Speech Recognition for Hassaniya Arabic: Adaptation, Matched Baselines, and Orthographic Variation
Year: 2026
Type: preprint
Abstract: Research preprint; not peer reviewed. Hassaniya Arabic remains underrepresented in speech-recognition resources, and variable spelling complicates its evaluation. We adapt Cohere Transcribe Arabic using 13,375 human-transcribed clips comprising 37.42 clip-hours, or approximately 30.3–32.3 hours of distinct audio after accounting for overlapping views. A preceding comparison on common adaptation data favours this backbone over fine-tuned Qwen3-ASR and Gemma configurations, although checkpoint selection introduces an explicit evaluation caveat. We compare the final model with four public recognisers on the same 408 media-test clips and 951 eligible clips from the Mauritanian Casablanca test split. Under a fixed orthography- and spacing-aware policy, adaptation reduces character error rate from 40.5% to 8.1% on the media test and from 42.0% to 18.5% on Casablanca. Corresponding word error rates are 20.7% and 43.6%. We also report less-normalised edit rates and specify the custom alignment used for the orthography-aware scores. These custom scores differ from conventional WER and do not measure semantic accuracy. Hassaniya Speech Test v1 contains mixed Mauritanian media, including television series and broadcast material. Reference anchoring and programme overlap favour the adapted model family; Casablanca provides a different recording domain, but its relation to exploratory model selection requires care. A 40-clip CPU measurement gives 6.1 times real-time throughput with INT8 weights. We plan free access to approximately four hours of general Hassaniya speech, a separate media test set, and scoring materials. An open-source model with redistributable data is planned for the near future; this does not promise release of the checkpoint evaluated here. The full training corpus cannot be redistributed. This Zenodo deposit contains the paper PDF only; datasets and model weights are not included. Author: Mohamed Lamine Bewa, MauriAI, Nouakchott, Mauritania. Contact: [email protected].
Author: Bewa, Mohamed Lamine
Institution: Instituto Superior da Maia
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7222334580
Cites:
Cited by:
Related to: