#!/usr/bin/env python
# coding: utf-8

# In[1]:


import numpy as np
from scipy.io.wavfile import write, read
from scipy import signal

import matplotlib.pyplot as plt
from matplotlib import pyplot as mplot

import IPython.display as ipd


# ### Raw data
# 
# https://commons.wikimedia.org/wiki/File:Hello_world_said_by_eSpeakNG.ogg

# In[2]:


rawData = 'resources/audio/original/Hello_world_said_by_eSpeakNG.wav'

originalSampleRate, originalWaveForm = read(rawData)
print('length of original recording = %.2f secs' % (originalWaveForm.shape[0] / originalSampleRate))
print(originalWaveForm.shape)

ipd.Audio(rawData)


# ### Downsample and select relevant data

# In[3]:


downData = 'resources/audio/original/downsampled.Hello_world_said_by_eSpeakNG.wav'

downSample = 2
sampleRate = int(originalSampleRate / downSample)

waveForm = originalWaveForm[::downSample]
waveForm = waveForm[500:6200]

numbPoints = waveForm.shape[0]
print(numbPoints)

write(downData, sampleRate, waveForm)
ipd.Audio(downData)


# ### Optional: *Visualize original audio data*

# In[4]:


plt.plot(np.array(range(numbPoints)) / sampleRate, waveForm, color='black')
plt.xlabel('Time [sec]');


# In[5]:


sample_rate, samples = read(downData)
frequencies, times, spectrogram = signal.spectrogram(samples, sample_rate, nperseg=65, nfft=2048)
plt.pcolormesh(times, frequencies, np.log10(spectrogram + 1), cmap='afmhot')
plt.ylabel('Frequency [Hz]')
plt.xlabel('Time [sec]')
plt.ylim([0, int(sampleRate / 2)]);


# ## Find value of $\alpha$ to fit the original waveform

# In[6]:


# use tau = 12

from helperFunctions import *
from sklearn.preprocessing import MinMaxScaler


# In[7]:


data = waveForm.reshape(-1, 1)

data = data.astype('float64')


# In[8]:


normalizerModel = MinMaxScaler().fit(data)
normalizedData = normalizerModel.transform(data)


# In[9]:


normalizedData.shape


# ### Find initial condition

# In[10]:


t1 = datetime.datetime.now()
decimalInitial = findInitialCondition(normalizedData)
t2 = datetime.datetime.now()
print((t2 - t1).total_seconds())


# In[11]:


decimalInitial


# In[12]:


len(str(decimalInitial))


# ### Generate wave form from initial condition

# In[13]:


t1 = datetime.datetime.now()
decodedValues = generateData(decimalInitial, len(normalizedData))
t2 = datetime.datetime.now()
print((t2 - t1).total_seconds())


# ### Verify error bounds

# In[14]:


maxExpectedError = np.pi / 2 ** (tau - 1)

max([abs(x - y) / maxExpectedError for x, y in zip(decodedValues, normalizedData[:,0])])


# ### Investigate generated values

# In[15]:


saveResultsName = 'resources/audio/generated/HelloWorld_generated.wav'


# In[16]:


write(saveResultsName, sampleRate, np.array(decodedValues))


# In[17]:


ipd.Audio(saveResultsName)


# In[18]:


plt.plot(np.array(range(numbPoints)) / sample_rate, samples, color='darkgreen', linewidth=1)
plt.xlabel('Time [sec]')
plt.title(r'$\alpha$ = %.20f $\, \cdots$' % float(decimalInitial), {'size': 14, 'fontname':'serif'});
plt.grid()

mplot.savefig('resources/audio/generated/generated_waveform.png', bbox_inches='tight')


# In[19]:


sample_rate, samples = read(saveResultsName)
frequencies, times, spectrogram = signal.spectrogram(samples, sample_rate, nperseg=120, nfft=8192)
plt.pcolormesh(times, frequencies, np.log10(spectrogram + 1e-7), cmap='hot', shading='gouraud')
plt.ylabel('Frequency [Hz]')
plt.xlabel('Time [sec]')
plt.title(r'Hello world', {'size': 16, 'fontname':'serif'});
plt.ylim([0, 4000]);

mplot.savefig('resources/audio/generated/generated_spectrogram.png', bbox_inches='tight')