Issues with MP3-like Compression: Quantization and File Size

Question

I’m trying to implement an MP3-like compression algorithm for audio and have followed the general steps, but I’m encountering a few issues with the quantization step. Here's the overall process I'm following:

1. Apply Hanning window to the audio
2. Apply a filter bank (mp3_forward_fbt).
3. Apply Discrete Cosine Transform (DCT).
4. Quantize the DCT coefficients.
5. Apply Inverse Discrete Cosine Transform (IDCT).
6. Apply the inverse filter bank (mp3_reverse_fbt).

The process works fine, but I’m running into problems during the quantization step. When I remove any number of coefficients (to reduce data), the audio quality degrades significantly. I’ve also tried converting the coefficient array from float to int32 or int16, but the output file size remains the same as the input file size.

I’m looking for insights into what I might be missing or suggestions on how to improve quantization and file size reduction. Here is the code I’m using:

import os
import numpy as np
import scipy.io.wavfile as wav
from scipy.fftpack import dct, idct
import mp3funcs as mp

file_path = os.path.abspath('Projekt_1/audio.wav')
fs, audio = wav.read(file_path)

left_audio = audio[:, 0]
right_audio = audio[:, 1]

window_size = 512
hop_size = window_size // 2
hanning_window = np.hanning(window_size)

def apply_hanning(audio):
    num_frames = (len(audio) - window_size) // hop_size + 1
    windowed_audio = np.zeros(len(audio), dtype=np.float64)

    for i in range(num_frames):
        start = i * hop_size
        end = start + window_size
        windowed_audio[start:end] += audio[start:end] * hanning_window

    return windowed_audio

left_audio = apply_hanning(left_audio)
right_audio = apply_hanning(right_audio)

left_audio = mp.mp3_forward_fbt(left_audio)
right_audio = mp.mp3_forward_fbt(right_audio)

left_dct = dct(left_audio, type=2, n=None, axis=-1, norm='ortho')
right_dct = dct(right_audio, type=2, n=None, axis=-1, norm='ortho')

# quantization here
left_quant = []
right_quant = []

padding_needed_left = (32 - (len(left_quant) % 32)) % 32
left_quant = np.pad(left_quant, (0, padding_needed_left), mode='constant')

padding_needed_right = (32 - (len(right_quant) % 32)) % 32
right_quant = np.pad(right_quant, (0, padding_needed_right), mode='constant')

left_audio = idct(left_quant, type=2, n=None, axis=-1, norm='ortho')
right_audio = idct(right_quant, type=2, n=None, axis=-1, norm='ortho')

left_audio = mp.mp3_reverse_fbt(left_audio)
right_audio = mp.mp3_reverse_fbt(right_audio)

min_length = min(len(left_audio), len(right_audio))  
left_audio = left_audio[:min_length]  
right_audio = right_audio[:min_length]

audio = np.column_stack((left_audio, right_audio))

output_file_path = 'Projekt_1/output.wav'
wav.write(output_file_path, fs, audio.astype(np.int16))

# Test
print("Size difference:", os.path.getsize(output_file_path) - os.path.getsize(file_path))