Mispronunciation Detection Without L2 Pronunciation Dataset in Low-Resource Setting: A Case Study in Finland Swedish

Anna von Zansen; Ekaterina Voskoboinik; Maria Kautonen; Mikko Kurimo; Mikko Kuronen; Nhan Phan; Riikka Ullakonoja; Tam\'as Gr\'osz; Yaroslav Getman

REVIEW 3 major objections 6 minor 37 references

Mispronunciation Detection Without L2 Pronunciation Dataset in Low-Resource Setting: A Case Study in Finland Swedish

T0 review · 3 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Finland Swedish can have a mispronunciation detector trained on native speech alone, with a post-inference calibration algorithm balancing precision and recall despite the lack of an L2 pronunciation corpus.

desk verdict Honest low-resource MD case study whose headline numbers rest on noisy proxy labels; worth refereeing with fixes to evaluation framing and ablation. read the letter →

arxiv 2506.01156 v1 pith:TYLIXYK5 submitted 2025-06-01 cs.CL cs.SDeess.AS

Nhan Phan , Mikko Kuronen , Maria Kautonen , Riikka Ullakonoja , Anna von Zansen , Yaroslav Getman , Ekaterina Voskoboinik , Tamás Grósz

show 1 more author

Mikko Kurimo

This is my paper · ORCID

classification cs.CLcs.SDeess.AS

keywords mispronunciationdetectionFinlandSwedishlow-resourcespeechtechnologycomputer-assistedpronunciationtrainingwav2vec2.0temperaturescalingtop-knormalizationentropyregularization

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The authors set out to show that a usable mispronunciation detector can be built for Finland Swedish, a low-resource language variety, without any second-language (L2) pronunciation dataset. They fine-tune a multilingual self-supervised speech model on roughly 89 hours of native Finland Swedish speech and test it on 33 minutes of L2 read-aloud speech, using the test set's verbatim transcriptions as stand-in labels for pronunciation errors. The paper's contribution is a post-inference calibration step, temperature scaling with top-$k$ normalization, that softens the model's overconfident character scores and lets Precision and Recall be traded off without L2 validation data. The tuned system catches 43.2% of the mispronunciations noted by the transcriber, with 29.8% of its alarms being correct (F1 35.0% at character level, 48.3% at word level), against the raw baseline's 77.5% Recall at 17.6% Precision. If the approach is right, it offers a route to pronunciation feedback for other under-resourced languages that lack phonetically annotated L2 corpora.

What carries the argument

The load-bearing mechanism is a two-step post-inference algorithm the paper calls simplified temperature scaling and top-$k$ normalization: divide the character-logit vector by a temperature $T$, apply the softmax, then take the $k$ most probable labels and rescale their probabilities by the top-1 probability so the top label stays at 1 and plausible alternatives rise from near zero to comparable scores. This counteracts the peaky, overconfident CTC (Connectionist Temporal Classification) output that a model trained only on L1 speech produces on out-of-domain L2 audio, and it does so without needing an L2 validation set to tune a threshold. Supporting it are the multilingual XLS-R checkpoint of wav2vec 2.0 (a self-supervised speech-representation architecture), CTC forced alignment for character-level scores, and maximum-entropy regularization (a training penalty against overconfident frame predictions) during fine-tuning, which flattens the output distribution so the post-processing has something to work with. The result is a continuous score instead of a hard threshold, letting a learner-facing application grade pronunciation as partially correct and change strictness simply by changing $T$.

What would settle it

Take a sample of the DigiTala read-aloud recordings, have trained phonetic experts produce word-level mispronunciation labels for them, and compare the model's decisions against these expert labels instead of against the original transcriptions; if the model matches the experts no better than it matches the original transcription proxy, or if expert-identified errors that the transcriber omitted are mostly missed, the central claim is not supported.

