Paper deep dive
Mitigating Structural Noise in Low-Resource S2TT: An Optimized Cascaded Nepali-English Pipeline with Punctuation Restoration
Tangsang Chongbang, Pranesh Pyara Shrestha, Amrit Sarki, Anku Jaiswal
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 95%
Last extracted: 7/20/2026, 12:29:14 PM
Summary
This paper addresses structural noise in low-resource cascaded Speech-to-Text Translation (S2TT) systems, specifically for Nepali-to-English. The authors propose an optimized pipeline incorporating a Punctuation Restoration Module (PRM) using a multilingual T5 (mT5) model between the Automatic Speech Recognition (ASR) stage (Wav2Vec2-XLS-R-300m) and the Neural Machine Translation (NMT) stage (MarianMT). The study demonstrates that unpunctuated ASR output causes a significant 20.7% relative BLEU drop, while the optimized pipeline with PRM achieves a 4.90 BLEU point gain over the baseline, validated by human assessment metrics for Adequacy and Fluency.
Entities (12)
Relation Signals (9)
Wav2Vec2-XLS-R-300m â achievesperformanceon â OpenSLR-54
confidence 95% ¡ a Wav2Vec2-XLS-R-300m model achieved a state-of-the-art 2.72% CER on OpenSLR-54
MarianMT â achievesperformanceon â FLORES-200
confidence 95% ¡ a multi-stage fine-tuned MarianMT model reached a 28.32 BLEU score on the FLORES-200 benchmark
ASR â feedsinto â Punctuation Restoration Module
confidence 95% ¡ applied the PRM directly to ASR output
Punctuation Restoration Module â feedsinto â NMT
confidence 95% ¡ The unpunctuated ASR output is passed through a preliminary mT5 model... The transcribed text is fed into the fine-tuned MarianMT
Punctuation Restoration Module â ispartofpipeline â S2TT
confidence 95% ¡ propose and evaluate an intermediate Punctuation Restoration Module (PRM)... The final S2TT pipeline
MarianMT â performstask â NMT
confidence 95% ¡ MarianMT NMT model
Wav2Vec2-XLS-R-300m â performstask â ASR
confidence 95% ¡ Wav2Vec2.0 XLS-R 300m ASR model
Punctuation Restoration Module â usesmodel â mT5
confidence 95% ¡ a Punctuation Restoration Module (PRM) using multilingual T5 (mT5) model
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Cascaded speech-to-text translation (S2TT) systems for low-resource languages can suffer from structural noise, particularly the loss of punctuation during the Automatic Speech Recognition (ASR) phase. This research investigates the impact of such noise on Nepali-to-English translation and proposes an optimized pipeline to mitigate quality degradation. We first establish highly proficient ASR and NMT components: a Wav2Vec2-XLS-R-300m model achieved a state-of-the-art 2.72% CER on OpenSLR-54, and a multi-stage fine-tuned MarianMT model reached a 28.32 BLEU score on the FLORES-200 benchmark. We empirically investigate the influence of punctuation loss, demonstrating that unpunctuated ASR output significantly degrades translation quality, causing a massive 20.7% relative BLEU drop on the FLORES benchmark. To overcome this, we propose and evaluate an intermediate Punctuation Restoration Module (PRM). The final S2TT pipeline was tested across three configurations on a custom dataset. The optimal configuration, which applied the PRM directly to ASR output, achieved a 4.90 BLEU point gain over the direct ASR-to-NMT baseline (BLEU 36.38 vs. 31.48). This improvement was validated by human assessment, which confirmed the optimized pipeline's superior Adequacy (3.673) and Fluency (3.804) with inter-rater reliability (Krippendorff's ${\alpha} {\geq}$ 0.723). This work validates that targeted punctuation restoration is the most effective intervention for mitigating structural noise in the Nepali S2TT pipeline. It establishes an optimized baseline and demonstrates a critical architectural insight for developing cascaded speech translation systems for similar low-resource languages.
Tags
Links
- Source: https://arxiv.org/abs/2602.21647v2
- Canonical: https://arxiv.org/abs/2602.21647v2
Trouble viewing inline? Open PDF directly â
Full Text
49,098 characters extracted from source content.
Expand or collapse full text
Mitigating Structural Noise in Low-Resource S2T: An Optimized Cascaded Nepali-English Pipeline with Punctuation Restoration TANGSANG CHONGBANG, PRANESH PYARA SHRESTHA, AMRIT SARKI, and ANKU JAISWAL, De- partment of Electronics and Computer Engineering, Institute of Engineering, Pulchowk Campus, Nepal Cascaded speech-to-text translation (S2T) systems for low-resource languages can suffer from structural noise, particularly the loss of punctuation during the Automatic Speech Recognition (ASR) phase. This research investigates the impact of such noise on Nepali-to-English translation and proposes an optimized pipeline to mitigate quality degradation. We first establish highly proficient ASR and NMT components: a Wav2Vec2-XLS-R-300m model achieved a state-of-the-art 2.72% CER on OpenSLR-54, and a multi-stage fine-tuned MarianMT model reached a 28.32 BLEU score on the FLORES-200 benchmark. We empirically investigate the influence of punctuation loss, demonstrating that unpunctuated ASR output significantly degrades translation quality, causing a massive 20.7% relative BLEU drop on the FLORES benchmark. To overcome this, we propose and evaluate an intermediate Punctuation Restoration Module (PRM). The final S2T pipeline was tested across three configurations on a custom dataset. The optimal configuration, which applied the PRM directly to ASR output, achieved a 4.90 BLEU point gain over the direct ASR-to-NMT baseline (BLEU 36.38 vs. 31.48). This improvement was validated by human assessment, which confirmed the optimized pipelineâs superior Adequacy (3.673) and Fluency (3.804) with inter-rater reliability (Krippendorffâsíź âĽ0.723). This work validates that targeted punctuation restoration is the most effective intervention for mitigating structural noise in the Nepali S2T pipeline. It establishes an optimized baseline and demonstrates a critical architectural insight for developing cascaded speech translation systems for similar low-resource languages. CCS Concepts:⢠Computing methodologiesâSpeech recognition; Machine translation; Language resources; Natural language generation. Additional Key Words and Phrases: Speech-to-Text Translation, Low-Resource Languages, Nepali, ASR, NMT, Nepali-to-English, Punctuation Restoration, Wav2Vec2, MarianMT, Cascaded Pipeline. ACM Reference Format: Tangsang Chongbang, Pranesh Pyara Shrestha, Amrit Sarki, and Anku Jaiswal. 2026. Mitigating Structural Noise in Low-Resource S2T: An Optimized Cascaded Nepali-English Pipeline with Punctuation Restoration. ACM Trans. Asian Low-Resour. Lang. Inf. Process. 