Use Wand (ImageMagick python) before rendering in matplotlib (with spectrograms)

Question

I want to print on paper some spectrograms and i made this function. I want to keep the high quality pdf of text, but at the same this i need to apply some modification (remove white background) to the spectrogram BEFORE rendering in order to keep quality of vector text.

I tried this way, but it doesnt work. Can anyone know if it's possible?

from wand.image import Image, Color
from io import BytesIO
import numpy as np
import matplotlib
matplotlib.use('Agg')  # Use a non-GUI backend
import matplotlib.pyplot as plt
import librosa
import random
import subprocess

def process_plot_spectrum(input, output, onset, dur, color):

    y, sr = librosa.load(input, mono=False, sr=None, offset=onset, duration=dur)
    duration = librosa.get_duration(y=y, sr=sr)

    channels = y.shape[0]

    n_fft = 4096
    hop_length = n_fft // 32

    #result is 3149 × 1773
    #width = width + width - 3149
    #height = height + height - 1773
    #dpi = width / (height / 25.4)
    dpi = 300

    spectrograms = []
    for channel in range(channels):
        #y, n_fft, hop_length = args
        D = librosa.stft(y[channel], n_fft=n_fft, hop_length=hop_length)
        S_db = librosa.amplitude_to_db(np.abs(D), top_db=120, ref=np.max)
        #args = (y, n_fft, hop_length)
        #spectrograms = list(executor.map(lambda channel: plot_spectrogram(channel, args), range(channels)))
        #y_harm, y_perc = librosa.effects.hpss(y[channel])
        spectrograms.append(S_db)

    fig, ax = plt.subplots(nrows=channels, dpi=dpi, figsize=(width // dpi, height // dpi))
    plt.subplots_adjust(hspace=0)

    linewidth = .25
    # Generate logarithmically spaced array of frequencies
    num_freqs = 11  # Number of frequency points
    min_freq = 20  # Minimum frequency in Hz
    max_freq = sr / 2  # Maximum frequency (Nyquist frequency)
    log_freqs = np.logspace(np.log10(min_freq), np.log10(max_freq), num=num_freqs)[:-1]

    times = format_time(duration, onset)
    #print([format_mmss(t) for t in times])
    for channel, S_db in enumerate(spectrograms):
        #print(S_db)
        # Generate film grain noise with the same shape as the spectrogram
        film_grain_noise = np.random.normal(loc=0, scale=5, size=S_db.shape)
        clipped_noise = np.clip(film_grain_noise, -10, 10)  # Adjust the range as needed

        # Add the noise to the spectrogram
        noisy_spectrogram = S_db + clipped_noise
        # Define the parameters for gamma correction

        librosa.display.specshow(noisy_spectrogram,
                        x_axis='time',
                        y_axis='log',
                        ax=ax[channel],
                        n_fft=n_fft,
                        hop_length=hop_length,
                        sr=sr,
                        cmap=plt.get_cmap(color)
                        )

        #ax[channel].set_axis_off()
  
        buf = BytesIO()
        plt.savefig(buf, format='png', dpi=dpi)
        buf.seek(0)

        # Use Wand to enhance exposure
        with Image(file=buf, resolution=dpi) as img:
            fuzz = .105
            img.transparent_color(Color('white'), 0, fuzz=int(np.iinfo(np.uint16).max * fuzz))
            img.trim()
            # Convert the modified image back to a Matplotlib plot
            ax[channel].cla()  # Clear the axes
            ax[channel].imshow(np.asarray(img))

        ax[channel].tick_params(labelsize=7, labelfontfamily='Andale Mono', width=linewidth)

        #ax[channel].grid(True, which='major', axis='x', linewidth=linewidth, color='black')  # Add gridlines
        ax[channel].set_xticks(times-onset)
        ax[channel].set_xticklabels([format_mmss(t) for t in times])  # Set y-axis tick labels

        if channel == 0:
            ax[channel].xaxis.set_ticks_position('top')
        else:
            ax[channel].set_xticks([])
            ax[channel].set_xticklabels([])  # Set y-axis tick labels

        ax[channel].set_xlabel('')
        ax[channel].set_ylabel(f'ch{channel+1}')
        #librosa.display.waveshow(y_harm, sr=sr, ax=ax[channel], color='b')
        #librosa.display.waveshow(y_perc, sr=sr, ax=ax[channel], color='w')
        # Customize the tick labels to add 's' for seconds

        ax[channel].set_yticks(log_freqs)  # Set y-axis ticks as logarithmically spaced frequencies
        ax[channel].set_yticklabels([format_frequency(freq) for freq in log_freqs], fontsize=3.5)  # Set y-axis tick labels

        #ax[channel].set_ylim([0, sr / 2])  # Adjust the y-axis limit to the Nyquist frequency
        ax[channel].xaxis.set_minor_locator(plt.NullLocator())
        ax[channel].yaxis.set_minor_locator(plt.NullLocator())

        # Remove outside borders
        ax[channel].spines['top'].set_visible(False)
        ax[channel].spines['right'].set_visible(False)
        ax[channel].spines['left'].set_visible(False)

        #line_x_adjust = 1/2048  # Adjust the horizontal position of the line
        ax[channel].spines['bottom'].set_capstyle('butt')
        #plt.axhline(y=log_freqs[-1]*(channel+1), color='black', linewidth=linewidth, xmin=line_x_adjust, xmax=1-line_x_adjust)  # Adjust the y-coordinate as needed
        ax[channel].spines['bottom'].set_visible(True if channel < channels-1 else False)

    fig.savefig(output, dpi=dpi, bbox_inches='tight', pad_inches=1/8, format=EXPORT_FORMAT)
    plt.close(fig)
    #return output

process_plot_spectrum(input, output, 0, 1, 'YlGnBu'):