Watch

Extended reading notes

Core claim

The central claim is that mispronunciation detection for a low-resource language variety does not require a phonetically annotated L2 corpus: a wav2vec 2.0 model (specifically the multilingual XLS-R checkpoint) fine-tuned on native speech with CTC and maximum-entropy regularization, followed by a simple two-step post-processing algorithm, can detect a useful share of the mispronunciations that human transcribers note in L2 read-aloud speech. The algorithm divides the per-frame logits by a temperature $T$, applies the softmax, then rescales the top-$k$ probabilities by the top-1 probability so plausible alternative characters receive scores comparable to the predicted one; with $T{=}10$ and $k{=}3$ the model balances Recall and Precision (43.2% and 29.8% at character level) and improves F1 over the uncalibrated baseline. The authors also report, on a small expert-recorded set, word-level Precision of 80% at 33% Recall, and show the model flags some words whose Finland Swedish and Sweden Swedish pronunciations differ. The paper's position is that the reported numbers are relative, not absolute, because no proper L2 pronunciation corpus exists to give ground truth.

Load-bearing premise

The load-bearing premise is that the verbatim orthographic transcriptions in the DigiTala test set reliably mark mispronunciations, even though the transcriber was not required to note every mistake and was not a phonetic expert; if this stand-in fails, the reported recall and precision do not measure mispronunciation detection at all.

Editorial extensions

If this is right

A computer-assisted pronunciation training tool for Finland Swedish can be built entirely from existing native-speech corpora, with no costly phonetic annotation of learner speech.
The same L1-only pipeline and post-processing algorithm can be applied to other low-resource languages or varieties, since nothing in it depends on an L2 dataset or on shallow orthography.
The model accepts some Sweden Swedish pronunciations (for example, of fara and göra) as correct, which the authors view as aligned with intelligibility-oriented teaching, while for words like sju and djur it flags the Sweden Swedish form as a likely mispronunciation.
Because the temperature can be changed after inference at almost no cost, the same application can serve different learners, from beginners who need lenient feedback to advanced users who want stricter practice, and can adjust dynamically as a learner improves.
Word-level feedback is the reliable level for learners, since the grapheme-to-phoneme mapping is not modeled; character-level scores are an internal representation, not learner-facing feedback.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Beyond the paper: because the test labels come from a transcriber who was not required to mark every error, the reported recall is best read as an optimistic bound and the reported precision as a pessimistic one; an expert-annotated sample would show how far apart they are.
Beyond the paper: the temperature-and-top-$k$ recipe could be applied to any CTC-based sequence model whose outputs are overconfident and which lacks a validation set, turning binary accept/reject decisions into graded confidence scores for tasks beyond pronunciation.
Beyond the paper: the word-level detection rate on a fixed list of Finland Swedish versus Sweden Swedish words could be used as a continuous accent probe, since the model flags some Sweden Swedish forms more often than average words; this is a testable extension the paper does not pursue.
Beyond the paper: the language-independence claim could be checked cheaply by running the same pipeline on another low-resource variety that has L1 audio and a small L2 read-aloud sample, and verifying that the default temperature and top-$k$ settings still balance precision and recall without tuning.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Quick take: this is a useful, honest case study for a real low-resource problem. It doesn't overclaim—actually it undercuts its own headline numbers—and the code is public, which I respect.

What's new: nobody has built an MD system for Finland Swedish before, and the trick of using verbatim orthographic transcriptions as noisy mispronunciation labels is a sensible low-cost workaround. The temperature-scaling-plus-top-k normalization is a small but real heuristic: it softens overconfident CTC outputs without needing an L2 validation set. The paper is also careful to separate character-level from word-level metrics and to test on expert recordings.

The soft spots are exactly where the stress test points. The evaluation labels are incomplete—the transcriber wasn't required to note errors and wasn't a phonetician—so the headline Recall 43.2% / Precision 29.8% may not measure mispronunciation at all. The paper admits this in Section 4.2, which is good, but the abstract still leads with those numbers. That's a framing problem. The 34-sentence expert test is genuinely independent, but too small to settle anything: 16 positive, 18 negative sentences gives a wide confidence interval. Also, the top-k normalization isn't ablated, so we don't know how much of the F1 gain comes from entropy regularization vs. the new algorithm. T and k are hand-set, though the claim that T is robust is plausible from Figure 2.

