Low-Resource Nepali ASR
Parameter-efficient fine-tuning of Whisper for Nepali speech recognition.
Overview
A master's research track comparing LoRA-based parameter-efficient fine-tuning against full fine-tuning of Whisper for Nepali automatic speech recognition, and building a spontaneous-speech benchmark the language currently lacks.
The problem
Nepali is a low-resource language: transcribed audio is scarce, existing benchmarks are read-speech heavy, and full fine-tuning of a large multilingual model is expensive on the hardware most researchers in Nepal can reach.
Approach
- Compare LoRA / PEFT against full fine-tuning across accuracy, trainable parameter count and compute cost.
- Design a spontaneous-speech evaluation benchmark for Nepali — collection protocol, transcription guidelines and dataset preparation.
- Evaluate with Word Error Rate across read and spontaneous conditions to expose the gap between them.
- Document resource-efficiency trade-offs so the pipeline is reproducible on modest GPUs.
Outcome
An evaluation methodology and fine-tuning recipe aimed at making Nepali ASR trainable and measurable without large-scale infrastructure.