Sometin Beta Pass Notin (SBPN): Improving Multilingual ASR for Nigerian Languages via Knowledge Distillation

Sewade Ogun

arxiv: 2605.17710 · v1 · pith:QOWGFSZFnew · submitted 2026-05-18 · 💻 cs.CL · eess.AS

Sometin Beta Pass Notin (SBPN): Improving Multilingual ASR for Nigerian Languages via Knowledge Distillation

Sewade Ogun This is my paper

Pith reviewed 2026-05-19 22:01 UTC · model grok-4.3

classification 💻 cs.CL eess.AS

keywords multilingual ASRknowledge distillationNigerian languageslow-resource speech recognitionpseudo-labelingword error rateself-improvement

0 comments

The pith

Knowledge distillation from monolingual models followed by self-improvement on pseudo labels improves multilingual ASR for Nigerian languages.

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

The paper develops a two-stage process to build better speech recognition systems for languages like Yoruba, Hausa, Igbo, Nigerian Pidgin, and Nigerian English. It starts by distilling knowledge from existing monolingual models using language-specific N-gram models, then refines the student model iteratively with its own pseudo-labeled data. This approach aims to close the performance gap that current multilingual systems show compared to high-resource languages, which matters because it could make voice technology more accessible in regions with scarce labeled speech data. A sympathetic reader would care if this leads to more accurate transcription of tonal languages and code-switched speech without needing massive new datasets.

Core claim

The authors present Sometin Beta Pass Notin (SBPN), a multilingual ASR model trained via student-teacher knowledge distillation conditioned on robust language-specific N-gram language models, followed by iterative self-improvement using pseudo-labelled data. This yields an average relative Word Error Rate reduction of 29% over monolingual baselines and outperforms state-of-the-art multilingual models on benchmarks like Common Voice and Fleurs. The framework covers Yoruba, Hausa, Igbo, Nigerian Pidgin, and Nigerian English, released in Base (120M) and Large (600M) parameter versions as open foundation models.

What carries the argument

Two-stage distillation process consisting of student-teacher knowledge distillation from monolingual models conditioned on N-gram language models, followed by iterative self-improvement using pseudo-labelled data.

If this is right

SBPN models achieve better accuracy than previous monolingual and multilingual systems on Nigerian language benchmarks.
The approach bridges the gap for low-resource languages with challenges like tonal diacritics and code-switching.
Open release of SBPN-Base and SBPN-Large enables further research into phonetic and cultural aspects of the region.
Performance gains hold across major benchmarks including Common Voice and Fleurs.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

This method might generalize to other low-resource language families facing similar data scarcity and orthographic issues.
Future work could test if combining this with larger unlabeled audio corpora further reduces the need for manual transcription.
Voice applications in healthcare or education in Nigeria could benefit from these improved recognition rates.

Load-bearing premise

The iterative self-improvement works only if the pseudo-labels from the first model are accurate enough to avoid compounding errors in later rounds.

What would settle it

A test showing that running the second stage on the pseudo-labeled data increases word error rate on a held-out evaluation set compared to stopping after the first stage would falsify the benefit of the iterative refinement.

Figures

Figures reproduced from arXiv: 2605.17710 by Sewade Ogun.

**Figure 1.** Figure 1: Flow diagram showing the pseudo-label generation pipeline from unprocessed audio data to processed audio segments with pseudo-labels Next, all long segments were split into 30 s segments using a silence threshold of −50 dB. This ensures that the segments are compatible with ASR models used for pseudo-labelling. The total number of hours after processing was about 10000 h. Gigaspeech (Chen et al., 2021) was… view at source ↗

**Figure 2.** Figure 2: Comparing WER (%) on the validation sets of Hausa (ha), Igbo (ig), Yor`ub´a (yo), and Nigerian Pidgin (pcm) when different CTC decoder libraries are used. ha, ig, and yo were evaluated on Fleurs while pcm was evaluated on the Nigerian Pidgin validation set the Nigerian English subset of the International Corpus of English (ICE)18 The ICE dataset contains texts extracted from Nigerian media and newspapers, … view at source ↗

