Kenei/whisper-small-nagamese-v2

Whisper Small — Nagamese ASR

Fine-tuned from openai/whisper-small on a Nagamese speech corpus using LoRA (r=32, alpha=64) and 8-bit quantization on a Kaggle T4 GPU.

The LoRA adapter has been merged into the base model, so this works out-of-the-box without PEFT installed.

Quick Start

from transformers import WhisperForConditionalGeneration, WhisperProcessor
import torch, librosa
 
model     = WhisperForConditionalGeneration.from_pretrained("Kenei/whisper-small-nagamese-v2")
processor = WhisperProcessor.from_pretrained("Kenei/whisper-small-nagamese-v2")
 
audio, _ = librosa.load("your_audio.wav", sr=16000)  # must be 16kHz mono
 
inputs = processor(audio, sampling_rate=16000, return_tensors="pt")
with torch.no_grad():
    predicted_ids = model.generate(inputs.input_features)
 
print(processor.batch_decode(predicted_ids, skip_special_tokens=True)[0])

Training Details

Setting	Value
Base model	openai/whisper-small
Language	Nagamese (Roman script)
LoRA rank / alpha	32 / 64
LoRA target modules	q_proj, v_proj, k_proj, out_proj, fc1, fc2
Effective batch size	8
Max steps	1200
Learning rate	0.001
Precision	fp16 + 8-bit base