python类mfcc()的实例源码

def wav_to_input( wav_file_name ):
    input_data, f_s = sf.read(wav_file_name)
    # mfcc_feat = MFCC_input(mfcc(input_data,f_s))
    mfcc_feat = mfcc(input_data,f_s)
    #Delta features
    delta_feat = mfcc_feat[:-1]-mfcc_feat[1:]
    #Delta-Delta features
    deltadelta_feat = delta_feat[:-1]-delta_feat[1:]

    #Removing the first two frames
    mfcc_feat = mfcc_feat[2:]
    delta_feat = delta_feat[1:]

    #Concatenating mfcc, delta and delta-delta features
    full_input = np.concatenate((mfcc_feat,delta_feat,deltadelta_feat), axis=1)

    return full_input

def get_data(rootdir = TIMIT_main_dir): 
    inputs = []
    targets = []
    for dir_path, sub_dirs, files in os.walk(rootdir):
        for file in files:          
            if (os.path.join(dir_path, file)).endswith('.wav'):
                wav_file_name = os.path.join(dir_path, file)
                input_data, f_s = sf.read(wav_file_name)
                # mfcc_feat = MFCC_input(mfcc(input_data,f_s))
                mfcc_feat = mfcc(input_data,f_s)
                inputs.append(mfcc_feat)#Rakeshvar wants one frame along each column but i am using Lasagne

                text_file_name = wav_file_name[:-4] + '.txt'
                target_data_file = open(text_file_name)
                target_data = str(target_data_file.read()).lower().translate(None, '!:,".;?')
                # target_data = str(target_data_file.read()).lower().translate(str.maketrans('','', '!:,".;?'))
                target_data = target_data[8:-1]#No '.' in lexfree dictionary
                targets.append(target_data)
    return inputs, targets

def get_data(rootdir = TIMIT_main_dir): 
    inputs = []
    targets = []
    for dir_path, sub_dirs, files in os.walk(rootdir):
        for file in files:          
            if (os.path.join(dir_path, file)).endswith('.wav'):
                wav_file_name = os.path.join(dir_path, file)
                input_data, f_s = sf.read(wav_file_name)
                # mfcc_feat = MFCC_input(mfcc(input_data,f_s))
                mfcc_feat = mfcc(input_data,f_s)
                #Delta features
                delta_feat = mfcc_feat[:-1]-mfcc_feat[1:]
                #Delta-Delta features
                deltadelta_feat = delta_feat[:-1]-delta_feat[1:]

                #Removing the first two frames
                mfcc_feat = mfcc_feat[2:]
                delta_feat = delta_feat[1:]

                #Concatenating mfcc, delta and delta-delta features
                full_input = np.concatenate((mfcc_feat,delta_feat,deltadelta_feat), axis=1)

                inputs.append(np.asarray(full_input, dtype=theano.config.floatX))#Rakeshvar wants one frame along each column but i am using Lasagne

                text_file_name = wav_file_name[:-4] + '.txt'
                target_data_file = open(text_file_name)
                target_data = str(target_data_file.read()).lower().translate(None, '!:,".;?')
                # target_data = str(target_data_file.read()).lower().translate(str.maketrans('','', '!:,".;?'))
                target_data = target_data[8:-1]#No '.' in lexfree dictionary
                targets.append(target_data)
    return inputs, targets

def read_wavs_trng(emotions, trng_path, pickle_path, use_pickle=False):
    """
        Utility function to read wav files, convert them into MFCC vectors and store in a pickle file
        (Pickle file is useful in case you re-train on the same data changing hyperparameters)
    """
    trng_data = {}
    if use_pickle and os.path.isfile(pickle_path):
        write_pickle = False
        trng_data = pickle.load(open(pickle_path, "rb"))
    else:
        write_pickle = True
        for emo in emotions:
            mfccs = []
            for wavfile in glob.glob(trng_path + '/' + emo + '/*.wav'):
                rate, sig = wvf.read(wavfile)
                mfcc_feat = mfcc(sig, rate)
                mfccs.append(mfcc_feat)
            trng_data[emo] = mfccs
    if write_pickle:
        pickle.dump(trng_data, open(pickle_path, "wb"))
    return trng_data

def test_emo(test_file, gmms):
    """
        NOTE: Use only after training.
        Test a given file and predict an emotion for it.
    """
    rate, sig = wvf.read(test_file)
    mfcc_feat = mfcc(sig, rate)
    pred = {}
    for emo in gmms:
        pred[emo] = gmms[emo].score(mfcc_feat)
    return emotions_nbest(pred, 2), pred

def run_tests(test_files):
    # Classify input data
    for test_file in test_files:
        # Read input file
        sampling_freq, signal = wavfile.read(test_file)

