This work addresses the frequent insertion of English words by Nepali speakers in everyday conversation, a scenario where conventional multilingual speech recognizers often fail at language switches and the script used for English terms. We adapt SraVaani, a 449‑million‑parameter hybrid TDT‑CTC model, using a 172.8‑hour corpus that spans six Nepali speech domains.\n\nThe pretrained tokenizer lacks support for digits and several common symbols. Instead of replacing the tokenizer or discarding recordings containing them, we generate both Nepali and English pronunciations for numbers and employ zero‑shot ASR output to select a plausible form.\n\nFull fine‑tuning reduces the test word error rate (WER) from 48.6% to 17.1%. After a 12‑epoch run, WER reaches 16.3%, comparable to the fine‑tuned Whisper large‑v3‑turbo at 16.4%. Freezing the encoder raises WER to 25.5%, while omitting numeric clips increases numeric‑test WER from 14.3% to 21.3%.\n\nWe also evaluate on the FLEURS benchmark using both its original references and mixed‑script alternatives. The comparison indicates that transcription conventions account for part of the out‑of‑domain performance loss, but not all.\n\nReview: The paper demonstrates that careful data preparation combined with targeted fine‑tuning can dramatically improve recognition accuracy in code‑mixed settings, offering a practical roadmap for leveraging large speech models in low‑resource languages.