1, 1 (March 2026), 16 pages. https://doi.org/10.1145/n.n 1 Introduction Speech-to-text translation (S2T) converts spoken speech from a source into written text in a target language. It is crucial for bridging communication barriers, enhancing digital accessibility and enabling applications like voice-command systems, automatic subtitling, virtual assistants, and cross-lingual communication. While Automatic Speech Recognition Authorsâ Contact Information: Tangsang Chongbang, 077bei047.tangsang@pcampus.edu.np; Pranesh Pyara Shrestha, 077bei030.pranesh@pcampus.edu.np; Amrit Sarki, 077bei049.amrit@pcampus.edu.np; Anku Jaiswal, anku.jaiswal@pcampus.edu.np, Department of Electronics and Computer Engineering, Institute of Engineering, Pulchowk Campus, Lalitpur, Bagmati, Nepal. Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. Copyrights for components of this work owned by others than the author(s) must be honored. Abstracting with credit is permitted. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. Request permissions from permissions@acm.org. Š 2026 Copyright held by the owner/author(s). Publication rights licensed to ACM. Manuscript submitted to ACM Manuscript submitted to ACM1 arXiv:2602.21647v2 [cs.CL] 2 Mar 2026 2Tangsang Chongbang, Pranesh Pyara Shrestha, Amrit Sarki, and Anku Jaiswal (ASR) and Machine Translation (MT) have seen significant advancements in high-resource languages, integrated S2T systems remain underdeveloped and unexplored for low-resource languages (LRLs) like Nepali. Nepali, spoken by approximately 19 million native speakers, and an additional 14 million second-language speakers [Eberhard et al.2024], presents a range of challenges for speech translation. These include limited annotated data, frequent code-switching with English [Gurung 2019], regional dialect variation, and sociolinguistic features such as complex honorifics without direct English equivalents (e.g.,tvs.tpAboth translate as âyouâ but differ in levels of formality). These characteristics complicate both ASR and MT for Nepali. End-to-end direct Speech-to-Speech translation (S2ST) models are gaining attention in recent years, for their advantages of architectural simplicity, and potential to reduce information loss, and error propagation [Bentivogli et al. 2021]. However, such approaches demand large-scale, parallel speech and text data [Sarim et al.2025], which makes it impractical for Nepali and most LRLs. In contrast, cascaded pipelines, which first transcribe speech with ASR and then translate it with MT, offer a more practical alternative in low-resource settings. Their modularity allows optimization of each component, resulting in more powerful performance [Sarim et al.2025]. Modern self-supervised ASR models like Wav2Vec2 have shown strong performance even with limited labeled data [Yi et al.2020], and multilingual NMT models like MarianMT and NLLB can be adapted even with modest amounts of LRL data to achieve good performance [Liu et al.2020; Verma et al.2022]. Consequently, for low-resource scenarios, cascaded pipelines offer greater data efficiency. However, cascaded systems introduce their own unique challenges. Error propagation is the most prominent example. A particularly underexplored issue inNepaliâ Englishliterature is the lack of punctuation in ASR outputs. While ASR models primarily focus on word accuracy, NMT models are trained on punctuated text, relying on punctuation marks for cues such as sentence boundaries, clause separation, emphasis, disambiguation. Their presence or absence can have significant impact on translation quality [Jwalapuram 2023]. In this work, we develop and analyze cascadedNepaliâ EnglishS2T system combining three main components: (1) a Wav2Vec2.0 XLS-R 300m ASR model (2) a Punctuation Restoration Module (PRM) using multilingual T5 (mT5) model for basic punctuation restoration, and (3) a multi-stage fine-tuned MarianMT NMT model. We confirm the component quality, with the ASR achieving a state-of-the-art 2.72% CER on OpenSLR-54 test set and the NMT achieving 28.32 BLEU on the FLORES-200. Critically, we demonstrate that the absence of punctuation causes 20.7% relative BLEU drop in translation quality, empirically validating the need for the PRM. The end-to-end evaluation shows that the PRM-optimized pipeline yields a 4.90 BLEU point gain over the direct ASR-to-NMT baseline. The key contributions of this work are as follows: â˘Establishing a new, optimized S2T baseline for NepaliâEnglish. We develop the first publicly documented cascaded NepaliâEnglish S2T system, demonstrating highly competitive component performance with a 2.72% CER for ASR and a 28.32 BLEU score for NMT on standard benchmarks. â˘Validating Punctuation Restoration as the Optimal Intervention. We quantitatively demonstrate that the loss of punctuation leads to a 20.7% relative BLEU drop, and show that incorporating an intermediate PRM yields robust 4.90 BLEU point gain in the end-to-end pipeline. The rest of the paper is structured as follows: Section 2 reviews related work. Section 3 outlines the methodology, including data preparation and experimental setups. Section 4 presents results and findings from the component-level evaluations, the end-to-end S2T scenarios and the human evaluation results. Section 5 discusses the findings, and Section 6 finally concludes the paper. Manuscript submitted to ACM Mitigating Structural Noise in Low-Resource S2T: An Optimized Cascaded Nepali-English Pipeline with Punctuation Restoration3 2 Related Work This section reviews prior research in ASR, NMT, and their integration into S2T pipelines. We specifically highlight the unique challenges and existing gaps in the literature concerning Nepali, for which a unified S2T system has yet to be publicly documented. 