wins on F1_all (hear) because of the recall-heavy metric issue noted above.
  • There is a strong teacher and rule bias in the benchmark evaluation. In PronCoach-Bench, the reference correction targets are generated by Gemini 3.1 Pro Preview and filtered by custom rule sets (Section 5 and Appendix C.1)—the exact same teacher model and filtering rules used to construct PronCoach's training data. Zero-shot public ALMs have no knowledge of these dataset-specific exclusion rules (e.g., which particles or voicing substitutions are ignored, or the cap of 3–5 errors per utterance), so they are naturally penalized on localization precision and FC.
  • Several key components and claims are left unevaluated. First, Section 4 describes a GRPO training stage, but there is no ablation comparing SFT-only vs. SFT+GRPO, making it unclear what GRPO contributes or whether the recall-biased F1.2 reward in GRPO is what caused the high false-alarm rate in Mandarin. Second, the abstract and introduction highlight "emergent cross-lingual correction capabilities", multi-turn coaching, grammar correction, and actionable practice plans, yet none of these are evaluated in Section 6 (which only evaluates single-turn position/phone matching). Finally, Appendix C.7 describes an "Auxiliary HuPER Perception Protocol" on 300 Mandarin recordings, but the actual results for this experiment appear to be missing from the paper.
Questions for authors
  1. In Table 8, PronCoach has a localization precision (PlocP_{\text{loc}}) of 0.1930 and Clean-Speech accuracy of 0.1728 in Mandarin, whereas the Discrete targets variant in Table 3 achieves F1loc=0.6348F1_{\text{loc}} = 0.6348, FC=0.0234\text{FC} = 0.0234, and ACCc=0.5309\text{ACC}_c = 0.5309. Why does using probabilistic supervision cause such a severe drop in precision and clean-speech restraint in Mandarin?
  2. What is the performance of the SFT-only checkpoint before applying GRPO? Does the recall-weighted F1.2F_{1.2} reward in GRPO contribute to the high false-correction rate on Mandarin speech?
  3. Why was localization precision (PlocP_{\text{loc}}) excluded from the definition of F1hearF1_{\text{hear}} and the overall score F1allhearF1_{\text{all}}^{\text{hear}}? A model that flags almost every word as an error can achieve high recall RR and high ACChear\text{ACC}_{\text{hear}} on matched positions while having very poor precision.
  4. Where are the quantitative or qualitative evaluations for (a) the "emergent cross-lingual correction capabilities" claimed in the abstract, (b) grammar error detection and coaching explanation quality, and (c) the 300-recording auxiliary HuPER evaluation described in Appendix C.7?
4 comments
AuthorsOpenReview ↗

Response to Reviewer VKjd [1/4]

W1 Contribution

All of the rules you mention apply only during annotation, not as evaluation criteria.

The 70/20 threshold mentioned in the appendix. This is the trigger for retaining disagreement evidence when constructing Mandarin supervision: candidate weights are normalized to a 100-point scale, and when the first choice is below 70 and the second choice is at least 20, the disagreement at that position is left for the teacher to judge using the audio and context. When constructing the dataset, we found that without explicitly specifying the numbers, Gemini tended to annotate conservatively and omit most annotations, retaining only very high-confidence probabilistic corrections. In fact, a softmax probability of 20 already represents a relatively evident deviation, so we added hard requirements to ensure that there would be no obvious omissions. This is not a hard bar, but a prompt requirement for ensuring correction recall.

This and the benchmark's ignore filtering belong to two different stages. Given the relatively conservative huper, it is necessary to specify a threshold to ensure that gemini does not ignore pronunciations with seemingly low probabilities.

The four-letter short-vowel exemption applies only to English. It is intended to avoid unnecessary corrections caused by large numbers of similar-vowel substitutions, such as the frequently occurring words of, at, and the. The exemption we set is limited to similar vowel pronunciations that do not affect comprehension, rather than obvious mispronunciations, which will still be corrected normally. This exemption is also based on training results. Without the exemption, the model learns large numbers of these short-vowel substitutions and persistently corrects various minor, near-equivalent vowel errors in the final bench, deviating from the original intent of a large model for genuine semantic and pronunciation correction. The specific exemption therefore prevents deterioration in model robustness. The final model's ability to correct short-vowel errors is also demonstrated in PronCoach: its ability to correct vowel substitutions in long words generalizes well to short words, without overcorrection.