**Figure 3.** Figure 3: Performance of SBPN-Large on test set samples across several speaking rates (0.8x to 2x). Average WER (%) computed on the Fleurs test sets and Nigerian pidgin test set. needs to be done on this front to reduce the WER gap between accented and unaccented predictions in Yor`ub´a language. For the Igbo language, the gap is not as large as that of the Yor`ub´a language. Tones are often omitted in standard Igbo… view at source ↗

**Figure 4.** Figure 4: Average WER (%) of SBPN and Teacher models on the Fleurs test sets before and after removing diacritical marks from Yor`ub´a and Igbo predicted texts. The teacher models are the monolingual baselines [PITH_FULL_IMAGE:figures/full_fig_p014_4.png] view at source ↗

read the original abstract

Although modern multilingual Automatic Speech Recognition (ASR) systems support several Nigerian languages, their performance consistently lags behind high-resource languages like English and French. Nigerian languages present unique modelling hurdles, including acute data scarcity, inconsistent orthography, tonal diacritics, diverse accents, frequent code-switching, and localized named entities. To address these challenges, we developed a multilingual ASR framework utilizing a two-stage distillation process. First, we employ student-teacher knowledge distillation from existing monolingual models, conditioned on robust language-specific N-gram language models. Second, we perform iterative self improvement using pseudo-labelled data to further refine accuracy. Our method significantly bridges the performance gap, achieving on average a relative Word Error Rate (WER) reduction of 29 % over monolingual baselines. Our models also outperform state-of-the-art multilingual models across major benchmarks, including Common Voice and Fleurs. We introduce Sometin Beta Pass Notin (SBPN), a foundational multilingual ASR model covering Yor\`ub\'a, Hausa, Igbo, Nigerian Pidgin, and Nigerian English. SBPN is released in two sizes: SBPN-Base (120 M parameters) and SBPN-Large (600 M parameters). By releasing these as open foundation models, we aim to provide ASR resources for further research into the rich phonetic and cultural landscape of the region.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper applies standard distillation plus pseudo-label self-training to five Nigerian languages and releases the resulting models, but the reported 29% WER gain rests on thin experimental details.

read the letter

The main takeaway is that this work takes knowledge distillation from monolingual teachers plus an iterative self-training stage and applies it to Yoruba, Hausa, Igbo, Nigerian Pidgin, and Nigerian English. They release both a 120M and 600M parameter model and claim the approach cuts WER by 29% relative to monolingual baselines while beating existing multilingual systems on Common Voice and Fleurs. That combination and the open release are the concrete new pieces here. The authors correctly flag the real difficulties these languages pose—tonal diacritics, code-switching, localized entities, and data scarcity—and the two-stage pipeline is a reasonable way to leverage existing monolingual resources. Releasing the models openly is the part that actually helps other researchers. The soft spots sit in the experimental support. The abstract gives headline numbers but leaves out dataset sizes, how the baselines were trained, statistical significance, or error bars. The second stage depends on the first-stage model producing usable pseudo labels; in tonal or code-switched data any consistent errors on tones or names could get reinforced, and the description does not mention filtering, thresholding, or an ablation that isolates the iterative step. Those gaps make the size of the improvement hard to judge from the given information. This paper is for people building or evaluating ASR for low-resource African languages who need starting checkpoints or benchmark numbers. It is coherent enough and grounded enough in public datasets that a serious editor should send it to peer review rather than desk-reject it. The claims are testable once the full experimental section and code are examined.

Referee Report

2 major / 2 minor

Summary. The manuscript introduces SBPN, a multilingual ASR framework for five Nigerian languages (Yorùbá, Hausa, Igbo, Nigerian Pidgin, Nigerian English) that employs a two-stage knowledge distillation process. The first stage distills from monolingual teacher models conditioned on language-specific N-gram LMs; the second stage performs iterative self-improvement on pseudo-labeled data. The central empirical claims are an average 29% relative WER reduction versus monolingual baselines and outperformance of existing SOTA multilingual models on Common Voice and Fleurs, with open release of 120 M and 600 M parameter models.

