sinus_dataset.py

# lets build an audio dataset of random sinus chunks!

def sinus_dataset_generator(num_examples, fs, samples, frequency_range):
    """Builds a dataset of sinus.
    
    Args:
        num_examples: number of examples to generate (int)
        fs: sample rate of the sinus
        samples: number of samples to generate (int)
        frequency_range: a list of two values defining [lower, upper] frequency range (int)
    Returns:
        A numpy array of sinus examples.
    """
    # first example
    sinus_data = (np.sin((2*np.pi*np.arange(samples)*440.0/fs) + 0.0)).astype(np.float32)
    sinus_data = np.reshape(sinus_data, newshape=(1, 1, samples, 1))
    
    for idx in range(0, num_examples-1): 
        # random frequency
        f = np.random.randint(frequency_range[0], frequency_range[1])
        # random phase shift
        phase = np.random.random() * np.pi
        # random gain
        gain = np.random.uniform(0.5, 1.0)

        sinus = (np.sin((2*np.pi*np.arange(samples)*f/fs) + phase) * gain).astype(np.float32)
        # add some noise, mu = 0, sigma = 0.1
        s = np.random.normal(0, 0.1, samples)
        sinus = sinus + s
        # bring it into shape for the model 
        sinus = np.reshape(sinus, newshape=(1, 1, samples, 1))
        sinus_data = np.append(sinus_data, sinus, axis=0)
    return sinus_data

sinus_data = sinus_dataset_generator(4000, RATE, AUDIO_CHUNK_SIZE, [30, 8000])
# split into train and eval dataset, roughly a 70/30 split
split = int(num_examples * 0.7)
train_dataset = tf.data.Dataset.from_tensor_slices((sinus_data[:split], sinus_data[:split]))
eval_dataset = tf.data.Dataset.from_tensor_slices((sinus_data[split:], sinus_data[split:]))