Paper deep dive
ArtNet: A JEPA-Like Articulatory Predictive Framework for Robust Zero-Shot Phoneme Recognition
Zeqian Hu, Fuliang Weng, Shu Shang, Yaqian Zhou
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 99%
Last extracted: 6/20/2026, 8:37:43 AM
Summary
ArtNet is a JEPA-like (Joint-Embedding Predictive Architecture) framework designed for robust zero-shot cross-lingual phoneme recognition. It addresses the fragility of direct acoustic-to-symbol mapping by predicting structured articulatory features instead. The framework integrates an articulatory predictor with a Variational Information Bottleneck (VIB) to suppress language-specific variations and uses a Vector-Space Inventory Alignment (VSIA) strategy for zero-shot inference. Experimental results on seven unseen languages show that ArtNet, especially with VSIA, significantly reduces Phoneme Error Rate (PER) and Phoneme Feature Error Rate (PFER) compared to standard SSL-based baselines.
Entities (9)
Relation Signals (5)
ArtNet → predicts → articulatory features
confidence 100% · a framework that explores a structured feature prediction task based on articulatory features
Panphon → provides → articulatory attributes
confidence 100% · we utilize the Panphon database to decompose each IPA symbol into a 24-dimensional vector of articulatory attributes
ArtNet → synergizedwith → Vector-Space Inventory Alignment
confidence 100% · ArtNet, particularly when synergized with the proposed vector-space inventory alignment (VSIA) strategy
ArtNet → uses → Variational Information Bottleneck
confidence 100% · ArtNet integrates an articulatory predictor... with a variational information bottleneck (VIB)
ArtNet → utilizes → mHuBERT-147
confidence 100% · In this work, we employ mHuBERT-147 as the SSL backbone for speech representation extraction.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Zero-shot cross-lingual phoneme recognition is often hindered by the fragility of direct acoustic-to-symbol mapping, which is susceptible to language-specific variations. Echoing joint-embedding predictive architecture (JEPA) work in vision, we propose ArtNet, a framework that explores a structured feature prediction task based on articulatory features to enhance acoustic robustness. Specifically, ArtNet integrates an articulatory predictor, designed to extract universal articulatory representations from self-supervised learning (SSL) features, with a variational information bottleneck (VIB) to suppress language-specific variations. Experiments on seven unseen languages demonstrate that ArtNet, particularly when synergized with the proposed vector-space inventory alignment (VSIA) strategy, significantly outperforms competitive baselines, achieving a 20.56\% relative reduction in phoneme error rate (PER) and 7.01\% in phoneme feature error rate (PFER).
Tags
Links
- Source: https://arxiv.org/abs/2606.16595v1
- Canonical: https://arxiv.org/abs/2606.16595v1
Trouble viewing inline? Open PDF directly →
Full Text
26,315 characters extracted from source content.