Significance. If the reported gains are robustly validated, the work would supply useful open foundation models for ASR in low-resource languages that exhibit tonal diacritics, code-switching, and orthographic inconsistency. The public release of both model sizes would enable downstream research on Nigerian-language speech technology.

major comments (2)

[Experimental Evaluation] Experimental section: the headline 29 % relative WER reduction and the claim of outperforming SOTA multilingual models on Common Voice and Fleurs are presented without dataset sizes, exact baseline implementations, statistical significance tests, or error bars. These omissions make it impossible to determine whether the central performance claims are supported by the data or sensitive to experimental choices.
[Method Description] Method section (iterative self-improvement stage): the performance delta is stated to depend on refinement using pseudo-labels generated by the first-stage model. No description of pseudo-label filtering, confidence thresholding, or an ablation that isolates the contribution of the iterative stage is provided. In the presence of tonal and code-switched data this assumption is load-bearing for the reported gains.

minor comments (2)

[Abstract] The abstract introduces the acronym SBPN without explaining its etymology or intended meaning.
[References] Ensure that all benchmark citations (Common Voice, Fleurs) are accompanied by complete references in the bibliography.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive and detailed feedback. We have reviewed the comments carefully and agree that additional experimental details and methodological clarifications will strengthen the manuscript. We address each major comment below and will incorporate the necessary revisions.

read point-by-point responses

Referee: [Experimental Evaluation] Experimental section: the headline 29 % relative WER reduction and the claim of outperforming SOTA multilingual models on Common Voice and Fleurs are presented without dataset sizes, exact baseline implementations, statistical significance tests, or error bars. These omissions make it impossible to determine whether the central performance claims are supported by the data or sensitive to experimental choices.

Authors: We agree that the experimental section requires more detail to substantiate the central claims. In the revised manuscript, we will add the training and evaluation dataset sizes for each of the five languages, precise specifications of the monolingual teacher models (including architectures, pre-training data, and N-gram LM integration), and the exact configurations of the SOTA multilingual baselines from Common Voice and Fleurs. We will also report error bars from multiple random seeds and include statistical significance testing (e.g., paired bootstrap or t-tests) for the reported 29% relative WER reduction and outperformance results. revision: yes
Referee: [Method Description] Method section (iterative self-improvement stage): the performance delta is stated to depend on refinement using pseudo-labels generated by the first-stage model. No description of pseudo-label filtering, confidence thresholding, or an ablation that isolates the contribution of the iterative stage is provided. In the presence of tonal and code-switched data this assumption is load-bearing for the reported gains.

Authors: We acknowledge that the iterative self-improvement stage needs a more explicit description, especially given the tonal and code-switching characteristics of the languages. In the revised version, we will expand the method section to detail the pseudo-label generation process, including confidence thresholding (based on token-level probabilities and alignment scores) and any filtering steps to mitigate errors from tonal diacritics or code-switched segments. We will also add an ablation study comparing first-stage distillation performance against the full two-stage model to isolate the iterative stage's contribution. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical performance gains on external benchmarks

full rationale

The paper presents an empirical two-stage knowledge distillation pipeline for multilingual ASR, with results consisting of measured WER reductions (29% relative) against monolingual baselines and SOTA models on independent benchmarks (Common Voice, Fleurs). No equations, first-principles derivations, or predictions are claimed that reduce by construction to fitted parameters or self-referential inputs inside the paper. The iterative pseudo-labeling stage is a standard semi-supervised technique whose validity is assessed via external test sets rather than being tautological; success depends on data and training outcomes, not definitional equivalence. No self-citations, uniqueness theorems, or ansatzes serve as load-bearing steps that collapse the central claim into its own assumptions.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The work relies on standard assumptions of knowledge distillation and self-training in ASR; no new physical entities or ad-hoc mathematical axioms are introduced. The main unstated premises are that the teacher monolingual models are already high-quality and that N-gram language models provide useful conditioning without introducing bias.

free parameters (1)

model size (120M / 600M)
Chosen architecture scales; the paper does not derive these sizes from first principles.

