On October 24, 2026, a study focusing on automatic speech recognition (ASR) for Nepali–English code-switching was published on the arXiv platform (paper ID 2610.30860v1). Addressing key challenges such as script confusion arising from the mixed use of Devanagari and Latin scripts in spontaneous conversations, as well as sharply increased error rates at language-switching points, the study constructed and publicly released a brand-new benchmark dataset.
This benchmark comprises 13,044 spontaneous Nepali–English mixed-language utterances derived from real podcasts, totaling 15.45 hours; the test set contains 10,214 utterances, and the development set includes 2,830. All data have been manually audited and verified to ensure there are no machine‑translation leaks or acoustic loops. The researchers further propose a two‑stage synthetic‑to‑real adaptation framework: in the first stage, supervised fine-tuning (SFT) is performed using 0.95 hours of synthetic speech generated by a multi‑speaker, dual‑engine TTS pipeline; in the second stage, building on the SFT‑converged model, a group‑relative policy optimization (GRPO) approach is employed to jointly optimize transcription accuracy and script‑consistency rewards.
Experiments show that synthetic speech SFT alone can significantly improve cross-domain performance: the mixed error rate (MER) decreases from 23.40% to 21.52%, and the switch-point error rate (SPER) drops from 35.89% to 28.77%. Subsequently, applying GRPO to 173 test sentences (representing 1.69% of the test set) reduces the English script confusion rate (SCR_en) slightly from 6.45% to 6.36% (p ≤ 0.001), while SPER further declines to 28.70% (p = 0.022, cluster-corrected). Ablation analysis reveals that, in the encoder–decoder architecture trained on SFT, since the variance of the synthetic data scripts approaches zero, adding an additional script reward term yields no performance gain. On the monolingual benchmark, SFT improves the Nepali MER by 2.71 percentage points; after GRPO, it stabilizes at 18.42% (a 2.38-percentage-point improvement over the baseline), whereas the English word error rate (WER) remains essentially unchanged (3.38% vs. 3.33%). All validation texts, timestamps, podcast metadata, and reconstructed scripts have been made open source.