Voice Recognition MATLAB Projects (Ideas with Complete Code)

% 1. Load Speech & Additive Gaussian Noise
[x, fs] = audioread('speech_sample.wav');
x_noisy = x + 0.05 * randn(size(x));

% 2. 4th-Order Bandpass Butterworth Filter (300 Hz - 3400 Hz)
f_low = 300 / (fs/2); f_high = 3400 / (fs/2);
[b, a] = butter(4, [f_low, f_high], 'bandpass');
x_clean = filtfilt(b, a, x_noisy);

% 3. Formant & Pitch Feature Extraction
f0 = pitch(x_clean, fs, 'Method', 'SRH');
f0_mean = mean(f0(~isnan(f0)));

% 4. FFT Spectral Peak Detection for Formants
N = 2048; X_mag = abs(fft(x_clean(1:min(N, end)), N));
freq_axis = (0:N/2-1)*(fs/N);
[pks, locs] = findpeaks(X_mag(1:N/2), freq_axis, 'MinPeakDistance', 200, 'NPeaks', 3);

fprintf('Estimated Mean Pitch (F0): %.2f Hz\n', f0_mean);
fprintf('Detected Formants: F1 = %.1f Hz, F2 = %.1f Hz\n', locs(1), locs(2));

% 1. Read Unknown Test Word & Reference Template
[test_audio, fs] = audioread('spoken_command.wav');
[ref_audio, ~]   = audioread('template_open.wav');

% 2. Voice Activity Detection (Trim Silence)
idx_test = detectSpeech(test_audio, fs);
test_trimmed = test_audio(idx_test(1,1):idx_test(1,2));

% 3. Compute 13-Coefficient MFCC Feature Matrices
mfcc_test = mfcc(test_trimmed, fs, 'WindowLength', round(0.03*fs), 'OverlapLength', round(0.015*fs));
mfcc_ref  = mfcc(ref_audio, fs, 'WindowLength', round(0.03*fs), 'OverlapLength', round(0.015*fs));

% 4. Dynamic Time Warping (DTW) Distance Alignment
[dist, ix, iy] = dtw(mfcc_test', mfcc_ref');

figure; dtw(mfcc_test(:,1)', mfcc_ref(:,1)');
title(sprintf('DTW Alignment Path (Min Cumulative Distance: %.2f)', dist));

% 1. Load Speech Utterance
[x, fs] = audioread('isolated_digit_three.wav');
frame_len = round(0.025 * fs); % 25 ms frame
frame_hop = round(0.010 * fs); % 10 ms hop

% 2. Frame Signal into Overlapping Segments
frames = buffer(x, frame_len, frame_len - frame_hop, 'nodelay');
num_frames = size(frames, 2);

% 3. Compute Short-Time Energy (STE) & Zero Crossing Rate (ZCR)
ste = sum(frames.^2, 1);
zcr = sum(abs(diff(sign(frames), 1, 1)), 1) / (2 * frame_len);

% 4. Threshold-Based Word Boundary Detection
vad_mask = (ste > 0.05 * max(ste)) & (zcr < 0.4);
word_duration_sec = sum(vad_mask) * (frame_hop / fs);
fprintf('Detected Active Word Duration: %.3f seconds\n', word_duration_sec);

% 1. Extract MFCC Feature Matrices for Query and Database Templates
[query_audio, fs] = audioread('query_command.wav');
[db_audio, ~]     = audioread('reference_command.wav');

coeffs_query = mfcc(query_audio, fs, 'NumCoeffs', 13);
coeffs_db    = mfcc(db_audio, fs, 'NumCoeffs', 13);

% 2. Normalized 2D Cross-Correlation Pattern Matching
q_norm = (coeffs_query - mean(coeffs_query(:))) / std(coeffs_query(:));
db_norm = (coeffs_db - mean(coeffs_db(:))) / std(coeffs_db(:));
corr_matrix = xcorr2(q_norm, db_norm) / (numel(q_norm));

[max_corr, max_idx] = max(corr_matrix(:));
fprintf('Peak Cross-Correlation Coefficient: %.4f\n', max_corr);
if max_corr > 0.85
    disp('Status: Confirmed Pattern Match!');
else
    disp('Status: No Speech Pattern Match.');
end

