--- title: "Advancing Arabic Speech Recognition — Abusheikh; 2025" description: "Arabic speech recognition has little transcribed audio, especially for dialects, and hand transcription is slow and expensive." authors: ["Mahmoud Salhab; Marwan Elghitany; Shameed Sait; Syed Sibghat Ullah; Mohammad Abusheikh;", "Hasan Abusheikh"] year: 2025 venue: "arXiv:2504.12254" source: "https://arxiv.org/abs/2504.12254" doi: "10.48550/arxiv.2504.12254" arxiv: "2504.12254" licence: "CC BY 4.0 (arXiv)" licence_url: "http://creativecommons.org/licenses/by/4.0/" summary_licence: "Elamur Cosmic License" canonical: "https://research.elamur.ai/free/en/papers/salhab-2025-advancing-arabic-speech-recognition-through/" --- [Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/salhab-2025-advancing-arabic-speech-recognition-through/) [home](https://md.research.elamur.ai/free/en/) / [papers](https://md.research.elamur.ai/free/en/papers/) / [2025](https://md.research.elamur.ai/free/en/papers/2025/) / Advancing Arabic Speech Recognition Through Large-Scale Weakly Supervised Learning # Advancing Arabic Speech Recognition Through Large-Scale Weakly Supervised Learning **Authors:** Mahmoud Salhab; Marwan Elghitany; Shameed Sait; Syed Sibghat Ullah; Mohammad Abusheikh;, Hasan Abusheikh · **Year:** 2025 · **Venue:** arXiv:2504.12254 · **Licence:** [CC BY 4.0 (arXiv)](http://creativecommons.org/licenses/by/4.0/) **Links:** [Publisher's page](https://arxiv.org/abs/2504.12254) · [PDF](https://arxiv.org/pdf/2504.12254) · [DOI 10.48550/arxiv.2504.12254](https://doi.org/10.48550/arxiv.2504.12254) · [arXiv 2504.12254](https://arxiv.org/abs/2504.12254) · [OpenAlex](https://openalex.org/W4414464546) **Citations:** 4 citations (openalex-snapshot · 2026-10-14) · [4](#cited-by) in our library, of the 140884 papers whose bibliographies we read ## Summary Arabic speech recognition has little transcribed audio, especially for dialects, and hand transcription is slow and expensive. The authors avoid manual labels altogether. They train a Conformer model from scratch on about 15,000 hours of automatically labeled Arabic speech covering both MSA and dialects. This data was filtered from roughly 30,000 hours of unlabeled in-house recordings. Labels come from a multi-stage pipeline: - VAD splits the audio into short segments. - Several existing ASR systems each produce a transcript. - The pipeline keeps only segments where these transcripts largely agree and picks the most consensual one. - A language-model perplexity check rejects unlikely text. - Surviving pieces are merged into longer training chunks. The process ran for two rounds, with each newly trained model joining the labeling ensemble. With greedy decoding and no external LM, the model scores best on average across six public Arabic benchmarks, beating open-source and commercial systems. This suggests noisy, agreement-filtered labels can replace costly human transcription for low-resource languages. ## Key points - **Data and labeling ensemble:** - About 30K hours of unlabeled in-house audio were reduced to about 15K hours of weakly labeled training data. - Clean public sets served only for development and testing. - The first round's transcribers were a proprietary ASR, FastConformer and Whisper. - Whisper was removed after round one because it hallucinated text. - **Filtering rules:** - Segments were capped at 5 s, and those with overlapping speakers were dropped. - A segment was also dropped if the average pairwise WER/CER between transcripts exceeded a threshold. - The chosen transcript minimizes the summed Levenshtein distance to the other hypotheses. - Merged training chunks are capped at 15 s. - **Benchmarks:** evaluation used the Open Universal Arabic ASR Leaderboard code on SADA, Common Voice 18.0, MASC (clean and noisy), MGB-2 and Casablanca. - **Headline numbers:** - The model's average WER is 26.68 and CER is 10.05. - The strongest baseline, Nvidia Conformer CTC Large V3, scores 34.74 and 13.37. - That is a 23.19% relative WER reduction and 24.78% CER reduction. - The largest gain is on SADA (WER 27.71 vs 47.26). - **Commercial systems:** ElevenLabs Scribe, Microsoft Azure STT and GPT-4o transcribe are all clearly worse (best average