axioms (1)

domain assumption Monolingual ASR teachers provide useful supervisory signal for multilingual student
Invoked in the first stage of distillation; treated as given rather than proven in the abstract.

pith-pipeline@v0.9.0 · 5770 in / 1360 out tokens · 30062 ms · 2026-05-19T22:01:06.027531+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Cost/FunctionalEquation washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

two-stage distillation process. First, we employ student-teacher knowledge distillation from existing monolingual models, conditioned on robust language-specific N-gram language models. Second, we perform iterative self improvement using pseudo-labelled data
IndisputableMonolith/Foundation/RealityFromDistinction reality_from_one_distinction unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

SBPN improves transcription accuracy on several Nigerian languages relative to existing monolingual baselines

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

31 extracted references · 31 canonical work pages

[1]

doi: 10.21437/Interspeech.2023-466. Josh Meyer, David Adelani, Edresson Casanova, Alp ¨Oktem, Daniel Whitenack, Julian Weber, Salomon KABONGO KABENAMUALU, Elizabeth Salesky, Iroro Orife, Colin Leong, Perez Ogayo, Chris Chinenye Emezue, Jonathan Mukiibi, Salomey Osei, Apelete AGBOLO, Victor Akinode, Bernard Opoku, Olanrewaju Samuel, Jesujoba Alabi, and Sha...

work page doi:10.21437/interspeech.2023-466 2023
[2]

Hellina Hailu Nigatu, Atnafu Lambebo Tonja, Benjamin Rosman, Thamar Solorio, and Monojit Choudhury

European Language Resources Association. Hellina Hailu Nigatu, Atnafu Lambebo Tonja, Benjamin Rosman, Thamar Solorio, and Monojit Choudhury. The Zeno’s paradox of ‘low-resource’ languages. In Yaser Al- Onaizan, Mohit Bansal, and Yun-Nung Chen, editors,Proceedings of the 2024 Con- ference on Empirical Methods in Natural Language Processing, pages 17753–177...

work page doi:10.18653/v1/2024.emnlp-main.983 2024
[3]

20 Sometin Beta Pass Notin (SBPN) Appendix A

doi: 10.1109/TASLP.2023.3306709. 20 Sometin Beta Pass Notin (SBPN) Appendix A. List of Pidgin English variants. Phrases and words on the left were replaced with phrases and words on the right during training

work page doi:10.1109/taslp.2023.3306709 2023
[4]

beforebefore→bifor bifor

work page
[5]

chewing gum→ chingum

work page
[6]

everybody’s→evri- bodi’s

work page
[7]

gavernments→gov- ernments

work page
[8]

gobernment→govern- ment

work page
[9]

gouverment→govern- ment

work page
[10]

gouvernment→gov- ernment

work page
[11]

govaenment→govern- ment

work page
[12]

govenment→govern- ment

work page
[13]

goverment→govern- ment

work page
[14]

goverments→govern- ments

work page
[15]

governmen→govern- ment

work page
[16]

governmet→govern- ment

work page
[17]

governmint→govern- ment

work page
[18]

pickin’s→pikins 22 Sometin Beta Pass Notin (SBPN)

work page
[19]

tomorrow’s→to- moro’s

work page
[20]

waiting dey→wetin dey

work page
[21]

wan welcome to→wan welcome

work page
[22]

we de for→wey dey for

work page
[23]

List of Pidgin words with their homophones considered for replacements using the N-gram Pidgin language model

yu→you Appendix B. List of Pidgin words with their homophones considered for replacements using the N-gram Pidgin language model

work page
[24]

becoming — become hin

work page
[25]

convex — con vex — com vex

work page
[26]

fellow — folo — follow

work page
[27]

hear — here — hia — ear

work page
[28]

tory — tori — touring — thory

work page
[29]

way — wey — we — whey

work page
[30]

wear — wia — were — where

work page
[31]

yo — you 24 Sometin Beta Pass Notin (SBPN) Appendix C. Table showing the hyper-parameters selected for each model variant of SBPN Hyperparameter SBPN-Base SBPN-Large Base learning rate 1e−4 3e−4 Self improvement learning rate 1e−5 1e−5 Number of layers 17 24 No. of pred. RNN layers 1 2 Encoder feature dimension 512 1024 No. of attention heads 8 8 Weight d...