2.1 Automatic Speech Recognition (ASR) for Low-Resource Languages ASR in low-resource settings suffers from labeled data scarcity, diverse dialects and frequent code-switching. Traditional ASR approaches for Nepali relied on Hidden Markov Models (HMMs), as in the early work by [Ssarma et al.2017], which achieved 74.49% accuracy for single-word inputs and 55.55% for three-word phrase inputs Nepali recognition using Voice Activity Detection (VAD), primarily limited by their ability to model complex acoustic variations effectively. The advent of deep learning introduced more sophisticated models like RNN-CTC and CNN-RNN hybrids. [Regmi et al.2019] implemented an RNN-CTC model trained on a 1,320-word custom dataset, that achieved a 34% CER, but struggled with speaker variability and generalization. [Bhatta et al.2020] reported 1.83% CER and 11% WER using a CNN-GRU architecture on the OpenSLR43 text-to-speech (TTS) dataset. However, the studio-quality nature of OpenSLR43 limits real-world applicability. Other efforts, such as [Dhakal et al.2022], utilized bidirectional LSTM paired with ResNet and one-dimensional CNN to achieve a 17.06% CER on OpenSLR dataset. While these studies laid foundational groundwork for Nepali ASR, the models were less capable of generalizing to natural speech variations. Recent self-supervised models like Wav2Vec2 and Whisper have reshaped ASR for LRLs. By leveraging large-scale unlabeled data for pre-training, they enable better generalization with limited examples [Fatehi 2023; Hsu et al.2024; Zhu et al.2021]. Our work builds upon this paradigm by utilizing the Wav2Vec2-XLS-R-300m model on OpenSLR54 and Common Voice v17. This approach enables us to establish a competitive performance benchmark for Nepali ASR. 2.2 Machine Translation (MT) for Low-Resource Languages Similar to ASR, NMT for LRLs is constrained by lack of high-quality parallel corpora and linguistic complexities. Traditional Statistical Machine Translation (SMT) and rule-based methods dominated early efforts, but transformer- based NMT has since become the standard. However, NMT models require large datasets and are sensitive to rare words, long sentences, domain mismatch and word-alignment issues [Koehn and Knowles 2017]. To address these limitations, researchers have explored transfer learning, backtranslation, data augmentation, and pivot languages [Haddow et al.2022; Talwar and Laasri 2025; Zoph et al.2016]. For instance, MarianMT and NLLB are pretrained multilingual models that support rapid adaptation to LRLs, even with modest amounts of data [Liu et al.2020; Verma et al.2022]. Recent work by [Verma et al.2022] demonstrated the effectiveness of multi-stage fine-tuning strategy involving multilingual pre-training followed by language-pair specific fine-tuning, followed by domain fine-tuning. Inspired by these successes, our approach applies a similar multi-stage fine-tuning strategy to a MarianMT mul- en model using filtered NLLB data, synthetic parallel corpora, and a high-quality dataset to achieve substantial improvements in translation quality for Nepaliâ English. 2.3 Challenges in ASR-MT Integration and Punctuation Restoration Cascaded ASR-MT pipelines, while practical for LRLs, are inherently susceptible to error propagation, where tran- scription errors negatively affect downstream translation. Since ASR models like Wav2Vec2 primarily focus on word Manuscript submitted to ACM 4Tangsang Chongbang, Pranesh Pyara Shrestha, Amrit Sarki, and Anku Jaiswal accuracy, and are not inherently trained to predict punctuation, the resultant degradation from directly feeding the un- punctuated ASR output to NMT models is an under-explored issue, especially for LRLs. Punctuation conveys structural and semantic cues like sentence boundaries, emphasis, disambiguation, which NMT models heavily rely on for accurate translation. The work by [Jwalapuram 2023] quantified the effects of punctuation on MT specifically for German-English, Japanese-English and Ukrainian-English language pairs and concluded that models are heavily sensitive to punctuation. While prior works in high-resource languages have addressed punctuation restoration as a standalone task or via joint training, its role in ASR-MT pipelines for Nepali remains unexplored. Our work explicitly quantifies the degradation in translation performance due to the absence of punctuation, thereby underscoring the necessity for an intermediate structural noise mitigation component. Based on this gap, Section 3 details the comprehensive methodology developed to create and evaluate an optimized, punctuation-aware cascaded Nepaliâ English S2T system. 3 Methodology This section details the design, implementation, and training procedures of our cascaded Nepali S2T system. We first outline the overall system architecture. This is followed by a comprehensive description of the datasets used and the fine-tuning strategies applied to the ASR and NMT components. We then describe the implementation of the preliminary punctuation restoration module and define the specific evaluation scenarios employed. 3.1 Overall System Architecture The proposed system adopts a cascaded architecture (Figure 1) consisting of three sequential stages. â˘Stage 1 - ASR Transcription: Nepali speech is processed by the fine-tuned Wav2Vec2-XLS-R-300m model to generate an unpunctuated Nepali text transcription. â˘Stage 2 - Punctuation Restoration (Conditional): The unpunctuated ASR output is passed through a prelimi- nary mT5 model for the restoration of basic punctuation, primarily sentence-ending full stops. This stage is used only in evaluation scenarios to quantify the impact of punctuation recovery. â˘Stage 3 - NMT Translation: The transcribed text is fed into the fine-tuned MarianMT mul-en model, which produces the final English text translation. Audio/Speech (Nepali Audio) Wav2Vec2 (ASR) mT5 (Punctuation Restoration) MarianMT (Translation) Translated English text Fig. 1. Cascaded Nepali-to-English Speech-To-Text-Translation System Architecture 3.2 Datasets This section details the construction and preprocessing of the main corpora utilized in this study: two for ASR fine- tuning, three for NMT multi-stage training, one for punctuation restoration, and one for the specialized end-to-end evaluation. 3.2.1 ASR Datasets. The ASR model was fine-tuned on a combination of two publicly available Nepali speech corpora to increase speaker diversity and utterance coverage (see Table 1). Manuscript submitted to ACM Mitigating Structural Noise in Low-Resource S2T: An Optimized Cascaded Nepali-English Pipeline with Punctuation Restoration5 Table 1. Nepali ASR datasets used for fine-tuning. DatasetTotal Utterances/ClipsHours (Validated) Used Utterances Common Voice v17 a 742 (validated) + 595 (internally reviewed âotherâ split) = 1,337 1.01,337 OpenSLR54 b âź157,000 (initial)â148,188 (post-numerals-filtering)âź143 (pre-5s filter) 136,095 (post-5s filter) a https://huggingface.co/datasets/mozilla-foundation/common_voice_17_0 b https://huggingface.co/datasets/iamTangsang/OpenSLR54-Nepali-ASR From the Common Voice v17 (ne-NP) dataset, we used 1,337 utterances (approximately 2 hours) from 32 speakers. We combined the âvalidatedâ split with internally reviewed and validated samples from the âotherâ split for initial fine-tuning. All audio files were uniformly resampled to 16 kHz. For OpenSLR54, utterances containing numerals were removed to ensure the model outputs numbers in words rather than digits, since inconsistent transcriptions were common. For example, the numeral007appeared as "Ejro Ejro sAt " in one recording, "s y s y sAt" in another, and simply "sAt" elsewhere. To avoid confusing the model, all such utterances were excluded. In addition, since sequences longer than 5 seconds caused excessive processing time and frequent out-of-memory errors on the Colab environment, these were also removed. After filtering, 136,095 utterances remained for final fine-tuning. 