sonique.AUDIO LAB / EXPERIMENT 001About the recipe

THE LISTENING ROOM

Same voice.
A different environment.

Hear how clean speech changes with synthetic noise, room reflections, and reduced bandwidth. Switch a condition. Listen for the difference.

20
English speech excerpts
One clean reference. Seven corruptions.
01 — EXPLORE THE SAMPLESHeadphones recommended

RECORDING 01 / 20

Loading the listening room

Preparing audio and spectrograms.
SIGNAL-TO-NOISE RATIO
Less noise More noise

Additive noise at 10 dB. A lower SNR means more noise relative to the voice.

A

Clean reference

ORIGINAL
Clean speech spectrogram: frequency over time
0:00 / 0:00
CLEAN
B

Noise · 10 dB

SYNTHETIC
Corrupted speech spectrogram: frequency over time
0:00 / 0:00
NOISE
Actual spectrograms · 0–24 kHz · −100 to 0 dBFS · matching time scalesPlay one side at a time
Spectrogram scale & playback notes

Linear frequency, 0–24 kHz. Fixed −100 to 0 dBFS amplitude scale with the same colors for every plot. Hann window, 1,024-point FFT, 240-sample hop. Audio is 48 kHz mono; one common export gain preserves relative levels. The pipeline’s safety limiter is retained and reported for each condition.

02 — UNDER THE HOOD

The training recipe,
made audible.

Each training example receives one effect. The clean speech stays the target; only the input is changed. These are corrupted inputs, not model denoising outputs.

75%

Additive noise

A different noise recording, mixed at a continuously sampled SNR from 0 to 20 dB. The five levels above are listening examples, not fixed training levels.

15%

Reverberation

Room impulse responses add reflections around the voice, while preserving the clean target.

10%

Bandwidth reduction

Reduced frequency range makes speech sound narrower. Compare the high frequencies in the spectrograms.