import os import numpy as np import librosa from sklearn.ensemble import RandomForestClassifier, GradientBoostingClassifier from sklearn.metrics import confusion$\_$matrix, classification$\_$report import pickle import soundfile as sf from sklearn.model$\_$selection import train$\_$test$\_$split, GridSearchCV, StratifiedKFold import sounddevice as sd from sklearn.preprocessing import StandardScaler from sklearn.feature$\_$selection import SelectKBest, f$\_$classif # Paths to audio files audio$\_$files $=\{$ "Ifrah": $"$C:$/$software test$/$voices$/$Ifra$5.$wav", "Sharonne": $"$C:$/$software test$/$voices$/$Sharonne$1.$wav", "Talha": $"$C:$/$software test$/$voices$/$Talha$\_$converted.wav" $\}$ # Define a function to load an audio file and extract features def extract$\_$features$($audio$\_$path$)$: y$,$ sr $=$ librosa.load$($audio$\_$path$)$ return extract$\_$features$\_$from$\_$array$($y$,$ sr$)$ # Extract features from an audio array def extract$\_$features$\_$from$\_$array$($y$,$ sr$)$: mfccs $=$ np$.$mean$($librosa$.$feature.mfcc$($y$=$y$,$ sr$=$sr$,$ n$\_$mfcc$=13).$T$,$ axis$=0)$ chroma $=$ np$.$mean$($librosa$.$feature.chroma$\_$stft$($y$=$y$,$ sr$=$sr$).$T$,$ axis$=0)$ spectral$\_$contrast $=$ np$.$mean$($librosa$.$feature.spectral$\_$contrast$($y$=$y$,$ sr$=$sr$).$T$,$ axis$=0)$ zero$\_$crossings $=$ np$.$mean$($librosa$.$feature.zero$\_$crossing$\_$rate$($y$=$y$).$T$,$ axis$=0)$ spectral$\_$rolloff $=$ np$.$mean$($librosa$.$feature.spectral$\_$rolloff$($y$=$y$,$ sr$=$sr$).$T$,$ axis$=0)$ rms $=$ np$.$mean$($librosa$.$feature.rms$($y$=$y$).$T$,$ axis$=0)$ mel$\_$spectrogram $=$ np$.$mean$($librosa$.$feature.melspectrogram$($y$=$y$,$ sr$=$sr$).$T$,$ axis$=0)$ spectral$\_$bandwidth $=$ np$.$mean$($librosa$.$feature.spectral$\_$bandwidth$($y$=$y$,$ sr$=$sr$).$T$,$ axis$=0)$ mfcc$\_$delta $=$ np$.$mean$($librosa$.$feature.delta$($librosa$.$feature.mfcc$($y$=$y$,$ sr$=$sr$,$ n$\_$mfcc$=13)).$T$,$ axis$=0)$ features $=$ np$.$hstack$([$mfccs$,$ chroma, spectral$\_$contrast, zero$\_$crossings, spectral$\_$rolloff, rms$,$ mel$\_$spectrogram, spectral$\_$bandwidth, mfcc$\_$delta$])$ return features # Augment audio data def augment$\_$audio$($y$,$ sr$)$: noise $=$ np$.$random.normal$(0,0\mathrm{.}005,$ len$($y$))$ # Reduced noise level y$\_$noisy $=$ y $+$ noise y$\_$pitched $=$ librosa.effects.pitch$\_$shift$($y$,$ sr$=$sr$,$ n$\_$steps$=1)$ # Smaller pitch shift y$\_$speed $=$ librosa.effects.time$\_$stretch$($y$.$astype$(\text{'}$float$32\text{'}),$ rate$=1\mathrm{.}05)$ # Less aggressive time stretch return $[$y$\_$noisy, y$\_$pitched, y$\_$speed$]$ # Function to record and save audio for a speaker def record$\_$and$\_$save$()$: duration $=10$ # seconds print$("$Please record your audio for speaker identification..."$)$ recording $=$ sd$.$rec$($int$($duration $*16000),$ samplerate$=16000,$ channels$=1,$ dtype$=$'float$32\text{'})$ sd$.$wait$()$ filename $=$ "speaker$\_$recorded.wav" sf$.$write$($filename$,$ recording, $16000)$ print$($f$"$Recording saved as $\{$filename$\}")$ return filename # Extract features from uploaded files data $=[]$ labels $=[]$ for label, audio$\_$path in audio$\_$files.items$()$: y$,$ sr $=$ librosa.load$($audio$\_$path$)$ # Original features features $=$ extract$\_$features$($audio$\_$path$)$ data.append$($features$)$ labels.append$($label$)$ # Augmented features augmented$\_$audios $=$ augment$\_$audio$($y$,$ sr$)$ for aug$\_$y in augmented$\_$audios: aug$\_$features $=$ extract$\_$features$\_$from$\_$array$($aug$\_$y$,$ sr$)$ data.append$($aug$\_$features$)$ labels.append$($label$)$ # Convert data and labels to NumPy arrays data $=$ np$.$array$($data$)$ labels $=$ np$.$array$($labels$)$ # Split data into training and testing sets X$\_$train, X$\_$test, y$\_$train, y$\_$test $=$ train$\_$test$\_$split$($data$,$ labels, test$\_$size$=0\mathrm{.}3,$ random$\_$state$=42)$ # Normalize features scaler $=$ StandardScaler$()$ X$\_$train $=$ scaler.fit$\_$transform$($X$\_$train$)$ X$\_$test $=$ scaler.transform$($X$\_$test$)$ # Feature selection to reduce dimensionality and remove noise selector $=$ SelectKBest$($score$\_$func$=$f$\_$classif, k$=50)$ # Increased k to retain more features X$\_$train $=$ selector.fit$\_$transform$($X$\_$train, y$\_$train$)$ X$\_$test $=$ selector.transform$($X$\_$test$)$ # Train a Random Forest Classifier with GridSearch for hyperparameter tuning param$\_$grid $=\{\text{'}$n$\_$estimators': $[50,100,200],$ 'max$\_$depth': $[$None$,10,20],$ 'min$\_$samples$\_$split': $[5,10],$ # Increased to prevent overfitting 'min$\_$samples$\_$leaf': $[2,4]$ # Increased to prevent overfitting $\}$ kfold $=$ StratifiedKFold$($n$\_$splits$=3)$ grid $=$ GridSearchCV$($RandomForestClassifier$($random$\_$state$=42),$ param$\_$grid, cv$=$kfold, refit$=$True, verbose$=3)$ grid.fit$($X$\_$train, y$\_$train$)$ # Use the best model from grid search rf$\_$classifier $=$ grid.best$\_$estimator$\_$ # Train a Gradient Boosting Classifier for comparison gb$\_$classifier $=$ GradientBoostingClassifier$($random$\_$state$=42)$ gb$\_$classifier.fit$($X$\_$train, y$\_$train$)$ # Predict on the test set with Random Forest y$\_$pred$\_$rf $=$ rf$\_$classifier.predict$($X$\_$test$)$ # Predict on the test set with Gradient Boosting y$\_$pred$\_$gb $=$ gb$\_$classifier.predict$($X$\_$test$)$ # Evaluate the Random Forest model conf$\_$matrix$\_$rf $=$ confusion$\_$matrix$($y$\_$test, y$\_$pred$\_$rf$)$ class$\_$report$\_$rf $=$ classification$\_$report$($y$\_$test, y$\_$pred$\_$rf$)$ print$("$Random Forest Confusion Matrix:"$)$ print$($conf$\_$matrix$\_$rf$)$ print$("$