3.2.2 NMT Datasets and Multi-Stage Strategy. The Nepali-to-English NMT model was fine-tuned using a three-stage approach. This strategy leverages broad language knowledge from noisy sources first, expands coverage with high- volume synthetic data, and concludes with a refinement stage using high-fidelity data (see Table 2). 3.2.3 Punctuation and Segmentation Dataset. The punctuation Restoration module was trained to simultaneously address two critical issues observed in the ASR output: missing punctuation and fused words (lack of segmentation). The training corpus was derived from the 210,875 high-quality sentence pairs from Stage 3 of the NMT training. â˘Input Sequence: The original Nepali sentences were systematically modified under two distinct conditions: (a) by removing both inter-word spaces and punctuation marks to produce completely fused character sequences (e.g., dAUdl(yssEnklAIBnEtmFlksrFjAyOfAUlrjonATnmrkok rA, and (b) by removing only punctuation marks while preserving boundaries. These modifications were designed to emulate the types of errors commonly encountered in our ASR outputs. â˘Target Sequence: The corresponding reference Nepali text, fully segmented and punctuated, representing the desired restoration output (e.g.,dAUdl (ys sEnklAI Bn ,â EtmFl ksrF jAyO fAUl r jonATn mrko k rA ?â ) The final dataset comprised a combination of both modification types, enabling the model to learn from varying degrees of degradation. By framing this as a unified text-to-text generation task, the mT5 model was trained to jointly recover accurate word boundaries and restore essential punctuation marks (commas, and full stops) in a single step. Manuscript submitted to ACM 6Tangsang Chongbang, Pranesh Pyara Shrestha, Amrit Sarki, and Anku Jaiswal Table 2. Corpora used at each stage of NMT training and their filtering/rationale. StageCorpus TypeSize (Pairs) Primary Source / FilteringRationale Stage1: Founda- tional FilteredWeb crawled NLLB a 708,000Initial 19.6M web-crawled NLLB cor- pus pairs filtered by LabSE cosine sim- ilarity (>0.80), and removal of Nepali numerals (to maintain consistency). Knowledge Transfer: To learn funda- mental translation nuances from fil- tered, noisy web data. Stage 2: Ex- pansion Synthetic Corpora b,c 5.02M Combinationof1.6Msynthetic pairs [Duwal and Bal 2019] and 3.42M pre-training pairs [Duwal et al.2024] filtered using a chrF++ cut-off of 50 (translated with 8-bit NLLB/IndicTrans2). Coverage Expansion: Increase expo- sure to diverse structures and vocab- ularies using high-volume, quality- filtered synthetic data. Stage 3: Re- finement Manually translated high-Quality corpora d,e 210,875OPUS (GNOME/KDE/Ubuntu), Bible, Global Voices, Penn Treebank + [Acharya and Bal 2018], Nepal Law Commission,EasyBible,Nepal Budget Speech 2081/82, Shirish ko Fool Translation. Quality Refinement: Final tuning for translation fluency and accuracy on high-fidelity human-translated pairs. a https://huggingface.co/datasets/iamTangsang/Nepali-to-English-Translation-Dataset b https://huggingface.co/datasets/sharad461/ne-en-synthetic-1.6m c Dataset described in [Duwal et al. 2024]; currently unavailable online. d https://huggingface.co/datasets/sharad461/ne-en-parallel-177k e https://github.com/BISHALTWR/Nepali-English-Translation-Dataset 3.2.4 Final End-to-End Evaluation Dataset. The final performance of the full cascaded system was assessed on a newly created, representative test set 1 to simulate real-world usage. The final performance of the full cascaded system was assessed on a newly created, representative test set to simulate real-world usage. This evaluation dataset consists of 900 total audio clips, comprising 300 unique Nepali sentences recorded by 3 different speakers. Sentence Selection: The 300 sentences were manually crafted to ensure linguistic diversity, including: ⢠Statements (150): Simple declarative sentences. ⢠Questions (60): To test the modelâs ability to recognize interrogative punctuation. ⢠Commands/Imperatives (30): To test varied sentence structures. ⢠Complex/Compound Sentences (30): To test handling of multiple clauses. ⢠Named Enities (30): To test proper noun transcription and translation. Annotation: All audio was manually transcribed, punctuated, and translated to establish the Gold-Standard reference used for final S2T evaluation. 3.3 Model Training and Fine-tuning All models were trained and fine-tuned using the Google Colab environment on a Tesla T4 GPU. The training strategies were modularly defined for the ASR, NMT, and Punctuation Restoration components to ensure optimal performance in the low-resource setting. 1 The final evaluation dataset can be accessed through Hugging Face at https://huggingface.co/datasets/iamTangsang/nepali_to_english_pipeline_ evaluation Manuscript submitted to ACM Mitigating Structural Noise in Low-Resource S2T: An Optimized Cascaded Nepali-English Pipeline with Punctuation Restoration7 3.3.1 ASR Model Fine-tuning. The Wav2Vec2.0 XLS-R model (300 million parameters variant) was selected as the foundational architecture. The fine-tuning process was executed in a three-stage sequential manner, incorporating iterative vocabulary and learning rate optimization. Initial Tuning on Common Voice v17: The model was initially fine-tuned on a combined set of 1,337 items from the âvalidatedâ and internally reviewed âotherâ splits of the Common Voice v17 ne-NP corpus. ⢠Training was performed for 30 epochs with a learning rate of 3e-4. ⢠Vocabulary: The initial vocabulary size was of 64 tokens. Fine-tuning on OpenSLR 54 (Stage 1 - Plateau): The model obtained from CV-17 was further fine-tuned on the OpenSLR54 dataset. â˘Initial Parameters: The learning rate was set to 3e-4 with a linear decay for 16 epochs. A large, exploratory vocabulary of 71 tokens (including special tokens <s>, </s>, __UNK__, __PAD__) was utilized. ⢠Observation: Training plateaued around the 16 th epoch (WER fluctuating 