Exclusion list Excluding unintentional discourse particles and grammatical particles, as well as erhua, is an important part of the Mandarin work. Discourse particles vary greatly across contexts and speakers' pronunciations. A reasonable pronunciation-correction model should not focus on these discourse particles. In Mandarin, “的”“地”“得” have multiple pronunciations when used as grammatical particles; this is not an important distinction between standard pronunciation and accent. Similar to the short-vowel exemption described above, the pronunciations of grammatical particles and erhua are not primary correction targets when they do not affect semantic comprehension. The grammatical-particle corrections observed in PronCoach-Bench are also largely covered through generalization. For mispronounced particles that affect semantics, our prompt for the teacher also calls for correction; this is not a hard exclusion list.

Methodological contribution.

We have focused on a pronunciation-correction problem that ALM and Omni-Model had not previously focused on, and conducted experiments in two representative languages. Using the annotated data, we trained a high-accuracy G2P HuPER model for data expansion, proposed an innovative method for identifying pronunciation errors from pronunciation probabilities, and constructed a complete data pipeline and a basic domain-specific baseline model. SFT and GRPO are general-purpose methods in the large-model field; using these methods should not be regarded as lacking innovation. On the contrary, the problem in the domain we study can be addressed well using our proposed method without changing the large-model architecture or training procedure.

AuthorsOpenReview ↗

Response to Reviewer VKjd [2/4]

W2 F1_all(hear) and overcorrection in Mandarin

Thank you for pointing out the design of F1_hear. F1_hear is the harmonic mean of recall and conditional hearing agreement and indeed does not include localization precision. Our initial intention was to have extra corrections reflected in the Clean-track, while overcorrection in Diagnose track sentences could be reflected through P_loc.

We will change the primary metric for the diagnose track to F1_loc, which includes precision, and move P_loc from the appendix to the main table in the main text.

ModelLanguageR ↑P_loc ↑F1_loc ↑ACC_hear ↑F1_hear ↑FC ↓FP ↓N_pjoint P ↑joint R ↑jointF1 ↑ACC_c ↑Overall F1_hear_all ↑
PronCoachMandarin0.69800.59130.64020.95000.80480.25237071,6340.47490.58880.52570.44440.7142
English0.49670.40540.44640.80000.61290.07851282030.28570.38410.32770.5670
Ablation
w/ Scale dataMandarin0.77620.78690.78150.97560.86450.11883331,3580.66860.68890.67860.34570.7060
English0.67550.30630.42150.83330.74620.16562703810.20730.52320.29700.2371
w/ Discrete targetsMandarin0.16160.86590.27240.95770.27660.0200562700.74440.15250.25310.95060.4283
English0.44370.36020.39760.52240.47980.08771432140.15420.21850.18080.6804
Baselines
Doubao-Seed-2.0-LiteMandarin0.22380.40970.28950.34240.27070.25167051,0080.09920.07590.08600.03700.2056
English0.44370.23180.30450.16420.23970.18963093830.02610.06620.03750.0206
Gemini 3.8 FlashMandarin0.20640.36760.26430.44120.28120.40191,1261,4210.08590.09260.08910.02470.2843
English0.31790.16720.21920.08330.13200.26874384990.00800.02650.01230.5979
Gemini 3.1 Pro PreviewMandarin0.22760.31580.26460.36000.27890.44861,2571,6180.06610.08120.07290.02470.3208
English0.19210.11280.14220.03450.05850.52158509220.00110.00660.00190.9072
Muse-Spark-1.2Mandarin0.20560.28800.23990.23620.21980.70811,9842,3120.02980.05240.03800.07410.2145
English0.58280.09380.16160.00000.00000.65711,0711,2180.00000.00000.00000.4227

W3 Probabilistic phoneme targets

We asked pronunciation experts to manually review the annotations of PronCoach-Bench, resolving some missing and incorrect annotations in the original ALM-assisted references. The review records cover all recordings in the Diagnose track and the Clean track.