Expand or collapse full text
ArtNet: A JEPA-Like Articulatory Predictive Framework for Robust Zero-Shot Phoneme Recognition Zeqian Hu ID 1,∗ , Fuliang Weng 2,∗ , Shu Shang ID 1 , Yaqian Zhou 1,∗ 1 Fudan University, China 2 Pedawise, Shanghai, China zqhu24@m.fudan.edu.cn, fuliang.weng@pedawise.cn, zhouyaqian@fudan.edu.cn Abstract Zero-shot cross-lingual phoneme recognition is often hin- dered by the fragility of direct acoustic-to-symbol mapping, which is susceptible to language-specific variations. Echoing joint-embedding predictive architecture (JEPA) work in vision, we propose ArtNet, a framework that explores a structured fea- ture prediction task based on articulatory features to enhance acoustic robustness. Specifically, ArtNet integrates an articula- tory predictor—designed to extract universal articulatory repre- sentations from self-supervised learning (SSL) features—with a variational information bottleneck (VIB) to suppress language- specific variations. Experiments on seven unseen languages demonstrate that ArtNet, particularly when synergized with the proposed vector-space inventory alignment (VSIA) strat- egy, significantly outperforms competitive baselines, achieving a 20.56% relative reduction in phoneme error rate (PER) and 7.01% in phoneme feature error rate (PFER). Index Terms:zero-shot transfer learning, cross-lingual, phoneme recognition, articulatory features 1. Introduction Modern end-to-end (E2E) automatic speech recognition (ASR) systems have achieved remarkable success in resource-rich lan- guages such as English and Mandarin. However, the vast ma- jority of the world’s languages lack sufficient paired audio-text data, posing a significant challenge for E2E models to achieve comparable precision. In multilingual and cross-lingual scenarios, a more data- efficient paradigm involves developing a robust phoneme rec- ognizer as an intermediate stage [1]. The prevailing architec- ture for training such phoneme recognizers typically follows an encoder-decoder framework. Under this paradigm, the en- coder is first pre-trained through self-supervised learning like Wav2Vec2 [2] and HuBERT [3], followed by a joint fine-tuning of both the encoder and decoder on a limited amount of paired audio-text data. Such a framework is hereafter referred to as an SSL-based phoneme recognizer. However, empirical analysis suggests of our phoneme rec- ognizer trained on English that the primary bottleneck in cross- lingual generalization stems not only from the presence of un- seen phonemes but, more significantly, from the brittleness of previously learned phonetic representations. As illustrated in Fig. 1(left), the macro-average decomposition of the PER across seven unseen languages indicates that the substitution error is the predominant component, accounting for 84.7% of the total errors. This confirms that the challenge lies in the categorical discriminability of the phonetic frontend. * These authors contributed equally. ** indicates the corresponding author. Sub 84.7% Del 11.8% Ins 3.5% Average Composition of PER IV-Sub 61.6% OOV-Sub 38.4% Average Composition of Sub: IV vs OOV Figure 1: Detailed error analysis of the phoneme recognizer in zero-shot scenario. The left chart illustrates the macro-average decomposition of the PER into Substitution (Sub), Deletion (Del), and Insertion (Ins) error rate. The right chart provides a breakdown of substitution errors, distinguishing between in- vocabulary (IV) and out-of-vocabulary (OOV) phonemes across all unseen languages. Crucially, a deeper examination of these substitutions in Fig. 1(right) reveals a counter-intuitive finding: while one might expect out-of-vocabulary (OOV) phonemes to be the primary source of error in zero-shot scenarios, 61.6% of substitutions are actually attributed to in-vocabulary (IV) phonemes. These observations suggest that the performance degrada- tion of the cross-lingual recognition models is largely driven by the lack of discrimination at the shared phoneme representa- tion level. Although existing work [4] has attempted to incor- porate auxiliary linguistic knowledge to mitigate this, they of- ten overlook the need to improve the intrinsic robustness of the phoneme recognizer itself. In line with the philosophy of JEPA in computer vision [5] and its recent extensions to the audio domain [6, 7], we posit that this