26%-29%). Fine-tuning on the OpenSLR54 (Stage 2 - Refinement and Optimization): To address the plateauing, the model was retrained for an additional 3 epochs using key optimizations: ⢠Learning Rate: Reduced significantly to 2e-5 with linear decay. ⢠Vocabulary Optimization: The final vocabulary size was reduced to 67 total tokens by removing characters not used in standard Nepali, which improved model stability and generalization. Model Regularization: The Wav2Vec2 XLS-R utilized the following regularization parameters: attention_dropout 0.1, hidden_dropout 0.1, layerdrop: 0.1. The CTC loss reduction was set to "mean". The key hyperparameters used were as shown in Table 3. Table 3. Key Hyperparameters for Wav2Vec2.0 Fine-Tuning HyperparameterValue Learning Rate (Final) 2Ă 10 â5 Batch Size (Per Device)16 Gradient Accumulation2 Warmup Steps500 3.3.2 Punctuation Restoration Module Fine-tuning. The preliminary Punctuation Restoration module employed the Google mT5 model fine-tuned on the dataset defined in Section 3.2.3. The training was framed as a single text-to-text task to simultaneously restore segmentation and punctuation. The key hyperparameters used were as shown in Table 4. Table 4. Key Hyperparameters for mT5 Fine-Tuning HyperparameterValue Learning Rate 2Ă 10 â5 Batch Size (Train)8 Weight Decay0.1 Manuscript submitted to ACM 8Tangsang Chongbang, Pranesh Pyara Shrestha, Amrit Sarki, and Anku Jaiswal 3.3.3 NMT Model Multi-Stage Fine-tuning. The MarianMT mul-en model underwent a sequential, multi-stage fine- tuning process to effectively utilize the increasingly high-quality data corpora. Stage 1 & 2: Pre-training (NLLB Filtered & Synthetic Datasets): The model was initially trained on the 708,000 filtered NLLB corpus (Stage 1) for a single epoch to establish foundational transfer knowledge. Training then continued on the combined 5.02 million synthetic corpora (Stage 2) to expand linguistic coverage. ⢠Learning Rate: 5e-5 was used across the pre-training stages for a balance between speed and stability. ⢠Token Limits: The input and output were capped at 256 tokens. Stage 3: Fine-tuning (High-Quality Dataset): The model was finally refined on the 210,875 high-quality parallel pairs for 12 epochs. Table 5 summarizes the hyperparameters used. Table 5. Hyperparameters for MarianMT Pre-training and Fine-tuning HyperparameterPre-trainingFine-tuning Learning Rate 5Ă 10 â5 2Ă 10 â5 Batch Size (Train/Eval)1616 Weight Decay0.010.01 Warmup Steps5% of total steps5% of total steps 3.4 Evaluation Metrics Performance was measured using task-specific metrics as in Table 6. Table 6. Evaluation Types and Corresponding Metrics Evaluation TypeMetrics ASRWER, CER NMT BenchmarkingBLEU, chrF++, METEOR Punctuation Impact & S2TÎBLEU,ÎchrF++ Human Evaluation1â5 Likert Scale 3.5 Experimental Setup and Evaluation Scenarios The system performance was rigorously evaluated across three core experimental setups: 1. Component Performance: Individual evaluation of ASR (on OpenSLR-54 test set split) and NMT (on FLORES- 200/Tatoeba and the high-quality test split) to establish component-level baselines. 2. Punctuation Impact Quantification: ⢠Procedure: The NMT model was evaluated on the FLORES-200 dev, devtest, and Tatoeba datasets. This was done by comparing the performance on (1) the original, fully punctuated source sentences against (2) a modified version of the same sentences with all punctuation removed (Simulated ASR output). â˘Goal: The resulting perforamnce delta (ÎBLEU andÎchrF++) quantifies the raw performance degradation directly attributable to the loss of source-side punctuation. Manuscript submitted to ACM Mitigating Structural Noise in Low-Resource S2T: An Optimized Cascaded Nepali-English Pipeline with Punctuation Restoration9 3.End-to-End (S2T) Evaluation: The full cascaded pipeline was tested on the 900-clip custom dataset across three scenarios: ⢠Scenario A (Direct Baseline): ASR(unpunctuated output)â NMT. ⢠Scenario B: ASRâ Punctuation Restoration Module (with space removal from ASRâs output)â NMT. ⢠Scenario C: ASRâ Punctuation Restoration Module (without space removal from ASRâs output)â NMT. Human Evaluation (S2T):. A dedicated human evaluation was conducted across all three S2T scenarios using the 900-clip custom test set. Three of the authors conducted a blind human evaluation to assess translation fluency and adequacy on a 1 â 5 Likert scale. The outputs from all system configurations were randomly mixed and anonymized prior to scoring, ensuring that evaluators were unaware of which system configuration produced each translation. After all the ratings were completed, the configuration identities were revealed for analysis. Scores were averaged across raters for each system. To validate the consistency of the human assessments, inter-rater reliability was calculated using Krippendorffâs Alpha(íź ) for both metric, treating the Likert scores as ordinal data. 4 Results and Findings This section presents the empirical results of the cascaded S2T system, structured arond the performance of its components, the quantified impact of noise, and the final end-to-end efficacy of the proposed pipeline. 4.1 Component-Level Performance 4.1.1 ASR Performance. The fine-tuned Wav2Vec2 XLS-R model was evaluated on the OpenSLR-54 test set (see Table 7). The resulting WER of 16.82% and a particularly low CER of 2.72% confirm the viability of the self-supervised transformer approach for low-resource Nepali ASR. This CER represents a substantial improvement over previous established supervised models on the same dataset (e.g., [Dhakal et al.2022] at 17.07% CER and [Banjara et al.2020] at 27.72% CER). Table 7. ASR Performance on OpenSLR-54 Test Set MetricResult Word Error Rate (WER)16.82% Character Error Rate (CER)2.72% 4.1.2 NMT Benchmarking. The final NMT model was benchmarked against widely used standard sets and the base model. The results are reported in Table 8. The NMT model achieved significant gains across all benchmarks, surpassing the previously reported state-of-the-art results [Duwal and Bal 2019] on FLORES-200 byâź8â9 BLEU points and drastically outperforming the base multilingual MarianMT model. This performance ensures that the NMT component is highly proficient when provided with high-quality, fully punctuated source text. 4.2 Quantifying Punctuation Impact on NMT To isolate the impact of noise common to ASR output, the NMT model was tested in two modes: with and without Nepali punctuation. Quantitative results are reported in Table 9 and visualized in Figure 2. The results confirm punctuation