For Mandarin, we added an expert review to the original annotations from three experts. The 296 Mandarin recordings contain 4,120 evaluable positions: 1,308 positions have pronunciation disagreements and 2,812 have agreement. The reviewed correction references cover 1,027 and 2,291 of these positions, respectively. ALM missed some real errors, which have now been addressed through human expert review. For the English review, we added an extra checking expert and resolved disagreements, adjusting the number of corrections in English sentences from 176 to 151.

After the manual review of PronCoach-Bench, Mandarin F1_loc is 0.6402 for the main model and 0.2724 for Discrete; the strict joint scores are 0.5257 versus 0.2531. On the new references, Discrete has a precision of 0.8659 and recall of 0.1616. It reports only 246 positions, so it appeared relatively precise only on the old references with many missed corrections.

ConfigurationSupervision targetSingle candidate (291)Multiple candidates (1,027)DifferenceOverall recall
PronCoachProbability distribution0.5980.726+0.1280.698
ScaleProbability distribution0.5090.852+0.3430.776
Discretearg-max single label0.1620.162+0.0000.162
AuthorsOpenReview ↗

Response to Reviewer VKjd [3/4]

W4 Teacher-model and rule biases

We completely isolate PronCoach-Bench from the teacher model and instead use human-reviewed reference corrections and diagnoses. There is also no rule bias: the annotation instructions we give to experts mention that they only need to annotate pronunciation errors that affect semantic comprehension. The ALM rules do not affect expert annotations. In the expert annotations for PronCoach-Bench, Mandarin corrections for “的”“一” are present, and PronCoach also makes these corrections during correction inference. In other words, these rules only regulate the training targets, while our model generalizes well to correction during inference.

Similarly, the prompt we use during inference contains no filtering rules and imposes no limit on the number of corrections. After full human annotation, the corrections produced by the various baseline models under natural-language prompts constitute a fair Bench, without rule bias.

In the Mandarin Diagnose track, there are disagreements at a total of 1,308 positions. The old ALM-assisted references recorded only 364 of them (27.8%); human review increased this to 1,027 (78.5%). Of the 979 newly added entries, 826 concern tones. HuPER itself is relatively conservative in tone prediction, and further filtering by ALM may have amplified this issue.

W5 & Q4 Key components claims

(a) GRPO.

The initial motivation for RL training was to resolve two issues. The first was misaligned corrections. During our SFT-only training, we found that the model trained on instruct data exhibited the following issue: it could produce almost correct results in the separate realized-pronunciation and target-pronunciation tasks, but showed severe missed corrections or extra corrections in the coach task—at positions where realized pronunciation and target pronunciation were the same, the model instead made corrections, or it missed corrections at positions where the preceding two separate tasks produced different pronunciations. We believed that this occurred because realized pronunciation and target pronunciation conflicted for the model in the coach task. We therefore hoped to resolve it using a reward that compares the two paths and rewards recall and canonical. However, as subsequent work proceeded in parallel, we found that CoT data could largely resolve this issue well, and that the capabilities from CoT data could transfer relatively well to the instruct task. We therefore did not further design a reward to strengthen this capability through RL. The second issue was json format errors. We found that json parse errors occurred from time to time in the coach_json task, so we added a format parse reward during RL. These were our two initial motivations for the RL stage.

We added a comparison experiment to demonstrate the role of RL. As shown, without RL and with only instruct data, the metrics for correct corrections, P_loc and ACC_hear, did not decrease much, but R and FC decreased substantially. The model made many false corrections, leading to much poorer performance.

ModelLanguageR ↑P_loc ↑F1_loc ↑ACC_hear ↑F1_hear ↑FC ↓ACC_c ↑Zero correctionsOverall ↑
PronCoachMandarin0.69800.59130.64020.95000.80480.25230.345736/810.7142
English0.49670.40540.44640.80000.61290.07850.567055/97
w/o GRPOMandarin0.54470.56950.55680.91870.68390.58420.04944/810.5865
English0.43050.23640.30520.38460.40630.14420.402139/97

(b) Cross-lingual, multi-turn, grammar, and practice plans.