Who should read this: people building CAPT tools for low-resource varieties, and anyone who needs a pragmatic example of how to do evaluation when a proper L2 corpus doesn't exist. It deserves serious peer review—the method is reproducible, the problem is important, and the limitations are stated. For the final version, I'd want the authors to move the caveat into the abstract, add an ablation of the top-k step, and ideally either build a small expert-labeled test set bigger than 34 sentences or report confidence intervals on the proxy-based metrics.

Referee Report

3 major / 6 minor

Summary. The paper proposes a mispronunciation detection (MD) model for Finland Swedish (FS) that avoids the need for an L2 pronunciation dataset. The model is trained on 89 hours of L1 spontaneous speech (Talko3 and Aalto FS Parliament corpora) using a multilingual wav2vec 2.0 (XLS-R) backbone with maximum-entropy regularization, followed by a post-inference temperature-scaling and top-k normalization algorithm. The test set consists of 33 minutes of L2 read-aloud speech from DigiTala, with verbatim orthographic transcriptions used as proxy labels. The authors report that the proposed algorithm improves the character-level precision/recall/F1 from 17.6% / 77.5% / 28.7% (baseline) to 29.8% / 43.2% / 35.0%, and word-level from 31.3% / 85.0% / 45.8% to 42.4% / 56.1% / 48.3%. They also present a small expert validation (34 sentences) and an analysis of FS vs. Sweden Swedish pronunciation differences using Common Voice Swedish. The central claim is that the proposed method provides a practical, language-independent MD solution for low-resource language varieties without requiring any L2 pronunciation corpus.

Significance. If the evaluation can be validated, this is a useful contribution to CAPT for low-resource languages: the method requires only L1 speech data, uses standard XLS-R features, and the calibration algorithm is simple and post-hoc, with no parameters fitted to the L2 test set. The open-source repository strengthens reproducibility, and the FS/SweS diagnostic analysis is a creative use of the model as a linguistic tool. However, the current evidence does not yet establish the claimed MD accuracy because the primary test labels are non-expert, incomplete orthographic transcriptions, and the expert validation is too small to confirm the reported numbers. The contribution is therefore promising but conditional on a more rigorous evaluation or a more cautious framing of the claims.

major comments (3)

[Section 2 and Table 1] The headline evaluation in Table 1 and the abstract is computed against proxy labels from DigiTala verbatim orthographic transcriptions. According to Section 2, the transcriber was encouraged but not required to note mistakes and was not a phonetic expert, so the label set is incomplete and potentially noisy. The paper itself states in Section 4.2 and the Table 1 caption that 'the results do not reflect the actual performance of those models.' As a result, the Recall (43.2%) and Precision (29.8%) figures do not by themselves establish that the system detects mispronunciations; they measure agreement with an imperfect transcription. The relative-comparison defense supports model selection only if the label noise affects all models equally, which is plausible but not demonstrated. To make the central claim load-bearing, the authors should either add an expert-annotated L2 evaluation set (even a small one) or explicitly restrict the claim to 'performance against transcriptions' in the abstract and conclusion.
[Section 4 (expert validation)] The independent expert validation is too small to resolve the proxy-label problem. With 34 sentences (16 positive, 18 negative), the reported word-level 33% recall and 80% precision correspond to approximately 5 true positives and 1–2 false positives; the exact binomial 95% confidence interval for recall spans roughly 11% to 62%. Moreover, these 34 sentences are not sampled from the DigiTala test set used in Table 1, so they do not validate the specific numbers in the abstract. The authors should provide confidence intervals, report the full confusion matrix, and, if possible, enlarge the expert-annotated sample or draw it from the same distribution as the test set.
[Section 3.3] The proposed temperature scaling and top-k normalization introduce hyperparameters T, k, and the threshold θ=50%. The paper argues that T is insensitive for large values and that k=3 is motivated by L1 development-set statistics, but no sensitivity analysis is reported for k or θ, and the development set contains only L1 speech. Since the algorithm is the main technical contribution, the reported improvement in Table 1 could be particular to the chosen settings. A robustness analysis (e.g., varying k from 2 to 5 and θ from 30% to 70%) would strengthen the claim that the method does not require L2 tuning.

minor comments (6)

