from transformers import pipeline

asr = pipeline(
    "automatic-speech-recognition",
    model  = "abidanoaman/whisper-tiny-urdu-merged-data",
    device = 0,   # 0 = GPU, -1 = CPU
)

result = asr("your_urdu_audio.wav")
print(result["text"])

Manual Inference

python

import torch
import torchaudio
from transformers import WhisperForConditionalGeneration, WhisperProcessor

model     = WhisperForConditionalGeneration.from_pretrained("abidanoaman/whisper-tiny-urdu-merged-data")
processor = WhisperProcessor.from_pretrained("abidanoaman/whisper-tiny-urdu-merged-data")
device    = "cuda" if torch.cuda.is_available() else "cpu"
model     = model.to(device)

# Load and preprocess audio
waveform, sr = torchaudio.load("your_urdu_audio.wav")
if waveform.shape[0] > 1:                          # stereo → mono
    waveform = torch.mean(waveform, dim=0, keepdim=True)
if sr != 16000:
    waveform = torchaudio.transforms.Resample(sr, 16000)(waveform)
waveform = waveform.squeeze().numpy()

# Transcribe
inputs   = processor(waveform, sampling_rate=16000, return_tensors="pt")
input_features = inputs.input_features.to(device)

with torch.no_grad():
    pred_ids = model.generate(
        input_features,
        language       = "urdu",
        task           = "transcribe",
        max_new_tokens = 225,
    )

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

Resume / Continue Fine-tuning

python

from transformers import (
    WhisperForConditionalGeneration,
    WhisperProcessor,
    Seq2SeqTrainer,
    Seq2SeqTrainingArguments,
    EarlyStoppingCallback,
)

model     = WhisperForConditionalGeneration.from_pretrained("abidanoaman/whisper-tiny-urdu-merged-data")
processor = WhisperProcessor.from_pretrained("abidanoaman/whisper-tiny-urdu-merged-data")

# Re-apply generation config
model.generation_config.language        = "urdu"
model.generation_config.task            = "transcribe"
model.generation_config.suppress_tokens = []

# Optional: unfreeze CNN layers if adapting to a new domain
# for param in model.model.encoder.conv1.parameters():
#     param.requires_grad = True
# for param in model.model.encoder.conv2.parameters():
#     param.requires_grad = True

# Then set up your dataset, collator, and Seq2SeqTrainer as before.
# See resume_config.json for the full training configuration.

Training Details

Parameter	Value
Base model	openai/whisper-tiny
Dataset	Unified Urdu Speech ASR
Learning rate	3e-06
Effective batch size	16 (8 × 2 grad accum)
Warmup steps	200
Best checkpoint step	2000
Early stopping	patience=3 (eval every 500 steps)
Mixed precision	FP16
Frozen layers	CNN front-end (conv1 + conv2)

Text Normalization

Applied before training and evaluation:

Urdu diacritics (harakat) removed
Arabic character variants standardized (e.g., أ إ آ → ا)
Non-Urdu characters stripped

Files in this Repository

File	Description
`config.json`	Model architecture config
`generation_config.json`	Generation settings (language, task, suppress_tokens)
`model.safetensors`	Fine-tuned model weights
`preprocessor_config.json`	Feature extractor config (mel spectrogram)
`tokenizer_config.json`	Tokenizer settings
`vocab.json`	Whisper multilingual BPE vocabulary
`training_info.json`	Training results and hyperparameters
`resume_config.json`	Full config for resuming/replicating training