WER 40.05). - **Weak spots:** CER is slightly higher than the best baseline on Common Voice (3.21 vs 3.05) and MASC clean (5.80 vs 5.63). ## Interesting ideas - Choosing the label as the "medoid" hypothesis, the one closest by edit distance to all other systems' outputs, gives a simple consensus label without any voting model. — For Mova (learn): Silver UD trees from several witness parsers could be chosen as the tree with the smallest total attachment/label disagreement to the others, and items with high average pairwise disagreement could be dropped instead of adjudicated. - Labeling in very short units (≤5 s) means a rejected unit wastes little data, and accepted pieces are glued back into longer examples afterward. — For Mova (tree): Silver treebank filtering could accept or reject at clause or sentence level rather than per document, then reassemble accepted parts so the parser still sees long inputs. - The labeler pool evolves: each round's trained model becomes a new witness, and a witness found to hallucinate is removed entirely. — For Mova (learn): The witness-parser set and LLM teacher tiers could be audited between self-training rounds, adding the freshly trained gparse and evicting any witness whose errors are systematic rather than random. - After agreement filtering, a separate fluency check (LM perplexity) catches cases where all systems agree on implausible text. — For Mova (back): Generated or compressed-then-regenerated text could pass a perplexity gate before acceptance, catching outputs that are structurally consistent with the graph but unnatural as English. - The model trains only on noisy labels but is tuned and judged solely on clean, human-verified sets, so label noise cannot inflate scores. — For Mova (learn): Silver-trained parsers should always be tuned and evaluated on gold UD treebanks only, keeping silver data out of dev and test splits as a built-in control. *Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.* Fields: [Linguistics & languages](https://md.research.elamur.ai/free/en/shelves/linguistics/) ## Cited by Cited by 4 papers of our library cite it (of the 140884 whose bibliographies we read); 1 of them have a page here, 3 not yet. - [Almieyar: A Culturally Grounded Benchmark for Multi-Dialect Arabic Speech Recognition](https://md.research.elamur.ai/free/en/papers/ghahroodi-2026-almieyar-culturally-grounded-benchmark-multi/) ## Cite The paper itself, not this page. BibTeX from the library's record of the work. ```bibtex @misc{abusheikh2025advancing, title = {{Advancing Arabic Speech Recognition Through Large-Scale Weakly Supervised Learning}}, author = {Abusheikh;, Mahmoud Salhab; Marwan Elghitany; Shameed Sait; Syed Sibghat Ullah; Mohammad and Abusheikh, Hasan}, year = {2025}, doi = {10.48550/arxiv.2504.12254}, eprint = {2504.12254}, archivePrefix = {arXiv}, url = {https://doi.org/10.48550/arxiv.2504.12254} } ``` - [← previous Advancing Arabic Reverse Dictionary Systems: A Transformer-Based Approach with Dataset Construction Guidelines](https://md.research.elamur.ai/free/en/papers/sibaee-2025-advancing-arabic-reverse-dictionary-systems/) - [next → Advancing Dialectal Arabic to Modern Standard Arabic Machine Translation](https://md.research.elamur.ai/free/en/papers/alabdullah-2025-advancing-dialectal-arabic-modern-standard/) - [↑ back to papers of 2025](https://md.research.elamur.ai/free/en/papers/2025/) ## References 1 references, 0 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet. 1. [1] Wajdan Algihab, Noura Alawwad, Anfal Aldawish, and Sarah AlHumoud. Arabic speech recognition with deep learning: A review. In Gabriele Meiselwitz, editor, Social Computing and Social Media. Design, Human Behavior and Analytics, pages 15–31, Cham, 2019. Springer International Publishing. [2] Hamza Kheddar, Mustapha Hemis, and Yassine Himeur. Automatic speech recognition using advanced deep learning approaches: A survey. Information Fusion, page 102422, 2024. [find](https://scholar.google.com/scholar?q=%5B1%5D+Wajdan+Algihab%2C+Noura+Alawwad%2C+Anfal+Aldawish%2C+and+Sarah+AlHumoud%2E+Arabic+speech+recognition+with+deep+learning%3A+A+review%2E+In+Gabriele+Meiselwitz%2C+editor%2C+Social+Computing+and+Social+Media%2E+Design)