work page

[1] [1]

doi: 10.21437/Interspeech.2023-466. Josh Meyer, David Adelani, Edresson Casanova, Alp ¨Oktem, Daniel Whitenack, Julian Weber, Salomon KABONGO KABENAMUALU, Elizabeth Salesky, Iroro Orife, Colin Leong, Perez Ogayo, Chris Chinenye Emezue, Jonathan Mukiibi, Salomey Osei, Apelete AGBOLO, Victor Akinode, Bernard Opoku, Olanrewaju Samuel, Jesujoba Alabi, and Sha...

work page doi:10.21437/interspeech.2023-466 2023

[2] [2]

Hellina Hailu Nigatu, Atnafu Lambebo Tonja, Benjamin Rosman, Thamar Solorio, and Monojit Choudhury

European Language Resources Association. Hellina Hailu Nigatu, Atnafu Lambebo Tonja, Benjamin Rosman, Thamar Solorio, and Monojit Choudhury. The Zeno’s paradox of ‘low-resource’ languages. In Yaser Al- Onaizan, Mohit Bansal, and Yun-Nung Chen, editors,Proceedings of the 2024 Con- ference on Empirical Methods in Natural Language Processing, pages 17753–177...

work page doi:10.18653/v1/2024.emnlp-main.983 2024

[3] [3]

20 Sometin Beta Pass Notin (SBPN) Appendix A

doi: 10.1109/TASLP.2023.3306709. 20 Sometin Beta Pass Notin (SBPN) Appendix A. List of Pidgin English variants. Phrases and words on the left were replaced with phrases and words on the right during training

work page doi:10.1109/taslp.2023.3306709 2023

[4] [4]

beforebefore→bifor bifor

work page

[5] [5]

chewing gum→ chingum

work page

[6] [6]

everybody’s→evri- bodi’s

work page

[7] [7]

gavernments→gov- ernments

work page

[8] [8]

gobernment→govern- ment

work page

[9] [9]

gouverment→govern- ment

work page

[10] [10]

gouvernment→gov- ernment

work page

[11] [11]

govaenment→govern- ment

work page

[12] [12]

govenment→govern- ment

work page

[13] [13]

goverment→govern- ment

work page

[14] [14]

goverments→govern- ments

work page

[15] [15]

governmen→govern- ment

work page

[16] [16]

governmet→govern- ment

work page

[17] [17]

governmint→govern- ment

work page

[18] [18]

pickin’s→pikins 22 Sometin Beta Pass Notin (SBPN)

work page

[19] [19]

tomorrow’s→to- moro’s

work page

[20] [20]

waiting dey→wetin dey

work page

[21] [21]

wan welcome to→wan welcome

work page

[22] [22]

we de for→wey dey for

work page

[23] [23]

List of Pidgin words with their homophones considered for replacements using the N-gram Pidgin language model

yu→you Appendix B. List of Pidgin words with their homophones considered for replacements using the N-gram Pidgin language model

work page

[24] [24]

becoming — become hin

work page

[25] [25]

convex — con vex — com vex

work page

[26] [26]

fellow — folo — follow

work page

[27] [27]

hear — here — hia — ear

work page

[28] [28]

tory — tori — touring — thory

work page

[29] [29]

way — wey — we — whey

work page

[30] [30]

wear — wia — were — where

work page

[31] [31]

yo — you 24 Sometin Beta Pass Notin (SBPN) Appendix C. Table showing the hyper-parameters selected for each model variant of SBPN Hyperparameter SBPN-Base SBPN-Large Base learning rate 1e−4 3e−4 Self improvement learning rate 1e−5 1e−5 Number of layers 17 24 No. of pred. RNN layers 1 2 Encoder feature dimension 512 1024 No. of attention heads 8 8 Weight d...

work page