vulnerability stems from the prevailing training paradigm that attempts to map raw acous- tic signals directly onto discrete symbols—a process inherently susceptible to non-essential acoustic fluctuations and at the sur- face level. Therefore, rather than focusing on OOV mapping, a more critical priority is to shift toward a predictive model- ing framework that extracts language-invariant features within a structured representation space, thereby strengthening the dis- criminative robustness of IV phonemes across diverse linguistic environments. The articulatory features, which describe the physical movements of speech organs during sound production, offer a potential solution to this generalization gap. This perspective aligns with the foundational theory of The Sound Pattern of En- glish [8], which posits that phonemes are not atomic symbols but bundles of universal distinctive features grounded in hu- man physiology. Unlike acoustic signals, articulatory features arXiv:2606.16595v1 [cs.SD] 15 Jun 2026 are fundamentally shared between languages. Existing frame- works have utilized these features as a universal bridge by de- composing phonemes into shared attribute spaces or employing phonological-vector based phone embeddings [9, 10, 11, 12]. Although promising, these models often fail to disentangle lin- guistic representations from language-specific acoustic charac- teristics and prosodic patterns. Consequently, they risk over- fitting to source linguistic environments, which limits their ro- bustness when encountering truly unseen languages with signif- icantly different acoustic variations. To address the limitations, we propose a novel framework to solve these problems. The primary contributions of this re- search are summarized as follows: • ArtNet: A JEPA-like articulatory framework: Mov- ing beyond the conventional paradigm of direct acoustic- to-symbol mapping, we reformulate cross-lingual phoneme recognition as a non-generative predictive task. By utiliz- ing a specialized articulatory predictor to map SSL fea- tures into a structured, physical articulatory space, our model learns representations that are inherently more stable across linguistic boundaries than raw acoustic embeddings. • Mitigation of the substitution error bottleneck:We demonstrate that ArtNet successfully mitigates the predom- inant substitution errors by filtering out language-specific variations. This ensures that the extracted articulatory cues remain robust and discriminative even when encountering significantly different acoustic context. • Novel inference strategy and superior performance: To maximize the utility of the learned articulatory space, we introduce vector-space inventory alignment (VSIA), as a zero-shot inference strategy. Driven by this methodological design, our approach achieves an average relative reduction of 20.56% in PER and 7.01% in PFER over a competitive SSL-based baseline across seven unseen languages. 2. Methodology 2.1. Construction of Structured Articulatory Targets To enable the explicit modeling of the articulatory space, we construct a structured target space that provides supervision without the need for manual phonetic annotations. Within this framework, the SSL backbone functions as a context encoder, which is first optimized via a connectionist temporal classifica- tion (CTC) objective to establish a foundational phoneme rec- ognizer. To unify the label space, orthographic transcripts are transduced into international phonetic alphabet (IPA) sequences using the Epitran G2P library [13]. Once initialized, the phoneme recognizer serves as a pseudo-label generator that provides stable, frame-level pho- netic alignments necessary for guiding the subsequent predic- tive learning process. Let H = (h 1 ,...,h T ) denote the se- quence of frame-level hidden representations extracted from the context encoder. For each frame, the pseudo-label generator produces a phoneme π t ∈ Φ via greedy decoding, where Φ represents the universal phoneme inventory. To facilitate learning in a structured target space, we utilize the Panphon database [14] to decompose each IPA symbol into a 24-dimensional vector of articulatory attributes, originally rep- resented by the trinary values +,−, 0. For network com- patibility, we transform these categorical labels into a numeri- cal encoding scheme where + is mapped to 1, − to −1, and0 remains 0. This process yields a static mapping matrix M ∈ −1, 0, 1 |Φ|×24 . Consequently, the ground-truth artic- ulatory vector v t for each time step is derived via a mapping functionM(·) that performs a row-lookup on M according to the generated pseudo-label π t : v t =M(π t )(1) By mapping the discrete pseudo-labels into this continuous, structured target space, we provide language-agnostic supervi- sion signals. This predictive paradigm forces the model to dis- entangle linguistic content from the language-specific variations inherent in the encoder representations, establishing a more uni- versal foundation for cross-lingual transfer. 2.2. ArtNet To further improve the robustness of feature disentanglement, we propose ArtNet, a dedicated neural framework for structured articulatory modeling. ArtNet is composed of an articulatory predictor (AP) equipped with a variational information bottle- neck (VIB). The AP is designed as a lightweight module to pre- dict articulatory vectors from a robust latent space. Instead of feeding the hidden state h t directly to the AP, we maps it to a stochastic latent encoding z t to suppress noise. We assume the latent distribution p(z t |h t ) follows a mul- tivariate gaussian distribution N (μ t ,σ 2 t ). The parameters are predicted by the VIB encoder, where f μ (·) and f σ (·) are imple- mented as fully connected layers: μ t = f μ (h t ),σ t = f σ (h t )(2) Using the reparameterization trick, we sample z t as: z t =μ t +σ t ⊙ε,ε∼N (0,I)(3) The AP then predicts the articulatory vector ˆ v t from this latent variable: ˆ v t = AP(z t )(4) The loss of reconstruction L AP measures the discrepancy between the predicted articulatory vector ˆ v t and the ground truth v t . We employ the mean squared error (MSE) as the dis- tance metric: L AP = 1 T T X t=1 ∥ v t − ˆ v t ∥ 2 2 (5) Information bottleneck regularization is defined using the Kullback-Leibler (KL) divergence. Since the prior r(z) is a standard normal distribution N (0,I), the divergence can be computed analytically: L VIB = 1 T T X t=1 D KL (p(z t |h t )∥ r(z)) (6) Finally, the total objective function integrates the recon- struction loss and the regularization via a hyperparameter β: L =L AP + βL VIB (7) 2.3. Zero-Shot Inference Strategy During the inference phase, we adopt a hybrid decoding strategy designed to handle the constraints of zero-shot scenarios. The CTC decoder first determines the topology of the sequence. For any time step t where the CTC prediction y ctc t is ⟨blank⟩, the frame is considered devoid of articulatory content and is explic- itly skipped. 2.3.1. Pooling Inference In zero-shot cross-lingual settings, the CTC model often ex- hibits limited capability in precisely determining phoneme boundaries, rendering frame-by-frame inference susceptible to noise and instability. To mitigate this, we introduce a pooling inference strategy. Instead of predicting articulatory features for every single frame, we group consecutive frames that are pre- dicted as the same non-blank class into a segment. For each segment S, we compute the mean of the corresponding encoder output vectors h t : ̄ h S = 1 |S| X t∈S h t (8) This aggregated representation ̄ h S is then fed into AP to predict a robust articulatory vector ˆ v S . Note that strictly speak- ing, ˆ v S resides in a continuous physical space and does not yet correspond to a discrete symbol. 2.3.2. Vector-Space Inventory Alignment (VSIA) To overcome the rigid phoneme-level inventory constraints im- posed by the tr2tgt strategy [15], we introduce VSIA to adapt continuous predictions directly to the target language. While the original strategy forces a many-to-one discrete mapping by minimizing the Hamming edit distance over binary articulatory attributes, our approach explicitly leverages the rich, continuous geometric structure learned by AP. Specifically, we utilize the underlying structure of the target language’s phoneme inventory, denoted as Φ tgt , to constrain the search space. Instead of discretizing the predicted vector ˆ v S immediately, we perform a nearest-neighbor search based on angular proximity in the continuous articulatory space. The final predicted phoneme ˆ φ is identified as the candidate in Φ tgt that maximizes the cosine similarity with the prediction: ˆ φ = argmax φ∈Φ tgt ˆ v S ·M(φ) ∥ ˆ v S ∥M(φ)∥ (9) Through soft alignment, VSIA mitigates phonological mis- matches and ensures the output adheres to the target language’s valid phonetic definitions. 