removal leads to notable degradation in translation performance across all datasets. Notably, FLORES-200 DevTest exhibits a 5.91 BLEU drop (20.3% relative), while Tatoeba suffers the most extreme degradation, exhibiting 11.26 Manuscript submitted to ACM 10Tangsang Chongbang, Pranesh Pyara Shrestha, Amrit Sarki, and Anku Jaiswal Table 8. Benchmarking of the Final MarianMT Model Against Standard Test Sets and Previous Best Results Test SetMetric Our Score Previous Best FLORES-200 DevTest BLEU29.0420.76 chrF++58.14N/A METEOR0.6314N/A FLORES-200 DevBLEU28.4819.37 chrF++58.07N/A METEOR0.5676N/A TatoebaBLEU39.663.5 a chrF++55.730.168 a METEOR0.5532N/A a Base MarianMT mul-en: https://huggingface.co/Helsinki-NLP/opus-mt-mul-en. BLEU drop (28.39% relative). This empirical evidence provides the primary justification for integrating a Punctuation Restoration module to mitigate error propagation. Table 9. Impact of Punctuation on Translation Quality Across Test Sets (MarianMT Model Evaluation) Test SetConditionBLEU chrF++ÎBLEUÎchrF++ FLORES-200 DevTest Punctuated29.0458.14â Unpunctuated (ASR)23.1355.25 -5.91-2.89 FLORES-200 DevPunctuated28.4858.07â Unpunctuated (ASR)24.1256.11 -4.36-1.97 TatoebaPunctuated39.6655.73â Unpunctuated (ASR)28.4051.16 -11.26-4.57 BLEUchrF++ 0 10 20 30 40 50 60 Score 29.04 58.14 23.13 55.25 FLORES-200 DevTest BLEUchrF++ 28.48 58.07 24.12 56.11 FLORES-200 Dev BLEUchrF++ 39.66 55.73 28.40 51.16 Tatoeba PunctuatedUnpunctuated Fig. 2. Comparative performance of NMT on punctuated and unpunctuated (Simulated ASR) inputs across three test sets. A consistent degradation is observed, confirming that punctuation loss substantially impacts translation quality. Manuscript submitted to ACM Mitigating Structural Noise in Low-Resource S2T: An Optimized Cascaded Nepali-English Pipeline with Punctuation Restoration11 4.3 End-to-End(S2T) Evaluation The full cascaded system was evaluated on the 900-clip custom test set. 4.3.1 Automatic Metrics (BLEU and chrF++). The raw ASR output on this custom set had an average WER of 37.36% and CER of 13.76%, reflecting the difficulty in accuracy of the ASR model in real-world usage despite good performance in the OpenSLR54 test set. The three scenarios tested the efficacy of the Punctuation Restoration module. Table 10 and Figure 3 summarizes the result. Scenario AScenario BScenario C 0 10 20 30 40 50 Score 31.48 51.84 32.77 51.05 36.38 54.56 BLEUchrF++ Fig. 3. End-to-End S2T Performance Comparison (BLEU & chrF++) Table 10. Impact of System Configuration on Translation Quality ScenarioBLEU chrF++ÎBLEU (vs. Baseline A) A (Baseline)31.4851.84â B (Proposed - Full Pipeline)32.7751.05+1.29 C (Optimal Pipeline)36.3854.56+4.90 Scenario A: ASRâ NMT Scenario B: ASRâ Punctuation Restoration with space removalâ NMT Scenario C: ASRâ Punctuation Restorationwithout space removalâ NMT. Scenario C (Punctuation Restoration on the raw ASR output (without removing spaces)) yielded the most significant performance gain, achieving a 4.90 BLEU point increase and a 2.72 chrF++ point increase over the direct baseline (Scenario A). The minor drop in performance of performance B suggests that the initial space removal and re-tokenization introduced more detrimental segmentation errors than it resolved. Or, it may be because of lack of enough data as we used onlyâź208íexamples to train our mT5 model. Scenario C, which effectively only restored punctuation on raw ASR output, is the optimal system configuration. Manuscript submitted to ACM 12Tangsang Chongbang, Pranesh Pyara Shrestha, Amrit Sarki, and Anku Jaiswal 4.3.2 Human Evaluation (Adequacy and Fluency). The human evaluation results, summarized in Table 11 and illustrated in Figure 4, confirm the trends observed in the automatic evaluation metrics. Table 11. Impact of System Configuration on Translation Fluency and Adequacy with Inter-Rater Reliability (íź ) ScenarioAvg. Fluency íź (Fluency) Avg. Adequacy íź (Adequacy) A (Baseline)3.6770.7233.5810.783 B (Proposed)3.7740.7463.6280.780 C (Optimal)3.8040.7483.6730.785 Scenario A: ASRâ NMT. Scenario B: ASRâ Punctuation Restoration (with space removal)â NMT. Scenario C: ASRâ Punctuation Restoration (without space removal)â NMT. Scenario AScenario BScenario C 0 1 2 3 4 5 Average Score (1 5) 3.677 3.581 3.774 3.628 3.804 3.673 FluencyAdequacy Fig. 4. Human evaluation of S2T output. Average fluency and adequacy scores (1â5 scale) for Scenarios A, B, and C. Scenario C consistently yields the highest-quality output, achieving the best scores for both Fluency (3.804) and Adequacy (3.673). This result validates that the BLEU/chrF++ gains translate directly into a perceptibly higher-quality, more readable, and more meaning-preserving translations. Inter-rater reliability was measured using Krippendorffâs Alpha(íź) for both metrics, yielding high agreement (íź âĽ0.723) across all systems, reinforcing the reliability of these findings. 4.3.3 Performance Analysis by Sentence Type. The breakdown by sentence category further highlights the impact of the punctuation restoration on specific linguistic structures. Table 12 summarizes the result. Manuscript submitted to ACM Mitigating Structural Noise in Low-Resource S2T: An Optimized Cascaded Nepali-English Pipeline with Punctuation Restoration13 Table 12. Sentence-Type-wise Human Evaluation Results Sentence Type Metric Scenario A (Baseline) Scenario B Scenario C (Optimal) StatementsAdequacy3.9433.8393.962 StatementsFluency3.9673.9214.048 CommandsAdequacy3.1673.4213.416 CommandsFluency2.9813.1713.226 QuestionsAdequacy3.3653.3683.538 QuestionsFluency3.5993.6493.781 Named Entities Adequacy3.2693.8873.269 Named Entities Fluency3.2593.9213.323 ComplexAdequacy3.1813.2573.438 ComplexFluency3.4163.4793.611 Evaluation by sentence category based on output type. Scores are averaged over three human evaluators using a 1â5 Likert scale. Higher scores indicate better perceived fluency or adequacy. While Scenario C generally performs the best, Scenario B provided a superior outcome for the translation of Named Entities (Adequacy 3.887 vs. 3.269 and Fluency 3.921 vs. 3.323), suggesting that the unique pre-processing step may benefit specialized recognition tasks. Conversely, Scenario C showed consistent, measurable improvement for complex and command sentences, confirming its broad effectiveness in restoring sentence structure necessary for NMT processing. 5 Discussion The experimental results reveal key factors influencing the performance of cascaded S2T for the low-resource Nepali â English translation pair. We analyze the effectiveness of component-level refinements, quantify error propagation, evaluate the Punctuation Restoration Module (PRM), and validate findings through human assessment. 