        # Extract MFCC features
        with warnings.catch_warnings():
            warnings.simplefilter('ignore')
            features_mfcc = mfcc(signal, sampling_freq)

        # Define variables
        max_score = -float('inf') 
        output_label = None 

        # Run the current feature vector through all the HMM
        # models and pick the one with the highest score
        for item in speech_models:
            model, label = item
            score = model.compute_score(features_mfcc)
            if score > max_score:
                max_score = score
                predicted_label = label

        # Print the predicted output 
        start_index = test_file.find('/') + 1
        end_index = test_file.rfind('/')
        original_label = test_file[start_index:end_index]
        print('\nOriginal: ', original_label) 
        print('Predicted:', predicted_label)

def make_split_audio_array(folder, num_splits = 5):
    lst = []
    for filename in os.listdir(folder):
        if filename.endswith('wav'):
            normed_sig = make_standard_length(filename)
            chunk = normed_sig.shape[0]/num_splits
            for i in range(num_splits - 1):
                lst.append(normed_sig[i*chunk:(i+2)*chunk])
    lst = np.array(lst)
    lst = lst.reshape(lst.shape[0], -1)
    return lst

# for input wav file outputs (13, 2999) mfcc np array

def make_normed_mfcc(filename, outrate=8000):
    normed_sig = make_standard_length(filename)
    normed_mfcc_feat = mfcc(normed_sig, outrate)
    normed_mfcc_feat = normed_mfcc_feat.T
    return normed_mfcc_feat

# make mfcc np array from wav file using librosa package

def make_librosa_mfcc(filename):
     y, sr = librosa.load(filename)
     mfcc_feat = librosa.feature.mfcc(y=y, sr=sr, n_mfcc=13)
     return mfcc_feat

# make mfcc np array from wav file using speech features package

def make_mfcc(filename):
    (rate, sig) = wav.read(filename)
    mfcc_feat = mfcc(sig, rate)
    mfcc_feat = mfcc_feat.T
    return mfcc_feat

# for folder containing wav files, output numpy array of normed mfcc

def make_mean_mfcc(filename):
    try:
        (rate, sig) = wav.read(filename)
        mfcc_feat = mfcc(sig, rate)
        avg_mfcc = np.mean(mfcc_feat, axis = 0)
        return avg_mfcc
    except:
        pass

# write new csv corresponding to dataframe of given language and gender

def make_mean_mfcc_df(folder):
    norms = []
    for filename in os.listdir(folder):
        (rate, sig) = wav.read(filename)
        mfcc_feat = mfcc(sig, rate)
        mean_mfcc = np.mean(mfcc_feat, axis = 0)
        #mean_mfcc = np.reshape(mean_mfcc, (1,13))
        norms.append(mean_mfcc)
    flat = [a.ravel() for a in norms]
    stacked = np.vstack(flat)
    df = pd.DataFrame(stacked)
    return df

def build_models(input_folder):
    # Initialize the variable to store all the models
    speech_models = []

    # Parse the input directory
    for dirname in os.listdir(input_folder):
        # Get the name of the subfolder 
        subfolder = os.path.join(input_folder, dirname)

        if not os.path.isdir(subfolder): 
            continue

        # Extract the label
        label = subfolder[subfolder.rfind('/') + 1:]

        # Initialize the variables
        X = np.array([])

        # Create a list of files to be used for training
        # We will leave one file per folder for testing
        training_files = [x for x in os.listdir(subfolder) if x.endswith('.wav')][:-1]

        # Iterate through the training files and build the models
        for filename in training_files: 
            # Extract the current filepath
            filepath = os.path.join(subfolder, filename)

            # Read the audio signal from the input file
            sampling_freq, signal = wavfile.read(filepath)

            # Extract the MFCC features
            with warnings.catch_warnings():
                warnings.simplefilter('ignore')
                features_mfcc = mfcc(signal, sampling_freq)

            # Append to the variable X
            if len(X) == 0:
                X = features_mfcc
            else:
                X = np.append(X, features_mfcc, axis=0)

        # Create the HMM model
        model = ModelHMM()

        # Train the HMM
        model.train(X)

        # Save the model for the current word
        speech_models.append((model, label))

        # Reset the variable
        model = None

    return speech_models

# Define a function to run tests on input files