whisper-medium-nepali-openslr-ct2

Faster-Whisper (CTranslate2) Version of Nepali Whisper Medium

This repository contains the Int8 Quantized version of the State-of-the-Art Nepali ASR model Dragneel/whisper-medium-nepali-openslr.

It is optimized for high-speed inference on CPUs and GPUs using the faster-whisper or ctranslate2 libraries.

Performance

• WER: 11.17% (OpenSLR Test Set)
• Speed: ~4x faster than the original FP16 model.
• Memory: Requires < 2GB RAM.

Usage

1. Install the library:

   pip install faster-whisper
   

2. Run inference:

   from faster_whisper import WhisperModel
   
   # Load model directly from Hugging Face
   model = WhisperModel("Dragneel/whisper-medium-nepali-openslr-ct2", device="cpu", compute_type="int8")
   
   segments, info = model.transcribe("audio.mp3", beam_size=5)
   
   for segment in segments:
       print(segment.text)
   

This research was supported by the High Performance Computing (HPC) facility at Tribhuvan University, Nepal. We acknowledge the Supercomputer Centre for providing the computational resources required for this wor