[Section 3.3] The description of step 2 ('Divide P by the scaled probability of top-1') is ambiguous because after normalization the scaled top-1 probability is always 1; please clarify whether the denominator is the original or the scaled probability.
[Table 1] The table reports Precision and Recall as whole percentages but F1 with one decimal place; use consistent decimal places for all metrics.
[Table 2] The p-values are not corrected for multiple comparisons across the three groups; please state whether the conclusions survive a multiple-comparison correction such as Bonferroni.
[Section 3.2] The notation 'β%' is used inconsistently (e.g., 'β=20%' vs. '20% entropy'); clarify what the percentage refers to and use a single notation throughout.
[Section 4.1] Reference [12] is attributed in the text as 'Otto-Ville', but the reference is to Raitolahti; please use the author's surname for consistency.
[Section 3.3 example] The worked example would benefit from a note that K remains unchanged because it is not in the top-3 set after scaling; this is not immediately obvious from the description.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: the reported MD result is not equivalent to the training or test labels by construction, and no fitted parameter is relabeled as a prediction.

full rationale

The paper's derivation chain is self-contained. The MD model is trained only on L1 orthographic transcripts (Talko3, Aalto Parliament) and post-processed with a hand-set threshold θ=50%, temperature T, and top-k normalization; k=3 is justified from the L1 development set, and T is not optimized, so no test-label-fitted parameter is relabeled as a prediction. The DigiTala test labels are the transcriber's verbatim orthographic deviations from the target; these are a proxy, and Section 4.2 concedes "Without a proper L2 pronunciation corpus, we cannot accurately evaluate the performance of the MD models." That is an external-validity limitation, not a circular reduction: the model's CTC scores are not defined in terms of those labels. The self-citation to Phan et al. [11] is motivational, and the entropy-regularization effect is independently re-demonstrated in Table 1, so it is not load-bearing. The FS/SweS Common Voice analysis is an internal consistency check on selected words, not a prediction derived from the test set. No equation equates the claimed output to an input by construction.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The system introduces no new physical entities. It depends on four hand-set hyperparameters and four domain assumptions about proxy labels, training distribution, standard variety, and word-level score meaning. The proxy-label assumption is the most consequential: without it, Table 1 does not measure mispronunciation detection.

free parameters (4)

Temperature T = 10
Hand-set to soften overconfident CTC logits; Figure 2 shows performance is stable for T > 10, so it is not optimized on the test set.
Top-k k = 3
Chosen because the padding token appears in the top-3 in 99% of dev-set frames when a non-padding label is top-1. This is a development-set selection, not a test-set fit.
Entropy regularization weight beta = 20%
Compared at 0% and 20%; 20% is used as the primary model, motivated by prior Finnish work. No explicit validation-set tuning procedure is described.
Threshold theta = 50%
Chosen as the simplest probability threshold; the authors note there is no validation set to optimize it.

assumptions (4)

domain assumption Verbatim orthographic transcriptions in DigiTala can serve as proxy labels for mispronunciations.
Section 2 test set bullets state the transcriber was encouraged but not required to note mistakes and was not a phonetic expert, so the labels are incomplete. If this proxy fails, the reported metrics do not measure MD.
domain assumption L1 spontaneous speech is a sufficient training distribution for detecting L2 read-aloud mispronunciations.
The pipeline trains only on Talko3 and Parliament L1 speech and relies on XLS-R pretraining to bridge the domain gap. This is stated as a mitigation rather than demonstrated.
domain assumption The standard Finland Swedish pronunciation target approximates the Central Nyland variety, and accepting dialect variants is desirable.
Section 3.1 first principle. This affects what the L1-trained model treats as correct and therefore shapes both training and evaluation assumptions.
domain assumption Word-level scores derived from character-level CTC outputs are meaningful without a grapheme-to-phoneme mapping.
Section 4 says character-level feedback is not reliable because there is no grapheme-phoneme mapping, yet word-level metrics are used for the application. The validity of that aggregation is not independently verified.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Mispronunciation Detection Without L2 Pronunciation Dataset in Low-Resource Setting: A Case Study in Finland Swedish." pith.science (2026). https://pith.science/paper/TYLIXYK5

