This paper describes the NVIDIA NeMo team’s submission to the IWSLT 2026 Simultaneous Speech Translation (SimulST) tracks. We use a cascaded architecture combining a dual-mode Unified ASR Transducer model with a multilingual Large Language Model (LLM). The ASR is trained to deliver stable transcriptions across wide range of latencies, providing a reliable foundation for high-quality LLM translation. Our submission participates in the English–German, English–Italian, and English–Chinese tasks, in both standard and contextualized settings, as well as the Czech–English standard track, covering both low- and high-latency scenarios. We further analyze how ASR and LLM design choices affect the system’s overall latency and translation quality.

Streaming multilingual speech recognition benefits from unified systems that produce numbers in their written form ‘34’ rather than their spoken form ‘thirty-four’. By generating digits directly, these systems eliminate the post-processing latency inherent in cascaded architectures that require a separate inverse text normalization (ITN) step. Arabic presents a formidable challenge for ITN; the system must not only determine the correct numerical value but also navigate complex rules for gender, number, and case marking that are determined by the counted noun. For instance, the digit ‘7’ (as in ‘47’) exhibits gender polarity: it must take a masculine form if modifying a feminine noun (e.g., Halala) and a feminine form if modifying a masculine noun (e.g., Riyal). While Arabic dialects typically exhibit simplified numeral systems by omitting case and gender markers, they vary significantly in verbalization patterns. This study explores the efficacy of a unified streaming Automatic Speech Recognition (ASR) system with integrated ITN features, comparing it against a traditional cascaded approach utilizing a post-processing rule-based ITN module. We utilize a FastConformer cache-aware streaming model trained on English and a diverse Arabic corpus spanning Modern Standard (MSA), dialectal, and Classical Arabic, while maintaining diacritics where contextually appropriate. We evaluate the system using Word Error Rate (WER) for ASR accuracy and exact match for ITN capability. Our results demonstrate that integrating ITN does not degrade core ASR performance and that the unified model achieves accuracy competitive with cascaded systems across Arabic variants. However, error analysis reveals that the primary failures in ITN are rooted in diacritization, gender polarity, and orthographic variation, highlighting the challenges of Arabic’s unique linguistic features in end-to-end modeling.