3. Experiment 3.1. Datasets To evaluate the cross-lingual generalization capability of our proposed method, we define distinct source and target domains. • Training Data (Source) We utilize the LibriSpeech [16] train-clean-100 dataset as our primary training corpus. This subset consists of approximately 100 hours of clean English speech. • Test Data (Target) For zero-shot evaluation, we utilize the standard test sets of seven non-English languages (German, Dutch, French, Spanish, Italian, Portuguese, and Polish) from the Multilingual LibriSpeech (MLS) dataset[17]. • Data Preprocessing To unify the label space across diverse languages, we employ Epitran to convert the orthographic transcriptions of all datasets (both LibriSpeech and MLS) into the IPA symbols. 3.2. Model Training and Implementation Details In this work, we employ mHuBERT-147 1 [18] as the SSL back- bone for speech representation extraction. It is a 95M parameter model consisting of 12 Transformer encoder layers, with an em- bedding dimension of 768 and 12 attention heads. The model was pre-trained on an extensive multilingual corpus encompass- ing 147 different languages. During the phoneme recognizer initialization phase, we uti- lize low-rank adaptation (LoRA) to efficiently optimize the con- text encoder and CTC head. In the subsequent articulatory fea- ture prediction phase, the phoneme recognizer’s parameters are frozen, and training focuses exclusively on the AP and VIB modules. To investigate the role of temporal context in feature extraction, we evaluate ArtNet across three architectural vari- ants: a Multi-Layer Perceptron (MLP), a Time Delay Neural Network (TDNN), and a Long Short-Term Memory (LSTM) network.These architectures correspond to distinct levels of temporal modeling—ranging from the context-independent MLP and the locally constrained TDNN to the LSTM, which captures long-range sequential dependencies. This comparison allows us to isolate the optimal level of temporal dynamics re- quired for robust articulatory modeling. For the VIB configu- ration, the dimension of the latent variable z is set to 128, with the trade-off parameter β set to 0.001 to balance reconstruction loss and KL-divergence regularization. Both stages are optimized using the Adam optimizer [19]. The learning rate is linearly warmed up for the first 10% train- ing steps and peaks at 1e-3. To ensure a strict zero-shot evalua- tion, the entire framework is trained exclusively on the English source data. 3.3. Evaluation Metrics Standard PER doesn’t consider the phonemic differences in terms of phonological similarity. To directly evaluate the qual- ity of the predicted physical articulatory attributes, we use the PFER metric as described in [20]. PFER measures the distance between the predicted feature vectors and the ground-truth ar- ticulatory features, providing a more granular assessment of the model’s phonetic disentanglement ability. 4. Results and Analysis 4.1. Zero-Shot Cross-Lingual Performance To evaluate the effectiveness of our proposed method in zero- shot scenarios, we conducted experiments on seven target lan- guages using a model trained solely on English. The perfor- mance is measured using PER and PFER. The main results, utilizing the TDNN-based ArtNet syner- gized with the VSIA strategy, are presented in Table 1. Com- pared to the standard SSL-based baseline, this combined ap- proach demonstrates consistent improvements across the ma- jority of target languages and metrics. Specifically, the integra- tion of ArtNet and VSIA yielded an average relative reduction of 20.56% in PER compared to the baseline. Notably, substan- tial gains are achieved in Romance languages such as Spanish, where the PER decreased by approximately 28.26% absolute points. The overall reduction in PFER (averaging 7.01% rel- ative improvement) further confirms that our proposed frame- work, by aligning robust articulatory predictions with the target 1 https://huggingface.co/utter-project/ mHuBERT-147 Table 1: Zero-shot cross-lingual performance comparison (in %) measured by PER and PFER. “tr2tgt” denotes the phoneme mapping strategy, and “ArtNet+VSIA” represents the proposed method. The best results are highlighted in bold. ↓ indicates lower is better. DutchFrenchGermanItalianPolishPortugueseSpanishAvg PER↓ Baseline59.6759.3352.6354.4355.0861.3858.7657.33 ArtNet58.6456.8451.6351.7350.2459.9455.5754.94 Baseline+tr2tgt56.1256.5450.4846.0741.0158.0234.2748.93 ArtNet+VSIA 55.40 53.7550.04 39.96 35.1853.9330.50 45.54 PFER↓ Baseline15.5421.9012.6610.1610.6414.1610.7913.69 ArtNet15.3321.7612.499.8110.2113.8910.3313.40 Baseline+tr2tgt15.8420.9612.7810.2310.5814.519.4113.47 ArtNet+VSIA16.08 20.8212.718.489.4913.637.9212.73 inventory, successfully captures more accurate articulatory fea- tures. 