acceptodds
Withdrawn from ICLR 2027

RC-MemNet: A Regional Context Memory Network for Cross-Regional Wu Vowel Recognition

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Cross-regional Wu vowel recognition aims to identify shared vowel categories in unseen regions. Existing methods often suffer from incomplete intra-regional feature representation, insufficient inter-regional dependency modeling, and the scarcity of fine-grained cross-regional Wu vowel resources. To address these issues, we propose a Regional Context Memory Network, termed RC-MemNet, which consists of a Regional Prompt Learning (RPL) module and a Region-Guided Memory Adapter (RGMA) module. Specifically, in the RPL module, learnable regional prompts are introduced to encode region-specific acoustic characteristics through a shared encoder, providing a more complete representation of intra-regional features. Additionally, in the RGMA module, a persistent memory mechanism is introduced to propagate acoustic knowledge across source regions, supporting effective modeling of inter-regional dependencies. To increase the diversity of fine-grained cross-regional Wu vowel resources, we construct WuVowelSet, which covers 45 survey locations across 14 Wu dialect regions. Comprehensive experiments are conducted on our presented WuVowelSet dataset. The experimental results demonstrate that RC-MemNet achieves state-of-the-art performance on WuVowelSet, which is conducive to promoting the dissemination of dialect culture.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
cross-regional, vowel recognition, memory network, representation learning

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_2L6YOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The model is trained to classify vowels using speech samples from different regions, with an intermediate token bank serving as a form of memory, where each token learns to capture region-specific characteristics.

When speech from an unseen region is introduced, the parameters of the rest of the model remain frozen and shared with the previously trained model. Only the weights that determine the degree of conditioning applied to intermediate features between layers in the Reginal Prompt Learning (RPL) are learned.

After a brief adaptation phase using speech from an unseen region, FiLM-based conditioning is applied using the memory obtained from Region-Guided Memory Adapter (RGMA) to enhance the features, which are then used for vowel classification.

Critical strengths
  • The proposed method achieves SOTA performance in the 4-shot setting across six unseen regions, outperforming the main baselines, including MAS-LoRA and WhisAID.
  • Ablation studies on each module demonstrate the necessity and effectiveness of the individual components.
Critical weaknesses
  • The authors constructed their own dataset, WuVowelSet, for both training and evaluation. However, since the dataset has not been publicly released, independent reproduction of the results is not possible.
  • Although different source-prompt composition weights are assigned to different unseen regions, there are no experiments investigating whether these differences actually arise from similarities between the regions.
Questions for authors
  • Additional experiments are needed to investigate whether the different source-prompt composition weights assigned to unseen regions are actually related to the similarities between those regions.
  • The paper reports few-shot experiments on pretrained models such as Whisper. How exactly was few-shot adaptation performed for these pretrained baseline models?
Reviewer_edGXOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The authors target the problem of vowel recognition in Wu, a variety of Chinese. Since there are no resources for this task, they collect and annotate recordings of vowels in Spoken Wu to train and evaluate models. To improve performance, they propose Regional Context Memory Network (RC-MemNet) , which consists of a Regional Prompt Learning (RPL) module and a Region-Guided Memory Adapter (RGMA) module. Experimental results on their collected dataset (WuVowelSet) suggest that the use of RC-Memnet outperforms other baseline methods for few-shot adaptation.

Critical strengths
  • The paper targets a low-resource problem domain (Wu Vowel Recognition) and releases a dataset to support future work on it
Critical weaknesses

Critical weaknesses

  • The proposed method is quite complicated for what is a relatively easy task (vowel classification), and its unclear if it can be generalized to other tasks or problems. The proposed method requires additional information to what is typically required for vowel recognition (region / accent labels), limiting its scalability and applicability. Injecting geographic or accent information into speech recognizers is not a new idea, and it has been already applied to ASR (a harder task) [1] and even without the use of manual accent/region labels [2].
  • It is very hard to gauge the effectiveness of the proposed method due to the specificity of the data used. If the authors want to show that their proposed method is an interesting ML technique, it would be more beneficial to evaluate the method on standard ASR/PR benchmarks like LibriSpeech, AISHELL, ML-SUPERB, or even Common Phone.

Other issues

  • the use of inline citations without paratheses (\cite instead of \citep), makes the text very hard to read
  • the problem of Wu vowel recognition is very niche, even within the domain of speech research, and highly unlikely to be interesting to the ICLR nor general ML research community. The paper would likely be of more interest to venues like LREC.

[1] Mehralian, P., Van hamme, H. (2025) Leveraging Geographic Metadata for Dialect-Aware Speech Recognition. Proc. Interspeech 2025 [2] Prabhu, Darshan, et al. "Improving self-supervised pre-training using accent-specific codebooks.". Proc. Interspeech 2024.

Questions for authors
  1. How are the baseline methods trained? ASR and SSL models cannot natively perform vowel recognition, so how were they adapted to this task?
Reviewer_g7CvOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces RC-MemNet, a framework for recognizing vowel categories across different Wu Chinese dialect regions. contributions of the paper: 1. RPL, learnable region-specific prompts into a pretrained speech encoder to capture fine-grained acoustic variations; 2. RGMA, a category-aware memory that aggregates acoustic information across source regions. WuVowelSet dataset

Critical strengths
  1. contribution to low-resource dialect recognition through a new dataset.
  2. Strong empirical performance supported by comprehensive evaluations and ablation studies. The experiments include comparisons with a diverse range of speech models, component-level ablations, and evaluations across different few-shot settings.
Critical weaknesses
  1. Several reported results are somehow inconsistent, e.g. in Fig. C.2, no Jingxian class has recall above about 63%, which would make 64.99% accuracy impossible. most values in Fig. C.1 values also appear lower than Table 2, which raises my concerns about the experimental results
  2. I have a few concerns on fair comparison vs baselines, At test time, RC-MemNet keeps its source-trained classifier and memory and tunes only composition scalars. The strongest baselines discard their source-trained classifiers and classify by 4-shot prototypes. The paper’s own ablation shows RC-MemNet without target adaptation already roughly matches MAS-LoRA, would this suggest much of the reported gains comes from this asymmetry rather than from RPL or RGMA?
Questions for authors

please first address my concerns on the experiments

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

Related papers

Discussion (0)

Sign in to comment.