The cross-lingual correction capability that emerges across languages is shown in Figure1. Our model can exhibit an emergent ability to correct English pronunciation using Chinese characters and to correct English pronunciation using the pronunciations of similar Chinese characters. For example:

“Pronunciation tips (with the aid of Chinese homophones):

伊特 沃兹 (wəz) 阿 奈斯 戴。 爱 温特 (wɛntə) 泽 帕克 安德 普雷德 (pleɪd) 威兹 泽 多格斯。”

For multi-turn data, we used CosyVoice3 to control and construct some multi-turn correction data for mixed training; there are also sentences formed by concatenating single turns. We used the Qwen-tts work to control the transfer of accent, timbre, and pronunciation. Because current open-source large models do not have explicit corrective practice plans, we did not evaluate practice plans, grammar correction, or correction performance on multi-turn data in PronCoach-Bench. This is an important direction for future work: human subjective evaluation can assess the efficiency and effectiveness of practice plans and multi-turn correction.

AuthorsOpenReview ↗

Response to Reviewer VKjd [4/4]

(c) Appendix C.7: the 300 entries.

ComponentTP / FP / FN / TNF1
Onset110 / 57 / 43 / 40980.6875
Rime72 / 34 / 22 / 41800.7200
Tone14 / 43 / 277 / 39740.0805

There is a specific reason for the low tone recall: tone slots have the most concentrated disagreements among the three-person panel, and a single adjudicated label itself is the least reliable. This is also consistent with real Mandarin accents: actual tone pronunciations are usually not fully formed, but are a probabilistic mixture of 2 or more tones. Confusion between neutral tone and the first and fourth tones is also an important reason for this disagreement. Nonstandard accented pronunciation also tends to concentrate on tones rather than complete onset/rime errors.

Q1 Probabilistic supervision

We resolved this concern after re-annotation by human experts.

On PronCoach-Bench with our reviewed annotations, we found that missing reference annotations were the main cause of the Mandarin F1_loc result. We re-evaluated the same checkpoint and obtained 0.6402 (R 0.6980, P_loc 0.5913, FC 0.2523).

For Clean 0.1728, our model does indeed still make a small number of corrections at some positions in standard recordings. Our reported comparison of FCR@1, FCR@2, and the average number of corrections can demonstrate this capability relatively well. Our model also has an advantage over baselines; it may still have room to improve correction restraint on the task overall. Data containing no errors account for a relatively small proportion of our training data, approximately 15%, so other data mixtures may better improve Full Clean capability. However, our model makes fewer corrections overall. In the experiment increasing the Full Clean proportion to 20%, our model achieves Clean ACC_c=0.4444.

PronCoach has a median of 1 correction and a mean of 1.51 corrections in Mandarin; 76.5% of recordings receive at most 1 correction, rather than indiscriminate correction of the entire sentence. The baselines are worse than every configuration of ours on this metric, and the gaps in the Clean column are order-of-magnitude differences.

SystemACC_cZero correctionsFCR@1FCR@2MedianMean
PronCoach0.444436/810.76540.814811.51
Doubao-Seed-2.0-Lite0.01231/810.01230.086443.88
Qwen3.5-Omni0.345728/810.50610.728332.11
Gemini 3.8 Flash0.03703/810.09880.185244.35

Q2 SFT-only and GRPO

We added a comparison experiment to demonstrate the role of RL. As shown, without RL and with only instruct data, the metrics for correct corrections, P_loc and ACC_hear, did not decrease much, but R and FC decreased substantially. The model made many false corrections, leading to much poorer performance.