5.1 Validation of Component-Level Refinements ASR Performance: The fine-tuned Wav2Vec2 XLS-R model achieved a low CER (2.72%), representing a significant improvement over previous state-of-the-art supervised models on the OpenSLR-54 dataset. This success is primarily attributed to transfer learning. The Wav2Vec2 model, pre-trained on vast amounts of unlabelled speech data, quickly adapted its generalized features to limited Nepali data. This can also be attributed to the fact that, Nepali is very similar to Hindi, and Wav2Vec2 has been pre-trained on large amounts of unlabeled Hindi data as well, effectively mitigitaing the effects of data scarcity. NMT performance: The subsequent NMT model demonstrated high translation proficiency, outperforming previously documented models on standard benchmarks (e.g.,âź8â9 BLEU points on FLORES-200). This confirms the success of the data curation and MarianMT fine-tuning stages, establishing the NMT component as highly capable when provided with clean, structured input, or large-scale high-quality training data. Manuscript submitted to ACM 14Tangsang Chongbang, Pranesh Pyara Shrestha, Amrit Sarki, and Anku Jaiswal 5.2 Quantifying Error Propagation The experiment isolating punctuation loss empirically confirms the necessity of the proposed PRM intervention. By comparing NMT performance on punctuated vs. unpunctuated gold-standard text, the results showed a severe performance degradation, peaking at a 11 BLEU point loss (28.39% relative drop) on the Tatoeba test set. This degradation occurs because punctuation marks are essential structural tokens that define sentence bondaries and syntactic relationships, which NMT models rely on for accurate translation. Their removal forces the NMT encoder to process long, coherent streams of tokens, leading to grammatical incoherence, and a copmplete loss of contextual cues (e.g., distinguishing between a statement and a question). This finding empirically validates the core premise of this resrearch: that directly addressing the loss of structural metadata (punctuation) is the most critical intervention point for improving cascaded S2T quality. 5.3 Analysis of End-to-End Pipeline Performance The S2T evaluation on the custom test set (with a ASR WER of 37.36%) reveals the optimal strategy for integrating the PRM: Optimal Configuration (Scenario C):. Scenario C (ASRâPunctuation Restoration (without removing spaces i.e. without addressing ASR segmentation issues)âNMT) achieved the highest S2T scores (36.38 BLEU), providing a 4.90 BLEU point gain over the direct baseline (Scenario A). This success indicates that the ASR componentâs internal language model, despite its high WER, was effective at generating adequate word segmentation (i.e., correct spacing between words). The PRM in Scenario C succeeded because it made the minimal effective intervention: focusing primarily on restoring the lost punctuation and sentence boundaries without disturbing the existing, reasonably correct word spacing. Sub-Optimal Configuration (Scenario B):. Scenario B (ASRâPunctuation Restoration with removing spaces from inputâNMT) performed sub-optimally, achieving lower score than Scenario C. This outcome demonstrates a crucial trade-off: the deliberate pre-processinng step of removing all spaces (token segmentation) and forcing the PRM to re-segment the entire utterance introduced new, cascading segmentation errors. These self-inflicted errors proved more detrimental to the downstream NMT model than the benefit gained from the restored punctuation alone. This finding suggests that for Nepali, preserving the ASRâs inherent word segmentation is critical and should not be discarded in favor of concurrent text-to-text re-segmentation unless the PRM is highly trained on a large-scale high-quality and diverse dataset for that specific task. 5.4 Validation by Human Assessment The results from the human evaluation strongly corroborate the automatic metrics. Scenario C consistently scored the highest in both Fluency (3.804) and Adequacy (3.673). This correlation validates that the 4.90 BLEU gain is not merely a statistical artifact but translates directly into a perceivably higher-quality, more readable, and meaning- preserving translation for the end-user. This is further supported by a high degree of consensus among evaluators, with a Krippendorffâs Alpha (íź ) of 0.748 for fluency and 0.785 for adequacy across the optimal configuration. The human breakdown also highlighted specific structural benefits. While Scenario B showed a surprising advantage in case of Named Entities (suggesting value in its forced segmentation for specific tasks), Scenario Câs superior Manuscript submitted to ACM Mitigating Structural Noise in Low-Resource S2T: An Optimized Cascaded Nepali-English Pipeline with Punctuation Restoration15 performance in Complex sentences and Commands confirms its broader robustness in restoring the sentence structure necessary for NMT. 5.5 Primary Contribution and Future Work 5.5.1 Primary Contribution. The primary contribution of this work is the empirical validation and successful deployment of a Punctuation Restoration Module as a targeted intermediary step in a low-resource S2T pipeline. We demonstrated that for Nepali, this strategy provides a robust and significant performance gain (+4.90 BLEU) by mitigating the most severe form of ASR noise â the loss of structural cues â thereby establishing a new, optimized baseline for Nepali-to-English S2T. Limitations and Future Work The main limitations stem from the cascaded nature of the system and high-quality data scarcity. ⢠Cascaded Error Floor: The systemâs performance ceiling is ultimately limited by the 37.36% WER of the ASR output on the custom test set. Future work should focus on integrating the PRM and NMT into a single, end-to-end Speech-to-Text Translation model to allow for joint training and attention, potentially overcoming cascaded error propagation. â˘PRM Robustness: The failure of Scenario B highlights the need for a more robust PRM capable of handling concurrent punctuaion and word segmentation. Future work should explore training the PRM on deliberately noisy, unsegmented data to improve its generalization across varying ASR output styles. â˘Data Generalization: The training data for all components remains relatively narrow. Future efforts must prioritize the creation of larger, more diverse, and domain-spanning Nepali speech and parallel text corpora. 