app/mfcc.py

import librosa
from transformers import AutoFeatureExtractor, Wav2Vec2BertModel
import soundfile as sf
from sklearn.metrics.pairwise import cosine_similarity
import numpy as np


# Model and feature extractor (same as before)
model_id = "facebook/w2v-bert-2.0"
feature_extractor = AutoFeatureExtractor.from_pretrained(model_id)
model = Wav2Vec2BertModel.from_pretrained(model_id)

def load_and_resample_audio(file_path, target_sample_rate=16000):
    audio_input, sample_rate = sf.read(file_path)
    if sample_rate != target_sample_rate:
        audio_input = librosa.resample(
            audio_input, orig_sr=sample_rate, target_sr=target_sample_rate
        )
    return audio_input, target_sample_rate

def calculate_mfcc(audio_data, sample_rate):
    mfccs = librosa.feature.mfcc(y=audio_data, sr=sample_rate, n_mfcc=13)
    mfccs_scaled = np.mean(mfccs.T, axis=0)  # Average across time dimension
    return mfccs_scaled

def calculate_similarity(mfccs1, mfccs2):
    similarity = cosine_similarity(mfccs1.reshape(1, -1), mfccs2.reshape(1, -1))
    return similarity[0][0]

def mfcc_similarty_check(original: str, recorded: str):
    correct_pronunciation_audio, _ = load_and_resample_audio(original)
    user_pronunciation_audio, sample_rate = load_and_resample_audio(recorded)

    # Extract MFCCs from audio data
    correct_mfccs = calculate_mfcc(correct_pronunciation_audio.flatten(), sample_rate)
    user_mfccs = calculate_mfcc(user_pronunciation_audio.flatten(), sample_rate)

    distance = np.linalg.norm(correct_mfccs.flatten() - user_mfccs.flatten())


    # Calculate cosine similarity using MFCCs
    similarity_score = calculate_similarity(correct_mfccs, user_mfccs)
    accuracy_percentage = similarity_score * 100

    return distance, accuracy_percentage