Papers

Papers and preprints.

We publish the method, not just the result: corpus construction, the metric, and the failures. Everything is available as a PDF.

Preprint2026Speech recognitionEvaluationDarijaOpen weights

fihrIA-ASR: Open-Weight Speech Recognition and Semantic Evaluation for Moroccan Darija

fihrIA

Modern speech-language models achieve near-human transcription accuracy on high-resource languages with standardized writing systems, yet they remain brittle when deployed on unstandardized, code-switched vernaculars. Moroccan Darija exemplifies this challenge: speakers interleave Arabic, French and English within a single clause, and the same spoken word admits multiple valid spellings across Arabic and Latin scripts. We introduce fihrIA-ASR (0.6B and 1.7B), open-weight encoder–decoder ASR models trained on 1.51 million decontaminated speech clips (≈ 10,000 hours), paired with an 89,000-token Darija-adapted byte-level BPE tokenizer that reduces subword fertility by 42% and accelerates inference at unchanged accuracy. To evaluate recognition fidelity independently of surface orthography, we formalize a span-level Semantic Error Rate (SER) protocol and validate LLM-based SER judges against 2,006 reference–hypothesis pairs annotated by seven native Darija speakers.

Cite this paper
@misc{fihria2026asr,
  title  = {fihrIA-ASR: Open-Weight Speech Recognition and Semantic Evaluation for Moroccan Darija},
  author = {{fihrIA}},
  year   = {2026},
  note   = {Preprint}
}