wav2vec_aligen.py

import torch
import librosa
import os
from model import Wav2Vec2BertForSequenceClassification
from transformers import AutoFeatureExtractor
# from optimum.bettertransformer import BetterTransformer

device = 'cuda' if torch.cuda.is_available() else 'cpu'
# os.environ['PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION'] = 'python'
# os.environ['TRANSFORMERS_NO_ADVISORY_WARNINGS'] = '1'
# os.environ['TRANSFORMERS_VERBOSITY'] = 'error'
torch.random.manual_seed(0); 
# protobuf==3.20.0

model_name = "arslanarjumand/wav2vec-reptiles"
processor = AutoFeatureExtractor.from_pretrained(model_name)
model = Wav2Vec2BertForSequenceClassification.from_pretrained(model_name).to(device)
# model = BetterTransformer.transform(model)

def load_audio(audio_path, processor):
    audio, sr = librosa.load(audio_path, sr=16000)

    input_values = processor(audio, sampling_rate=16000, return_tensors="pt").input_features
    return input_values
        
@torch.inference_mode()
def get_emissions(input_values, model):
    results = model(input_values,).logits[0]
    return results

def vlaidate_range(score):
    score = score if score <= 90 else 90
    score = score if score >= 10 else 10
    return score

def speaker_pronunciation_assesment(audio_path):
    input_values = load_audio(audio_path, processor)
    result_scores = get_emissions(input_values, model)

    pronunciation_score = vlaidate_range(round(result_scores[0].cpu().item()))
    fluency_score           = vlaidate_range(round(result_scores[1].cpu().item()))
    total_score               = vlaidate_range(round(result_scores[2].cpu().item()))
    content_scores         = vlaidate_range(round(result_scores[3].cpu().item()))

     

    result = {'pronunciation_accuracy': pronunciation_score,
              'content_scores': content_scores,
              'total_score': total_score,
              'fluency_score': fluency_score}
    return result

if __name__ == '__main__':
    pass