% 1. Extract MFCC Features from User Verification Utterance
[test_voice, fs] = audioread('user_access_attempt.wav');
test_mfcc = mfcc(test_voice, fs, 'NumCoeffs', 16);

% 2. Evaluate Log-Likelihood against Enrolled Speaker GMM
% Assume 'speaker_gmm_model' is a pre-trained gmdistribution
log_lik = mean(log(pdf(speaker_gmm_model, test_mfcc) + 1e-12));
threshold = -24.5; % Calibrated security decision threshold

if log_lik >= threshold
    fprintf('Access GRANTED: Speaker Verified (Score: %.2f)\n', log_lik);
else
    fprintf('Access DENIED: Impostor Detected (Score: %.2f). Triggering GSM...\n', log_lik);
    % 3. Send SMS Alert via GSM Modem (COM port)
    % s = serialport('COM3', 9600);
    % writeline(s, 'AT+CMGF=1'); pause(0.5);
    % writeline(s, 'AT+CMGS="+1234567890"'); pause(0.5);
    % writeline(s, ['ALERT: Unauthorized voice access detected! ' char(26)]);
end

% 1. Construct Deep Attention Biometric Architecture
layers = [
    imageInputLayer([64 64 1], 'Name', 'spec_input')
    convolution2dLayer(3, 32, 'Padding', 'same', 'Name', 'conv1')
    batchNormalizationLayer('Name', 'bn1')
    reluLayer('Name', 'relu1')
    maxPooling2dLayer(2, 'Stride', 2, 'Name', 'pool1')
    
    % Attention Weighting & Embedding Layers
    convolution2dLayer(3, 64, 'Padding', 'same', 'Name', 'conv2_att')
    reluLayer('Name', 'relu2')
    fullyConnectedLayer(128, 'Name', 'fc_embed')
    dropoutLayer(0.3, 'Name', 'drop1')
    fullyConnectedLayer(50, 'Name', 'fc_speakers') % 50 Enrolled Subjects
    softmaxLayer('Name', 'softmax')
    classificationLayer('Name', 'output')
];

% 2. Training Options & Optimizer
options = trainingOptions('adam', ...
    'MaxEpochs', 25, ...
    'MiniBatchSize', 32, ...
    'InitialLearnRate', 1e-3, ...
    'Plots', 'training-progress');

% 1. Train Universal Background Model (UBM) on Impostor Pool
ubm_components = 32;
ubm = fitgmdist(impostor_mfcc_pool, ubm_components, 'RegularizationValue', 1e-4, 'MaxIter', 100);

% 2. MAP Adaptation for Target Client Speaker
gamma = posterior(ubm, target_speaker_mfcc);
n_k = sum(gamma, 1); % Sufficient statistics
alpha_k = n_k ./ (n_k + 16); % Relevance factor r = 16
adapted_mu = alpha_k' .* (gamma' * target_speaker_mfcc ./ n_k') + (1 - alpha_k') .* ubm.mu;

% 3. Compute Log-Likelihood Ratio (LLR) for Test Utterance
log_p_target = mean(log(pdf(gmdistribution(adapted_mu, ubm.Sigma, ubm.ComponentProportion), test_mfcc) + 1e-12));
log_p_ubm    = mean(log(pdf(ubm, test_mfcc) + 1e-12));
llr_score    = log_p_target - log_p_ubm;

fprintf('Verification LLR Score: %.3f (Decision: %s)\n', llr_score, ternary(llr_score > 0.5, 'VERIFIED', 'REJECTED'));

% 1. Define Deep BiLSTM Architecture for Voice Embeddings
num_features = 40; % 40-band log-mel filterbank energies
num_classes  = 100;

layers = [
    sequenceInputLayer(num_features, 'Name', 'mel_seq')
    bilstmLayer(128, 'OutputMode', 'sequence', 'Name', 'bilstm1')
    bilstmLayer(128, 'OutputMode', 'last', 'Name', 'bilstm2')
    fullyConnectedLayer(128, 'Name', 'x_vector_layer') % Embedding Bottleneck
    reluLayer('Name', 'relu_embed')
    dropoutLayer(0.2, 'Name', 'drop')
    fullyConnectedLayer(num_classes, 'Name', 'fc_out')
    softmaxLayer('Name', 'prob')
    classificationLayer('Name', 'class_output')
];

