himalaya-ai/whisper-large-v3-turbo

Whisper large-v3-turbo -- Nepali ASR (fine-tuned)

Nepali fine-tune of openai/whisper-large-v3-turbo for transcribe.

Results

Metric	Value
Test WER	15.44%
Test CER	10.01%
Best eval WER	16.97%
Steps trained	9368

Training configuration

Parameter	Value
Base model	openai/whisper-large-v3-turbo
Language / task	nepali / transcribe
Epochs	1.0
Learning rate	1e-05
LR schedule	linear with 5% warmup ratio
Per-device train batch	8
Grad accumulation steps	8
Effective batch size	64
Per-device eval batch	8
Precision	bf16
Gradient checkpointing	False
Optimizer	adamw_torch
Weight decay	0.0
Max grad norm	1.0
Max label length	448 tokens
Generation max length	225
Eval / save strategy	per epoch (best model by WER kept)
Seed	42

Data

Parameter	Value
Dataset	lilgoose7777/slr-combined-nepali-tts2
Rows requested	0
Split	80% / 10% / 10% (train/val/test)
Train / val / test examples	599608 / 74951 / 74951
Per-epoch validation subset	500 (test metrics use 2000 test examples)
Audio sampling rate	16 kHz mono
Checkpoint backups	milanakdj/whisper-large-v3-turbo-nepali-checkpoints

The training corpus is clean single-speaker studio audio, so expect degraded accuracy on noisy real-world recordings with background noise, multiple speakers, or strong accents.

Usage

import torch
from transformers import WhisperForConditionalGeneration, WhisperProcessor

model_id = "milanakdj/whisper-large-v3-turbo-nepali-final-corpus"
device   = "cuda" if torch.cuda.is_available() else "cpu"

processor = WhisperProcessor.from_pretrained(model_id, language="nepali", task="transcribe")
model     = WhisperForConditionalGeneration.from_pretrained(model_id).to(device)

# audio: 1-D float32 numpy array at 16kHz
inputs = processor(audio, sampling_rate=16000, return_tensors="pt").to(device)
with torch.inference_mode():
    ids = model.generate(inputs.input_features, max_new_tokens=225)

print(processor.batch_decode(ids, skip_special_tokens=True)[0])