@misc{pith2026250601156,
  author       = {Pith},
  title        = {Pith review of: Mispronunciation Detection Without L2 Pronunciation Dataset in Low-Resource Setting: A Case Study in Finland Swedish},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TYLIXYK5}},
  note         = {Machine review of arXiv:2506.01156}
}

read the original abstract

Mispronunciation detection (MD) models are the cornerstones of many language learning applications. Unfortunately, most systems are built for English and other major languages, while low-resourced language varieties, such as Finland Swedish (FS), lack such tools. In this paper, we introduce our MD model for FS, trained on 89 hours of first language (L1) speakers' spontaneous speech and tested on 33 minutes of L2 transcribed read-aloud speech. We trained a multilingual wav2vec 2.0 model with entropy regularization, followed by temperature scaling and top-k normalization after the inference to better adapt it for MD. The main novelty of our method lies in its simplicity, requiring minimal L2 data. The process is also language-independent, making it suitable for other low-resource languages. Our proposed algorithm allows us to balance Recall (43.2%) and Precision (29.8%), compared with the baseline model's Recall (77.5%) and Precision (17.6%).

Figures

Figures reproduced from arXiv: 2506.01156 by the authors.

**Figure 1.** Distribution of regions in the Talko3 training and development sets. Most data is from Osterbotten and Nyland. ¨ gions where FS is mainly spoken in Finland: Osterbotten, Ny- ¨ land, Aboland, ˚ Aland, and language islands (small communi- ˚ ties where Swedish is spoken in an environment dominated by a Finnish-speaking population). Most data is from Osterbotten ¨ and Nyland, noting that the dialects spoken in Osterbot… view at source ↗

**Figure 2.** XLS-R model performance per T value, with entropy β=20% and top-k with k=3. As T increased larger than 10, performance metrics did not change significantly. We do not have to optimize for T; as long as T is large enough to compensate for the peakiness of CTC output, the choice of T does not significantly affect model performance (see [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

37 extracted references · 34 canonical work pages

[1]

Introduction Pronunciation is a central learning goal for second language (L2) speakers, as it has a major impact on how well a speaker can be understood [1]. However, learning pronunciation in L2 is challenging, especially for adult learners, and therefore, ap- plications that support learning by providing feedback to L2 speakers on their pronunciation a...

work page
[2]

Mispronunciation Detection Without L2 Pronunciation Dataset in Low-Resource Setting: A Case Study in Finland Swedish

Dataset Our FS data are derived from 3 corpora: the Talko3 spo- ken corpus 2, Aalto FS Parliament Automatic Speech Recog- nition (ASR) corpus [12], and DigiTala [13]. The Talko3 is the largest FS dataset and it contains interviews where both speakers have FS as their L1. The speakers come from 5 re- 1https://github.com/aalto-speech/FinSwedish/ 2https://ww...

work page Pith review arXiv 2025
[3]

For MD, we use Recall ( T R T R+F A), Precision ( T R T R+F R), and their harmonic mean F1 as our metrics [5]

Framework and Experiment We define the correct identification of mispronunciation as true rejection (TR), the failure to detect mispronunciation as false acceptance (FA), and correct pronunciation that is misidenti- fied as mispronunciation as false rejection (FR). For MD, we use Recall ( T R T R+F A), Precision ( T R T R+F R), and their harmonic mean F1 ...

work page
[4]

The predicted labels’ proba- bilities, P , are obtained by softmax functionσ: P = σ(z/T )

Divide the logit vector z by T . The predicted labels’ proba- bilities, P , are obtained by softmax functionσ: P = σ(z/T )

work page
[5]

Divide P by the scaled probability of top-1

Determine the set top-k of labels whose probabilities are the largest k in P . Divide P by the scaled probability of top-1. top-1 will be normalized to a score of 1, and top-k will be rescaled to higher values: top-kscore = P top-1P

work page
[6]

partially correct

Replace the probabilities of top-k with the normalized top-kscore, keep the original probabilities of the other labels. Considering a simple CTC output with four labels and their probabilities: L (0.998), R (1e-3), K (1e-5), and padding token PAD (9.9e-4). This is a typical CTC output, showing overconfi- dence in L and under-estimate the probability of R....

work page
[7]

