Training an Audio Quality Detection Model Using Synthetic Noise and PESQ Scoring

The training of an audio quality detection model faces a shortage of degraded audio datasets and difficulty in evaluating their quality scores. To address this, a method is proposed that creates low‑quality audio solely from a clean high‑quality speech corpus and assigns scores using the PESQ algorithm.

Subjective evaluation of speech quality typically relies on MOS (Mean Opinion Score), which requires many listeners to rate audio on a 1‑5 scale; scores above 4 indicate good quality, while below 3 denote unacceptable quality.

Objective evaluation methods fall into two categories: reference‑based (e.g., PESQ) and no‑reference (e.g., P.563). The presented approach adopts the PESQ algorithm, which compares a degraded signal with its clean reference, aligns them, applies auditory transformations, measures spectral distortion, and maps the result to a MOS‑like PESQ score ranging from –0.5 to 4.5.

Data generation steps:

Randomly select start and end positions for inserting noise into the clean audio.

Calculate the noise scaling factor based on a specified signal‑to‑noise ratio (SNR).

Insert the chosen noise segment into the selected portion of the clean audio.

Compute the PESQ score for the resulting degraded audio.

Implementation details (Python):

def random_sample(n1, n2):
    if n1 < n2:
        start = random.randint(0, n1)
        end = random.randint(start, n1)
    else:
        start = random.randint(0, n2)
        end = random.randint(start, n2)
    return start, end

def add_noise(x, d, SNR):
    P_signal = np.sum(abs(x) ** 2)
    P_d = np.sum(abs(d) ** 2)
    P_noise = P_signal / 10 ** (SNR / 10)
    k = np.sqrt(P_noise / P_d)
    return k

def make_noise_data(high_wave_data, noise_sample_data):
    c_start, c_end = random_sample(len(high_wave_data), len(noise_sample_data))
    n_start = random.randint(0, len(noise_sample_data) - (c_end - c_start))
    n_end = c_end - c_start + n_start
    k = add_noise(high_wave_data, noise_sample_data[n_start:n_end], -10)
    convert_data = high_wave_data[c_start:c_end] + k * noise_sample_data[n_start:n_end]
    new_wave_data = np.concatenate((high_wave_data[:c_start], convert_data, high_wave_data[c_end:]))
    librosa.output.write_wav("noise.wav", new_wave_data, sr)

score = pesq(sr, high_wave_data, low_wave_data, 'nb')

In real‑world scenarios, obtaining a clean reference for every degraded audio segment is impractical, making direct PESQ computation infeasible. By training a neural network on the synthetically generated paired data and their PESQ scores, the model learns to predict audio quality without needing a reference signal, enabling scalable, reference‑free quality assessment.