Init ModelLanguageR ↑P_loc ↑F1_loc ↑ACC_hear ↑F1_hear ↑FC ↓ACC_c ↑Zero correctionsOverall ↑
PronCoachMandarin0.68970.60240.64310.94940.79900.23660.345728/810.7397
English0.51660.43820.47420.74360.60960.07360.546453/97
Qwen3.5-OmniMandarin0.54470.56950.55680.91870.68390.58420.04944/810.5865
English0.43050.23640.30520.38460.40630.14420.402139/97

The F1.2 reward did not cause a high false-correction rate. It only makes the model align heard and target at the correct positions and perform corrections correctly; this is demonstrated by its capability on the clean bench.

Q3 F1_hear

The design goal of F1_hear is to measure "whether it hears correctly on the errors it has already covered." We hope to leave the measurement of extra corrections to the Clean track. We will change to F1_loc, which includes precision, plus joint F1, and list false-positive counts directly.

In actual examples, we observe that Muse is such a mode. Muse corrects every word, and the corresponding F1 score is consequently poorer after weighting diagnose and clean. We will move the supplementary results on the model's P_loc performance from the appendix into the main results table. These supplementary metrics measure whole-utterance perception.

Reviewer_8FPJOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents PronCoach, an audio-language model-based framework for pronunciation correction and coaching in Mandarin and English without requiring reference transcripts. Unlike conventional pronunciation assessment and mispronunciation detection and diagnosis (MDD) approaches, PronCoach aims to provide comprehensive pronunciation guidance by integrating intended-content recovery, realized-pronunciation inference, contextual comparison, correction decisions, and personalized feedback generation. The framework employs language-specific phonetic perception models to obtain probabilistic pronunciation evidence and constructs structured reasoning supervision for training an audio-language model. The model is trained using multitask supervised fine-tuning (SFT), followed by group relative policy optimization (GRPO) to improve structured pronunciation diagnosis. The authors also introduce PronCoach-Bench, which evaluates both pronunciation correction accuracy and the ability to avoid unnecessary corrections. Experimental results demonstrate substantial improvements over general-purpose audio-language models on the proposed benchmark. Ablation studies further investigate the contributions of probabilistic pronunciation supervision, structured reasoning data, and training-data scale.

Critical strengths
  1. A comprehensive formulation of pronunciation correction beyond conventional MDD. Different from previous works that tackle different CALL tasks independently, the paper extends pronunciation assessment beyond error detection and diagnosis by integrating intended-content recovery, realized-pronunciation perception, contextual reasoning, correction decisions, and actionable feedback generation into a unified framework. In particular, the distinction between pronunciation deviations and errors that actually warrant correction is meaningful and practically important. This formulation addresses a relevant limitation of existing pronunciation assessment systems and provides a useful perspective for developing more effective pronunciation coaching models.
  2. Effective structured reasoning supervision for pronunciation correction. The proposed training framework explicitly supervises the intermediate reasoning process, including pronunciation comparison, error identification, and correction decisions, rather than learning only the final feedback. The ablation results suggests that structured reasoning supervision can help audio-language models learn more reliable pronunciation correction behavior.
Critical weaknesses
  1. Concerns about the independence and reliability of the evaluation benchmark. Although PronCoach-Bench incorporates expert phonetic annotations, the reference correction decisions are generated with the assistance of Gemini 3.1 Pro Preview, which is also involved in constructing the training supervision. Consequently, the evaluation references are not fully independent of the ALM-based annotation methodology. This raises concerns about potential annotation biases and whether the reported improvements accurately reflect pronunciation correction quality as judged by human experts. The paper would be more convincing if the authors provided independent validation of the reference corrections by qualified pronunciation experts, reported inter-annotator agreement on correction decisions, and examined whether the model rankings remain consistent under independently established human references.
  2. Limited methodological novelty and insufficient evidence to isolate the proposed contributions. The framework primarily combines existing components and techniques, including Qwen3.5-Omni, HuPER-based phonetic perception, supervised fine-tuning, and GRPO. While their integration into a pronunciation coaching framework is useful, the methodological novelty beyond task-specific data construction and system integration appears relatively limited. More importantly, the current experiments do not sufficiently establish whether the substantial performance gains arise from the proposed structured reasoning framework or simply from domain-specific fine-tuning on carefully constructed pronunciation correction data. Overall, the novelty of this paper is not up to the level of ICLR.
Questions for authors

Q1. Reliability of the benchmark references. Have the reference correction decisions in PronCoach-Bench been independently verified by human pronunciation experts? If so, could the authors report the validation procedure and inter-annotator agreement? If not, could the authors provide an evaluation on a human-validated subset to demonstrate that the reported improvements are not primarily influenced by biases introduced during ALM-assisted reference construction?