4.2. Analysis of ArtNet It is important to note that both tr2tgt and VSIA strategy lever- age external linguistic knowledge to align source and target phoneme spaces. To decouple this external benefit and rigor- ously evaluate the intrinsic robustness of the articulatory fea- tures learned by ArtNet, the following analyses focus on the model outputs before the application of zero-shot inference strategy. 4.2.1. Generalization Analysis: Insights from Substitution Er- rors To assess adaptability, we analyze substitution errors (Table 2). Beyond the anticipated 5.67% error reduction for in-vocabulary (IV) tokens, ArtNet surprisingly reduced out-of-vocabulary (OOV) errors by 1.62%. This dual improvement indicates that rather than overfitting to training patterns, explicitly modeling articulatory features allows ArtNet to capture universal phonetic properties that generalize effectively to unseen tokens. Table 2: Analysis of the number of substitution error tokens. MethodTotalIVOOV Baseline1,078,964665,073413,891 ArtNet1,034,535627,342407,193 Relative Reduction4.12%5.67%1.62% 4.2.2. Ablation Study: Impact of ArtNet Backbone Architecture To determine the optimal architecture for extracting articulatory features, we evaluate three ArtNet backbone variants: a 2-layer MLP (no context), a 2-layer TDNN (local context), and a 2- layer LSTM (global context). We compare these against the baseline to quantify the gain derived from articulatory model- ing. The average PER across all seven target languages is sum- marized in Table 3. As shown in the results, all ArtNet variants outperform the baseline. The experimental results, with TDNN (54.94%) yielding the lowest error rate, followed by MLP (55.33%) and LSTM (56.51%), provide critical insights into cross-lingual modeling. Specifically, while LSTM is capable of modeling long- range dependencies, its higher error rate suggests that the global context from the source language contains language-specific phonotactic and prosodic biases. These biases likely act as high-level variations during zero-shot transfer to a target lan- guage with different rhythmic structures. In contrast, TDNN strikes an ideal balance by utilizing a constrained receptive field; it captures universal, local phonetic transitions while fil- tering out detrimental global biases. Furthermore, the fact that even the frame-level MLP outperforms the baseline underscores the inherent robustness of the articulatory space as a language- agnostic anchor, which proves more effective for cross-lingual mapping than raw acoustic features alone. Table 3: Average PER (%) of different ArtNet backbone archi- tectures across seven target languages. Model / Backbone Context Type Avg. PER↓ Baseline-57.33 ArtNet (LSTM)Global56.51 ArtNet (MLP)None55.33 ArtNet (TDNN)Local54.94 5. Conclusion In this work, we addressed the generalization challenge in zero- shot cross-lingual ASR by reformulating articulatory modeling as a JEPA-like structured prediction task. We proposed Art- Net, a framework that integrates an articulatory predictor with a VIB to project SSL representations into a universal, physical target space. This architecture proved critical in filtering out language-specific variations and implicitly handling structural redundancy, thereby significantly reducing substitution errors. Our comprehensive evaluations show that ArtNet, particularly when synergized with the proposed VSIA strategy, yields su- perior performance in terms of PER and PFER. Overall, our findings suggest that shifting towards non-generative, predic- tive modeling of articulatory features offers a promising direc- tion for enhancing the intrinsic robustness of next-generation multilingual speech systems. 