Noise reduction using pKDTW averaging principle from (2) noisy instances of a speech signal

Signal ("quatre vingt onze"(91))+noise: Signal to Noise Ratio (SNR) is 0dB

x1(t)=s(t)+ε1(t), where ε1 is an iid Gaussian noise with zero mean and variance σ. σ is selected such that the SNR is 0dB.

Signal+noise Signal+Noise Spectrogram Play WAV

Clean Signal and Spectrogram

Signal Spectrogram Play WAV

pKDTW average Signal and Spectrogram (from two noisy instances)

x1(t)=s(t)+ε1(t), x2(t)=s(t+τ)+ε2(t) where ε1 and ε2 are two iid Gaussian noises with zero mean and variance σ, τ is a delay randomly selected in [5,15]. σ is selected such that the SNR is 0dB.

Signal Signal+noise (2 instances) pKDTW average Play WAV

DBA average Signal and Spectrogram (from two noisy instances)

Same conditions as above.

Signal Signal+noise (2 instances) Euclidean average Play WAV

Euclidean average Spectrogram (from two noisy instances)

Same conditions as above. The two noisy instances are "aligned" using cross correlation and the Euclidean average is computed.

Signal Signal+noise (2 instances) Euclidean average Play WAV

Signal + Noise + Low pass filter and Spectrogram

Signal Signal+Noise Filtered Spectrogram Play WAV

Noisy Signal and Spectrogram

Signal Signal+Noise Spectrogram Play WAV