What WER Hides: A Closer Look at Algerian Code-Switched ASR
Abstract
AbstractAutomatic speech recognition for Algerian social-media speech is hard because of dialectal variation, Arabic-French code-switching, and non-standard orthography, yet these difficulties are usually summarised by a single word error rate that hides where recognition fails and whether gains reach information-bearing content. We evaluate seven systems on roughly 30 minutes of Algerian political social-media speech, annotated by language variety as Algerian dialect, Modern Standard Arabic, code-switched, or French: a Whisper-large-v3 baseline, five Whisper variants fine-tuned on different combinations of broadcast, dialectal, and code-switched speech, and the prompted Qwen3-Omni model. The best system reduces overall WER from 60.1% to 31.3%, but still reaches 44.0% on code-switched speech, and named entities remain harder than ordinary tokens, with fewer than half preserved. These results show that ASR difficulty on Algerian social-media speech cannot be read from overall WER alone; it must be measured separately by language variety and by whether named entities survive.