Q2. Contribution of the proposed training framework. Could the authors compare PronCoach with a simpler baseline obtained by directly fine-tuning the same Qwen3.5-Omni model on the same pronunciation correction data, without the proposed structured reasoning supervision? In particular, a comparison under matched training data, optimization steps, and computational budgets would help clarify whether the improvements are attributable to the proposed methodology rather than domain-specific fine-tuning.

Q3. Effectiveness of correction decisions. Despite the strong overall results, PronCoach achieves a relatively low Mandarin localization F1 (0.3124) and Clean-Speech accuracy (0.1728). Could the authors explain the main causes of these errors and clarify how the model distinguishes genuine pronunciation problems from acceptable pronunciation variations? This is particularly important given that avoiding unnecessary corrections is presented as a central motivation of the work.

Q4. The high ACChear score may not fully reflect the model's pronunciation perception capability. ACChear is calculated only at correctly localized error positions. For Mandarin, the model achieves a relatively low localization F1 (F1loc) of 0.3124, but obtains a very high ACChear of 0.9799. This suggests that the high ACChear score may primarily reflect the accuracy of pronunciation predictions at successfully localized error positions, rather than the model's overall ability to perceive pronunciation across an entire utterance. What is the reason? To provide a more comprehensive evaluation of pronunciation perception, I would suggest that the authors report additional metrics, such as full-utterance phoneme recognition accuracy, separate accuracy scores for Mandarin initials, finals, and tones, and confusion analyses at mispronounced positions.

4 comments
AuthorsOpenReview ↗

Response to Reviewer 8FPJ [1/3]

Thank you for your specific suggestions on benchmark reliability, the contribution of training, and the boundaries of correction.

Q1 & W1 Independently verified, reference

Sources of reference answers, manual verification, and agreement

We asked additional pronunciation experts to manually annotate and review PronCoach-Bench, resolving some missing and incorrect annotations in the original ALM-assisted references. The review records cover all recordings in the Diagnose track and the Clean track.

For Mandarin, we added one expert reviewer to the original three-expert annotation. The 296 Mandarin recordings contain 4,120 evaluable positions, of which 1,308 had pronunciation disagreements and 2,812 had unanimous judgments; the reviewed correction references cover 1,027 and 2,291 of these positions, respectively. The ALM missed some genuine errors, which have now been resolved through review by human experts. For English, we added one additional checking expert and resolved disagreements, adjusting the number of corrections in the English sentences from 176 to 151.

Among the newly added Mandarin annotations, 75.2% fall at positions where the original 3 experts disagreed, and 71.7% of the heard pronunciations match a pronunciation actually written by one of the annotators. These cases were rejected in the initial version by the ALM because it lacked fine-grained auditory perception, but human listeners can in fact hear the pronunciation errors, which were identified during the review.

Under the manually reviewed annotations, Main achieves a Mandarin F1_loc of 0.6402, compared with 0.2895 for the strongest external baseline, Doubao; the English results are 0.4464 versus 0.3045. The ranking is consistent with the original paper. Gemini 3.1 Pro Preview, which participated in constructing the training supervision, itself achieves only 0.2646 (Mandarin). After reviewing this result, we observed a Mandarin F1_loc of 0.2749 and an English result of 0.2351, far below Main. This is also consistent with the severe missing annotations observed in the original benchmark. This demonstrates that our gains do not arise from bias introduced when LALM assisted in constructing the reference answers.

Inter-annotator agreement:

LevelFull agreement among all threeDisagreement presentAgreement ratePairwise agreement rate
Whole syllable2,8321,3800.67240.7816
Onset4,0441680.96010.9734
Rime4,0231890.95510.9701
Tone3,1221,0900.74120.8275

Mandarin annotation disagreements

Disagreement slotNumber of positionsShare of disagreements
Tone only1,04375.6%
Rime only14410.4%
Onset only1319.5%
Rime + tone251.8%
Onset + tone171.2%
Onset + rime151.1%
All three slots disagree50.4%

W2 Limited methodological novelty

We address a pronunciation-correction problem that ALMs and Omni-Models had not previously addressed, and conduct experiments in two representative languages. Using the annotated data, we trained a highly accurate G2P HuPER model for data augmentation, proposed an innovative method for judging pronunciation errors based on pronunciation probabilities, and constructed a complete data pipeline and a foundational baseline model for this domain. SFT and GRPO are general-purpose methods in the large-model field, and their use should not be considered uninnovative. On the contrary, the problem in this domain can be addressed effectively with our proposed method without modifying the large-model architecture or training procedure.

Q2 Comparison with direct fine-tuning on the same data, without structured supervision

InitializationStructured supervisionLanguageRP_locF1_locACC_hearF1_hearFC ↓ACC_cZero correctionsOverall
PronCoachRetainedMandarin0.68970.60240.64310.94940.79900.23660.345728/810.7397
English0.51660.43820.47420.74360.60960.07360.546453/97
Qwen3.5-Omni baseRemovedMandarin0.77690.39720.52570.91110.83870.59060.03703/810.6809
English0.45700.27380.34240.34780.39500.12880.371136/97

We fine-tuned from the base Qwen3.5-Omni model using exactly the same configuration, removing only structured supervision. The results in the table above show that our structured supervision noticeably improves the ability to avoid erroneous corrections of actual pronunciation and substantially improves the model's performance on the clean track, resolving the previous model's problems of overcorrection and frequent hallucinations on actual pronunciation.

AuthorsOpenReview ↗

Response to Reviewer 8FPJ [2/3]

Q3 Mandarin F1_loc Clean

After reviewing and reannotating PronCoach-Bench, we found that missing reference annotations were the main reason for the Mandarin F1_loc result. We reevaluated the same checkpoint and obtained 0.6402 (R 0.6980, P_loc 0.5913, FC 0.2523).

For Clean 0.1728, our model does indeed still make a small number of corrections at some positions in standard recordings. Reporting FCR@1, FCR@2, and the average number of corrections provides a better comparison of this ability. Our model also has an advantage over the baselines. Across the task as a whole, there may still be room to improve our model's restraint in making corrections. The proportion of completely error-free data in our training set is relatively small, approximately 15%, so other data mixtures may better improve Full Clean performance; however, our model makes fewer corrections overall. When we increased the Full Clean proportion to 20%, our model achieved Clean ACC_c=0.4444.

PronCoach has a median of 1 correction and a mean of 1.51 corrections on Mandarin Clean track, and 76.5% of recordings receive at most 1 reported correction. It does not correct indiscriminately throughout the sentence. On this metric, the baselines perform worse than every configuration of our model, and the gap in the Clean column is an order-of-magnitude difference.

SystemACC_cZero correctionsFCR@1FCR@2MedianMean
PronCoach0.444436/810.76540.814811.51
Doubao-Seed-2.0-Lite0.01231/810.01230.086443.88
Qwen3.5-Omni0.345728/810.50610.728332.11
Gemini 3.8 Flash0.03703/810.09880.185244.35

Q4 ACC_hear

ACC_hear compares pronunciation candidates at correctly localized positions and measures agreement in auditory judgments at the positions that have been found. This and complete pronunciation output for the whole utterance measure two different objects.

SlotAccuracyNumber of errors
Onset0.989110
Rime0.984714
Tone0.8754114
Exact whole-syllable agreement0.8568131

Confusions at mispronounced positions: Tone confusions are distributed across tone pairs: the first tone is reported as the third tone 25 times, the fourth tone as the third tone 19 times, the fourth tone as the second tone 18 times, the second tone as the first tone 10 times, and the second tone as the fourth tone 10 times, with no single dominant pattern. There are 10 onset confusions, concentrated in alveolar versus retroflex sounds (sh→s 3 times, zh→z 2 times) and n/l; the 14 rime confusions mainly concern front versus back nasal endings (eng/en, ong/un).