Results Our results are shown in Table 1. The baseline model ( β=0%, T =0) has very low Precision, which is expected given the diver- sity of FS dialects and the limited amount of data available to adequately capture this variability. This limitation is evidenced by Otto-Ville [12] in his Talko3 region-wise L1 ASR perfor- mance analysis, with word error r...

work page
[8]

We relied entirely on L1 data and tested on a minimal L2 ASR corpus

Conclusion In this paper, we detailed our pipeline and algorithm for devel- oping MD models for a low-resource dialect. We relied entirely on L1 data and tested on a minimal L2 ASR corpus. Our model was able to detect some significant differences between FS and SweS while recognizing the variations within FS dialects. How- ever, there were cases where the...

work page

Show all 37 references

[9]

Technology-enhanced foreign and second-language learning of Nordic languages

Acknowledgements We would like to thank the following projects and funding agen- cies: NordForsk through the funding to “Technology-enhanced foreign and second-language learning of Nordic languages” (project number 103893); Research Council of Finland through the funding to “D...

2023
[10]

J. M. Levis, Intelligibility, oral communication, and the teaching of pronunciation. Cambridge University Press, 2018

2018
[11]

Kautonen and M

M. Kautonen and M. Kuronen, Uttalsinl¨arning med fokus p˚a sven- ska. Svenska litteraturs ¨allskapet i Finland, 2021

2021
[12]

Ideologies of standardisation: Finland Swedish and Swedish-language Finland,

J.-O. ¨Ostman and L. Mattfolk, “Ideologies of standardisation: Finland Swedish and Swedish-language Finland,” Standard lan- guages and language standards in a changing Europe, vol. 1, pp. 75–82, 2011

2011
[13]

Swedish quantity: Cen- tral Standard Swedish and Fenno-Swedish,

P. Helgason, C. Ringen, and K. Suomi, “Swedish quantity: Cen- tral Standard Swedish and Fenno-Swedish,”Journal of Phonetics, vol. 41, no. 6, pp. 534–545, 2013

2013
[14]

Mispronunciation detection and diagnosis using deep neural networks: A systematic review,

M. Lounis, B. Dendani, and H. Bahi, “Mispronunciation detection and diagnosis using deep neural networks: A systematic review,” Multimedia Tools and Applications, vol. 83, no. 23, pp. 62 793– 62 827, Jul. 2024

2024
[15]

L2-ARCTIC: A non-native English speech corpus,

G. Zhao, S. Sonsaat, A. Silpachai, I. Lucic, E. Chukharev- Hudilainen, J. Levis, and R. Gutierrez-Osuna, “L2-ARCTIC: A non-native English speech corpus,” in Interspeech 2018. ISCA, Sep. 2018, pp. 2783–2787

2018
[16]

speechocean762: An open-source non- native English speech corpus for pronunciation assessment,

J. Zhang, Z. Zhang, Y . Wang, Z. Yan, Q. Song, Y . Huang, K. Li, D. Povey, and Y . Wang, “speechocean762: An open-source non- native English speech corpus for pronunciation assessment,” in Interspeech 2021, 2021, pp. 3710–3714

2021
[17]

Phonetic-level mis- pronunciation detection in non-native Swedish speech,

P. Langlais, A.-M. ¨Oster, and B. Granstr¨om, “Phonetic-level mis- pronunciation detection in non-native Swedish speech,” in5th In- ternational Conference on Spoken Language Processing (ICSLP 1998), 1998, p. paper 0311

1998
[18]

Developing a benchmark for pronun- ciation feedback: Creation of a phonemically annotated speech corpus of isiZulu language learner speech,

A. O’Neil, N. Hjortnaes, F. Tyers, Z. Nkosi, T. Ndlovu, Z. Mlondo, and N. P. Pewa, “Developing a benchmark for pronun- ciation feedback: Creation of a phonemically annotated speech corpus of isiZulu language learner speech,” in Proceedings of LREC-COLING 2024. ELRA and ICCL, M...

