from transformers import WhisperProcessor, WhisperForConditionalGeneration, WhisperTokenizer
import torch
import librosa

# Load model and tokenizer
model = WhisperForConditionalGeneration.from_pretrained("Noobbbbb/whisper-small-bn")
tokenizer = WhisperTokenizer.from_pretrained("Noobbbbb/whisper-small-bn")
processor = WhisperProcessor.from_pretrained("Noobbbbb/whisper-small-bn")

# Load audio (must be 16kHz)
audio, sr = librosa.load("audio.wav", sr=16000)

# Extract features
input_features = processor.feature_extractor(
    audio, 
    sampling_rate=16000, 
    return_tensors="pt"
).input_features

# Generate transcription
with torch.no_grad():
    generated_ids = model.generate(input_features, max_length=448)
    
# Decode
transcription = tokenizer.decode(generated_ids[0], skip_special_tokens=True)
print(transcription)

Training Details

Training Data: openslr37
Language: Bengali (bn)
Training Steps: 2000
Batch Size: 4
Learning Rate: 1e-05
Optimizer: AdamW
eval_wer: 0.3080158337456705

Limitations

Optimized for Bengali speech only
Works best with clear audio at 16kHz sampling rate
May not perform well on heavily accented or noisy audio

Acknowledgments

Based on OpenAI's Whisper model: https://github.com/openai/whisper