6. Generative AI Use Disclosure During the preparation of this manuscript, the authors used Gemini solely for editing and polishing the English language. The tool was not used for producing any significant part of the manuscript. 7. References [1] S. Yusuyin, T. Ma, H. Huang, W. Zhao, and Z. Ou, “Whistle: Data-efficient multilingual and crosslingual speech recognition via weakly phonetic supervision,” IEEE Transactions on Audio, Speech and Language Processing, 2025. [2] A. Baevski, Y. Zhou, A. Mohamed, and M. Auli, “wav2vec 2.0: A framework for self-supervised learning of speech repre- sentations,” Advances in neural information processing systems, vol. 33, p. 12 449–12 460, 2020. [3] W.-N. Hsu, B. Bolte, Y.-H. H. Tsai, K. Lakhotia, R. Salakhut- dinov, and A. Mohamed, “HuBERT: Self-supervised speech representation learning by masked prediction of hidden units,” IEEE/ACM transactions on audio, speech, and language process- ing, vol. 29, p. 3451–3460, 2021. [4] H. Gao, J. Ni, Y. Zhang, K. Qian, S. Chang, and M. Hasegawa- Johnson, “Zero-shot cross-lingual phonetic recognition with ex- ternal language embedding,” in Proc. Interspeech 2021, 2021, p. 1304–1308. [5] M. Assran, Q. Duval, I. Misra, P. Bojanowski, P. Vincent, M. Rab- bat, Y. LeCun, and N. Ballas, “Self-supervised learning from im- ages with a joint-embedding predictive architecture,” in Proceed- ings of the IEEE/CVF conference on computer vision and pattern recognition, 2023, p. 15 619–15 629. [6] G. Ioannides, C. Constantinou, A. Chadha, A. Elkins, L. Pang, R. Shwartz-Ziv, and Y. LeCun, “JEPA as a neural tokenizer: Learning robust speech representations with density adaptive at- tention,” arXiv preprint arXiv:2512.07168, 2025. [7] L. Tuncay, E. Labb ́ e, E. Benetos, and T. Pellegrini, “Audio-JEPA: Joint-embedding predictive architecture for audio representation learning,” in ICME 2025, 2025. [8] N. Chomsky and M. Halle, The sound pattern of English. ERIC, 1968. [9] X. Li, S. Dalmia, D. Mortensen, J. Li, A. Black, and F. Metze, “Towards zero-shot learning for automatic phonemic transcrip- tion,” in Proceedings of the AAAI Conference on Artificial Intelli- gence, vol. 34, no. 05, 2020, p. 8261–8268. [10] C. Zhu, K. An, H. Zheng, and Z. Ou, “Multilingual and crosslin- gual speech recognition using phonological-vector based phone embeddings,” in 2021 IEEE Automatic Speech Recognition and Understanding Workshop (ASRU). IEEE, 2021, p. 1034–1041. [11] J. Lee, M. Mimura, and T. Kawahara, “Embedding articulatory constraints for low-resource speech recognition based on large pre-trained model,” in Proc. Interspeech 2023, 2023, p. 1394– 1398. [12] H. Yen, S. M. Siniscalchi, and C.-H. Lee, “Boosting end-to-end multilingual phoneme recognition through exploiting universal speech attributes constraints,” in ICASSP 2024-2024 IEEE Inter- national Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2024, p. 11 876–11 880. [13] D. R. Mortensen, S. Dalmia, and P. Littell, “Epitran: Precision G2P for many languages,” in Proceedings of the Eleventh Interna- tional Conference on Language Resources and Evaluation (LREC 2018), 2018. [14] D. R. Mortensen, P. Littell, A. Bharadwaj, K. Goyal, C. Dyer, and L. Levin, “Panphon: A resource for mapping IPA segments to articulatory feature vectors,” in Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, 2016, p. 3475–3484. [15] Q. Xu, A. Baevski, and M. Auli, “Simple and effective zero-shot cross-lingual phoneme recognition,” in Proc. Interspeech 2022, 2022, p. 2113–2117. [16] V. Panayotov, G. Chen, D. Povey, and S. Khudanpur, “Lib- riSpeech: an ASR corpus based on public domain audio books,” in 2015 IEEE international conference on acoustics, speech and signal processing (ICASSP). IEEE, 2015, p. 5206–5210. [17] V. Pratap, Q. Xu, A. Sriram, G. Synnaeve, and R. Collobert, “MLS: A large-scale multilingual dataset for speech research,” in Proc. Interspeech 2020, 2020, p. 2757–2761. [18] M. Zanon Boito, V. Iyer, N. Lagos, L. Besacier, and I. Calapode- scu, “mHuBERT-147: A compact multilingual HuBERT model,” in Proc. Interspeech 2024, 2024, p. 3939–3943. [19] D. P. Kingma and J. Ba, “Adam: A method for stochastic opti- mization,” arXiv preprint arXiv:1412.6980, 2014. [20] C. Taguchi, Y. Sakai, P. Haghani, and D. Chiang, “Universal au- tomatic phonetic transcription into the International Phonetic Al- phabet,” in Proc. Interspeech 2023, 2023, p. 2548–2552.