2024
[19]

Mispronunciation detection without nonna- tive training data,

A. Lee and J. Glass, “Mispronunciation detection without nonna- tive training data,” in Interspeech 2015, 2015, pp. 643–647

2015
[20]

CaptainA - a mobile app for practising Finnish pronunciation,

N. Phan, T. Gr ´osz, and M. Kurimo, “CaptainA - a mobile app for practising Finnish pronunciation,” in Proceedings of the 24th Nordic Conference on Computational Linguistics (NoDaLiDa) . University of Tartu Library, May 2023, pp. 265–270

2023
[21]

Finland Swedish automatic speech recogni- tion,

O.-V . Raitolahti, “Finland Swedish automatic speech recogni- tion,” Master’s thesis, Aalto University, 2022

2022
[22]

Automatic speaking assessment of spontaneous L2 Finnish and Swedish,

R. Al-Ghezi, K. V oskoboinik, Y . Getman, A. V on Zansen, H. Kallio, M. Kurimo, A. Huhta, and R. Hild ´en, “Automatic speaking assessment of spontaneous L2 Finnish and Swedish,” Language Assessment Quarterly , vol. 20, no. 4-5, pp. 421–444, 2023

2023
[23]

Ivars, Dialekter och sm ˚astadsspr˚ak, ser

A.-M. Ivars, Dialekter och sm ˚astadsspr˚ak, ser. Svenskan i Fin- land – i dag och i g ˚ar I:1. Skrifter utgivna av Svenska litter- aturs¨allskapet i Finland (798), 2015

2015
[24]

Changing contexts and shifting paradigms in pro- nunciation teaching,

J. M. Levis, “Changing contexts and shifting paradigms in pro- nunciation teaching,” TESOL quarterly, vol. 39, no. 3, pp. 369– 377, 2005

2005
[25]

L. F. Bachman, Fundamental Considerations in Language Test- ing. Oxford university press, 1990

1990
[26]

An overview of spoken language technology for education,

M. Eskenazi, “An overview of spoken language technology for education,” Speech Communication, vol. 51, no. 10, pp. 832–844, Oct. 2009

2009
[27]

Corrective feedback and teacher development,

R. Ellis, “Corrective feedback and teacher development,” L2 Jour- nal: An electronic refereed journal for foreign and second lan- guage educators, vol. 1, no. 1, 2009

2009
[28]

wav2vec 2.0: A framework for self-supervised learning of speech repre- sentations,

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, pp. 12 449–12 460, 2020

2020
[29]

XLS-R: Self-supervised cross-lingual speech representation learning at scale,

A. Babu, C. Wang, A. Tjandra, K. Lakhotia, Q. Xu, N. Goyal, K. Singh, P. von Platen, Y . Saraf, J. Pino, A. Baevski, A. Con- neau, and M. Auli, “XLS-R: Self-supervised cross-lingual speech representation learning at scale,” in Interspeech 2022, 2022, pp. 2278–2282

2022
[30]

A study on fine-tuning wav2vec2.0 model for the task of mispronunciation detection and diagnosis,

L. Peng, K. Fu, B. Lin, D. Ke, and J. Zhan, “A study on fine-tuning wav2vec2.0 model for the task of mispronunciation detection and diagnosis,” in Interspeech 2021, 2021, pp. 4448–4452

2021
[31]

CTC-segmentation of large corpora for German end-to-end speech recognition,

L. K ¨urzinger, D. Winkelbauer, L. Li, T. Watzel, and G. Rigoll, “CTC-segmentation of large corpora for German end-to-end speech recognition,” in International Conference on Speech and Computer. Springer, 2020, pp. 267–278

2020
[32]

Con- nectionist temporal classification: Labelling unsegmented se- quence data with recurrent neural networks,

A. Graves, S. Fern ´andez, F. Gomez, and J. Schmidhuber, “Con- nectionist temporal classification: Labelling unsegmented se- quence data with recurrent neural networks,” in Proceedings of the 23rd international conference on Machine learning, 2006, pp. 369–376

2006
[33]

Connectionist temporal classifica- tion with maximum entropy regularization,

H. Liu, S. Jin, and C. Zhang, “Connectionist temporal classifica- tion with maximum entropy regularization,” Advances in Neural Information Processing Systems, vol. 31, 2018

2018
[34]