Random Forest Classification Report:"$)$ print$($class$\_$report$\_$rf$)$ # Evaluate the Gradient Boosting model conf$\_$matrix$\_$gb $=$ confusion$\_$matrix$($y$\_$test, y$\_$pred$\_$gb$)$ class$\_$report$\_$gb $=$ classification$\_$report$($y$\_$test, y$\_$pred$\_$gb$)$ print$("$Gradient Boosting Confusion Matrix:"$)$ print$($conf$\_$matrix$\_$gb$)$ print$("$

Gradient Boosting Classification Report:"$)$ print$($class$\_$report$\_$gb$)$ # Save the trained models and scaler with open$("$rf$\_$classifier$\_$model.pkl$","$wb$")$ as model$\_$file: pickle.dump$($rf$\_$classifier, model$\_$file$)$ with open$("$gb$\_$classifier$\_$model.pkl$","$wb$")$ as model$\_$file: pickle.dump$($gb$\_$classifier, model$\_$file$)$ with open$("$scaler$.$pkl$","$wb$")$ as scaler$\_$file: pickle.dump$($scaler$,$ scaler$\_$file$)$ # Save the feature selector with open$("$selector$.$pkl$","$wb$")$ as selector$\_$file: pickle.dump$($selector$,$ selector$\_$file$)$ # Function to record and predict the speaker def record$\_$and$\_$predict$()$: # Record audio filename $=$ record$\_$and$\_$save$()$ # Extract features from the recorded audio features $=$ extract$\_$features$($filename$).$reshape$(1,-1)$ # Load scaler, selector, and model for prediction with open$("$scaler$.$pkl$","$rb$")$ as scaler$\_$file: scaler $=$ pickle.load$($scaler$\_$file$)$ features $=$ scaler.transform$($features$)$ with open$("$selector$.$pkl$","$rb$")$ as selector$\_$file: selector $=$ pickle.load$($selector$\_$file$)$ features $=$ selector.transform$($features$)$ with open$("$rf$\_$classifier$\_$model.pkl$","$rb$")$ as model$\_$file: model $=$ pickle.load$($model$\_$file$)$ prediction $=$ model.predict$($features$)$ print$($f$"$The speaker is predicted to be: $\{$prediction$[0]\}$

$")$ # Record and predict speaker once def main$()$: while True: record$\_$and$\_$predict$()$ cont $=$ input$("$Do you want to identify another speaker? $($yes$/$no$)$: $").$strip$().$lower$()$ if cont $!=$ 'yes': break # Run the main function if $\_\_$name$\_\_=="\_\_$main$\_\_"$: main$()$

Can you correct this code in order to allow the machine to correctly identify the person who records the Audio from the audio that we put in the code to teach the machine our voice the machine must allow the user to record an audio and identify if it is ifrah or sharonne or talha