The high value of Acc_hear arises because, for a fair comparison with base ALMs that lack probabilistic capabilities, we used only the first-choice phoneme as the evaluation criterion on the Mandarin test set. When this is relaxed to the 2-3 ordered phonemes in the human annotations, the results are as follows.

Mandarin auditory agreement under three candidate rules

A indicates that the model's candidates intersect with the candidate set in the manually reviewed annotations.

B indicates that the model's first candidate belongs to the candidate set in the manually reviewed annotations.

C indicates that the model's candidates are identical to the candidate set in the manually reviewed annotations in both order and candidates.

SystemHit positionsA: Any candidateB: First choice against allowed setC: Identical complete setAverage number of candidates
PronCoach9250.95030.84860.33951.68
w/ Scale data1,0250.97560.89170.41951.69
w/ Discrete targets2130.95770.95770.20661.00
Gemini 3.8 Flash2810.44480.44130.07471.06
Gemini 3.1 Pro Preview3190.34800.34800.04041.02
Doubao-Seed-2.0-Lite3000.33670.33670.06671.01
Muse-Spark-1.22940.23810.23470.02381.04

Because most ALMs report only a single candidate, we report metric A in the paper's evaluation. However, pronunciation does indeed have a probability distribution and cannot simply be represented by one deterministic sound. Most dialects and accents have pronunciation distributions "between pron1 and pron2." Therefore, in the revised manuscript, we will report the values for both "matching multiple candidates" and "matching the first choice."

Reporting only C is unfair to the baselines, but it shows the difference between our method and Discrete prediction.

AuthorsOpenReview ↗

Response to Reviewer 8FPJ [3/3]

We will move the supplementary results on the model's P_loc performance from the appendix into the main results table. These supplementary metrics measure whole-utterance perception.

ModelLanguageR ↑P_loc ↑F1_loc ↑ACC_hear ↑F1_hear ↑FC ↓FP ↓N_pjoint P ↑joint R ↑jointF1 ↑ACC_c ↑Overall F1_hear_all ↑
PronCoachMandarin0.69800.59130.64020.95000.80480.25237071,6340.47490.58880.52570.44440.7142
English0.49670.40540.44640.80000.61290.07851282030.28570.38410.32770.5670
Ablation
w/ Scale dataMandarin0.77620.78690.78150.97560.86450.11883331,3580.66860.68890.67860.34570.7060
English0.67550.30630.42150.83330.74620.16562703810.20730.52320.29700.2371
w/ Discrete targetsMandarin0.16160.86590.27240.95770.27660.0200562700.74440.15250.25310.95060.4283
English0.44370.36020.39760.52240.47980.08771432140.15420.21850.18080.6804
Baselines
Doubao-Seed-2.0-LiteMandarin0.22380.40970.28950.34240.27070.25167051,0080.09920.07590.08600.03700.2056
English0.44370.23180.30450.16420.23970.18963093830.02610.06620.03750.0206
Gemini 3.8 FlashMandarin0.20640.36760.26430.44120.28120.40191,1261,4210.08590.09260.08910.02470.2843
English0.31790.16720.21920.08330.13200.26874384990.00800.02650.01230.5979
Gemini 3.1 Pro PreviewMandarin0.22760.31580.26460.36000.27890.44861,2571,6180.06610.08120.07290.02470.3208
English0.19210.11280.14220.03450.05850.52158509220.00110.00660.00190.9072
Muse-Spark-1.2Mandarin0.20560.28800.23990.23620.21980.70811,9842,3120.02980.05240.03800.07410.2145
English0.58280.09380.16160.00000.00000.65711,0711,2180.00000.00000.00000.4227

Thanks for the updates. I'm keeping my rating at 2. The model still flags 45/81 clean Mandarin clips and has only 28.6% precision for English corrections. This contradicts the paper's goal of accurate feedback.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.