Explore wav2vec 2.0 for mispronunciation detection,

X. Xu, Y . Kang, S. Cao, B. Lin, and L. Ma, “Explore wav2vec 2.0 for mispronunciation detection,” in Interspeech 2021, 2021, pp. 4428–4432

2021
[35]

Phone-level pronunciation scoring and assessment for interactive language learning,

S. M. Witt and S. J. Young, “Phone-level pronunciation scoring and assessment for interactive language learning,” Speech com- munication, vol. 30, no. 2-3, pp. 95–108, 2000

2000
[36]

On calibra- tion of modern neural networks,

C. Guo, G. Pleiss, Y . Sun, and K. Q. Weinberger, “On calibra- tion of modern neural networks,” in International conference on machine learning. PMLR, 2017, pp. 1321–1330

2017
[37]

Common voice: A massively-multilingual speech corpus,

R. Ardila, M. Branson, K. Davis, M. Henretty, M. Kohler, J. Meyer, R. Morais, L. Saunders, F. M. Tyers, and G. Weber, “Common voice: A massively-multilingual speech corpus,” in Proceedings of the 12th Conference on Language Resources and Evaluation (LREC 2020), 2020, pp. 4211–4215

2020

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

Introduction Pronunciation is a central learning goal for second language (L2) speakers, as it has a major impact on how well a speaker can be understood [1]. However, learning pronunciation in L2 is challenging, especially for adult learners, and therefore, ap- plications that support learning by providing feedback to L2 speakers on their pronunciation a...

work page

[2] [2]

Mispronunciation Detection Without L2 Pronunciation Dataset in Low-Resource Setting: A Case Study in Finland Swedish

Dataset Our FS data are derived from 3 corpora: the Talko3 spo- ken corpus 2, Aalto FS Parliament Automatic Speech Recog- nition (ASR) corpus [12], and DigiTala [13]. The Talko3 is the largest FS dataset and it contains interviews where both speakers have FS as their L1. The speakers come from 5 re- 1https://github.com/aalto-speech/FinSwedish/ 2https://ww...

work page Pith review arXiv 2025

[3] [3]

For MD, we use Recall ( T R T R+F A), Precision ( T R T R+F R), and their harmonic mean F1 as our metrics [5]

Framework and Experiment We define the correct identification of mispronunciation as true rejection (TR), the failure to detect mispronunciation as false acceptance (FA), and correct pronunciation that is misidenti- fied as mispronunciation as false rejection (FR). For MD, we use Recall ( T R T R+F A), Precision ( T R T R+F R), and their harmonic mean F1 ...

work page

[4] [4]

The predicted labels’ proba- bilities, P , are obtained by softmax functionσ: P = σ(z/T )

Divide the logit vector z by T . The predicted labels’ proba- bilities, P , are obtained by softmax functionσ: P = σ(z/T )

work page

[5] [5]

Divide P by the scaled probability of top-1

Determine the set top-k of labels whose probabilities are the largest k in P . Divide P by the scaled probability of top-1. top-1 will be normalized to a score of 1, and top-k will be rescaled to higher values: top-kscore = P top-1P

work page

[6] [6]

partially correct

Replace the probabilities of top-k with the normalized top-kscore, keep the original probabilities of the other labels. Considering a simple CTC output with four labels and their probabilities: L (0.998), R (1e-3), K (1e-5), and padding token PAD (9.9e-4). This is a typical CTC output, showing overconfi- dence in L and under-estimate the probability of R....

work page

[7] [7]

Results Our results are shown in Table 1. The baseline model ( β=0%, T =0) has very low Precision, which is expected given the diver- sity of FS dialects and the limited amount of data available to adequately capture this variability. This limitation is evidenced by Otto-Ville [12] in his Talko3 region-wise L1 ASR perfor- mance analysis, with word error r...

work page

[8] [8]

We relied entirely on L1 data and tested on a minimal L2 ASR corpus

Conclusion In this paper, we detailed our pipeline and algorithm for devel- oping MD models for a low-resource dialect. We relied entirely on L1 data and tested on a minimal L2 ASR corpus. Our model was able to detect some significant differences between FS and SweS while recognizing the variations within FS dialects. How- ever, there were cases where the...

work page