% 2. Extract Embedding Vector & Compare via Cosine Distance
% emb1 = activations(trainedNet, sample1_seq, 'x_vector_layer');
% emb2 = activations(trainedNet, sample2_seq, 'x_vector_layer');
% cos_sim = 1 - pdist2(emb1, emb2, 'cosine');

% 1. Load Audio Sample (Live Speaker vs Replay Attack)
[audio_in, fs] = audioread('test_recording.wav');

% 2. Extract High-Frequency Spectral Flux, Centroid & Rolloff
sc = spectralCentroid(audio_in, fs);
ro = spectralRolloffPoint(audio_in, fs, 'Threshold', 0.95);
sf = spectralFlux(audio_in, fs);

feat_vec = [mean(sc), std(sc), mean(ro), mean(sf)];

% 3. Evaluate Liveness Using Pre-Trained TreeBagger Ensemble
% [pred_label, scores] = predict(anti_spoof_model, feat_vec);
% fprintf('Liveness Prediction: %s (Confidence: %.2f%%)\n', pred_label{1}, max(scores)*100);

% 1. Initialize Real-Time Audio Streaming Input
deviceReader = audioDeviceReader('SampleRate', 16000, 'SamplesPerFrame', 1024);
setup(deviceReader);

disp('Listening for voice commands... (Say "LIGHTS ON", "FAN OFF", "LOCK")');

% 2. Real-Time Circular Stream Processing Loop
buffer_size = 16000 * 1.5; % 1.5 second buffer
circular_buf = zeros(buffer_size, 1);

for frame = 1:200 % Streaming frames
    audio_frame = deviceReader();
    circular_buf = [circular_buf(1025:end); audio_frame];
    
    % Check for Voice Activity & Keyword Match
    if max(abs(audio_frame)) > 0.12
        % Feature Extraction & Fast Classification
        mfcc_feat = mfcc(circular_buf, 16000);
        % Trigger Hardware Actions...
    end
end
release(deviceReader);

% 1. Extract Acoustic Prosodic & Spectral Features
[audio_in, fs] = audioread('emotion_sample_angry.wav');

f0 = pitch(audio_in, fs); f0 = f0(~isnan(f0));
hnr = harmonicRatio(audio_in, fs);
coeffs = mfcc(audio_in, fs, 'NumCoeffs', 13);

% Statistical Pooling into Fixed-Length Feature Vector
features = [mean(f0), std(f0), max(f0)-min(f0), mean(hnr), mean(coeffs, 1), std(coeffs, 0, 1)];

% 2. Train Multi-Class SVM (ECOC) Classifier
% svm_model = fitcecoc(X_train, y_train_labels);
% pred_emotion = predict(svm_model, features);
% fprintf('Detected Emotion: %s\n', pred_emotion{1});

% 1. Load Speech Frame & Compute LPC All-Pole Coefficients
[x, fs] = audioread('source_speaker.wav');
frame = x(1000:1000+round(0.03*fs)-1) .* hamming(round(0.03*fs));
lpc_order = 12;
a = lpc(frame, lpc_order);

% 2. Extract Residual Glottal Excitation Error
res = filter(a, 1, frame);

% 3. Formant Pole Shifting for Voice Transformation
r = roots(a);
% Scale formant pole angles (frequencies) by factor 1.15
r_modified = abs(r) .* exp(1j * angle(r) * 1.15);
a_modified = poly(r_modified);

% 4. Resynthesize Transformed Speech Frame
frame_converted = filter(1, real(a_modified), res);

figure; [H, f] = freqz(1, a, 512, fs); [H_mod, ~] = freqz(1, real(a_modified), 512, fs);
plot(f, 20*log10(abs(H)), 'b', f, 20*log10(abs(H_mod)), 'r--', 'LineWidth', 1.5);
legend('Original Vocal Tract', 'Shifted Vocal Tract'); xlabel('Frequency (Hz)'); ylabel('Gain (dB)');