6 Conclusion This research successfully addressed the critical challenge of structural noise propagation in the Nepali-to-English cascaded S2T pipeline, a pervasive issue for low-resource languages that lack sophisticated end-to-end models. We first established the necessity of the intervention by empirically demonstrating that the removal of punctuation alone resulted in a substantial performance degradation, causing aâź6 BLEU point loss on standard NMT evaluation sets. To mitigate this core problem, we proposed and validated a Punctuation Restoration Module (PRM) as a crucial intermediate component. The optimized pipeline configuration (Scenario C), which deployed the PRM on the raw ASR output to restore structural cues, yielded a significant performance increase, achieving a 4.90 BLEU point gain over the direct ASR-to-NMT baseline. This result was further validated by human evaluation, with the optimized system scoring highest in both Adequacy (3.673) and Fluency (3.804), reinforced by inter-rater reliability (íź âĽ0.72) for all scenarios. The primary contribution of this work is the empirical proof that a targeted, text-to-text intervention, specifically, punctuation restoration, is an extremely effective strategy for overcoming noise inherent to low-resource ASR and significantly improving the overall quality of cascaded S2T. This work establishes an optimized baseline for Nepali-to- English S2T and provides a validated architectural blueprint for future research in similar low-resource language pairs. Manuscript submitted to ACM 16Tangsang Chongbang, Pranesh Pyara Shrestha, Amrit Sarki, and Anku Jaiswal Future work should focus on integrating the PRM and NMT into a single end-to-end Speech-to-Text Translation model to overcome the remaining limitations of cascaded error propagation and explore training a more robust PRM to concurrently handle both punctuation and word segmentation errors. References Praveen Acharya and Bal Krishna Bal. 2018. A Comparative Study of SMT and NMT: Case Study of English-Nepali Language Pair.. In SLTU. 90â93. Janardan Banjara, Kaushal Raj Mishra, Jayshree Rathi, Karuna Karki, and Subarna Shakya. 2020. Nepali speech recognition using cnn and sequence models. In 2020 IEEE International Conference on Machine Learning and Applied Network Technologies (ICMLANT). IEEE, 1â5. Luisa Bentivogli, Mauro Cettolo, Marco Gaido, Alina Karakanta, Alberto Martinelli, Matteo Negri, and Marco Turchi. 2021. Cascade versus direct speech translation: Do the differences still make a difference? arXiv preprint arXiv:2106.01045 (2021). Bharat Bhatta, Basanta Joshi, and Ram Krishna Maharjhan. 2020. Nepali speech recognition using CNN, GRU and CTC. In Proceedings of the 32nd conference on computational linguistics and speech processing (ROCLING 2020). 238â246. Manish Dhakal, Arman Chhetri, Aman Kumar Gupta, Prabin Lamichhane, Suraj Pandey, and Subarna Shakya. 2022. Automatic speech recognition for the Nepali language using CNN, bidirectional LSTM and ResNet. In 2022 International Conference on Inventive Computation Technologies (ICICT). IEEE, 515â521. Sharad Duwal and Bal Krishna Bal. 2019. Efforts in the development of an aug-mented englishânepali parallel corpus. In Proceedings of the 1st International Conference on Language Technologies for All. European Language Resources Association Paris, France, 375â378. Sharad Duwal, Suraj Prasai, and Suresh Manandhar. 2024. Domain-adaptative Continual Learning for Low-resource Tasks: Evaluation on Nepali. arXiv preprint arXiv:2412.13860 (2024). David M. Eberhard, Gary F. Simons, and Charles D. Fennig. 2024. Ethnologue: Languages of the World (27 ed.). SIL International. https://w.ethnologue. com Online version, accessed 2025-06-02. Kavan Fatehi. 2023. Self-supervised learning for automatic speech recognition In low-resource environments. Ph. D. Dissertation. University of Nottingham. Dinesh Gurung. 2019. Nepali-English code-switching in the conversations of Nepalese people: a sociolinguistic study. Ph. D. Dissertation. University of Roehampton. Barry Haddow, Rachel Bawden, Antonio Valerio Miceli Barone, JindĹich Helcl, and Alexandra Birch. 2022. Survey of low-resource machine translation. Computational Linguistics 48, 3 (2022), 673â732. Ming-Hao Hsu, Kuan Po Huang, and Hung-yi Lee. 2024. Meta-Whisper: Speech-Based Meta-ICL for ASR on Low-Resource Languages. arXiv preprint arXiv:2409.10429 (2024). Prathyusha Jwalapuram. 2023. Pulling Out All The Full Stops: Punctuation Sensitivity in Neural Machine Translation and Evaluation. In Findings of the Association for Computational Linguistics: ACL 2023. 6116â6130. Philipp Koehn and Rebecca Knowles. 2017. Six challenges for neural machine translation. arXiv preprint arXiv:1706.03872 (2017). Yinhan Liu, Jiatao Gu, Naman Goyal, Xian Li, Sergey Edunov, Marjan Ghazvininejad, Mike Lewis, and Luke Zettlemoyer. 2020. Multilingual denoising pre-training for neural machine translation. Transactions of the Association for Computational Linguistics 8 (2020), 726â742. Paribesh Regmi, Arjun Dahal, and Basanta Joshi. 2019. Nepali speech recognition using rnn-ctc model. International Journal of Computer Applications 178, 31 (2019), 1â6. Mohammad Sarim, Saim Shakeel, Laeeba Javed, Mohammad Nadeem, et al.2025. Direct Speech to Speech Translation: A Review. arXiv e-prints (2025), arXivâ2503. Manish K Ssarma, Avaas Gajurel, Anup Pokhrel, and Basanta Joshi. 2017. HMM based isolated word Nepali speech recognition. In 2017 International Conference on Machine Learning and Cybernetics (ICMLC), Vol. 1. IEEE, 71â76. Abhimanyu Talwar and Julien Laasri. 2025. Pivot Language for Low-Resource Machine Translation. arXiv preprint arXiv:2505.14553 (2025). Neha Verma, Kenton Murray, and Kevin Duh. 2022. Strategies for adapting multilingual pre-training for domain-specific machine translation. In Proceedings of the 15th biennial conference of the Association for Machine Translation in the Americas (Volume 1: Research Track). 31â44. Cheng Yi, Jianzhong Wang, Ning Cheng, Shiyu Zhou, and Bo Xu. 2020. Applying wav2vec2. 0 to speech recognition in various low-resource languages. arXiv preprint arXiv:2012.12121 (2020). Han Zhu, Li Wang, Jindong Wang, Gaofeng Cheng, Pengyuan Zhang, and Yonghong Yan. 2021. Wav2vec-S: Semi-supervised pre-training for low-resource ASR. arXiv preprint arXiv:2110.04484 (2021). Barret Zoph, Deniz Yuret, Jonathan May, and Kevin Knight. 2016. Transfer learning for low-resource neural machine translation. arXiv preprint arXiv:1604.02201 (2016). Manuscript submitted to ACM