STFT Phase Reconstruction in Voiced Speech for an Improved Single-Channel Speech Enhancement

Matlab code

Matlab code for the model-based reconstruction of the spectral phase of voiced speech proposed in

Martin Krawczyk, Timo Gerkmann,"STFT Phase Reconstruction in Voiced Speech for an Improved Single-Channel Speech Enhancement" , IEEE/ACM Trans. Audio, Speech, Language Proc., 2014, Vol. 22, No. 12, pp. 1931-1940, Dec. 2014.

can be found here.

---

Audio examples

The reconstructed phase can for example be employed as a priori information for a phase-sensitive estimation of the complex-valued clean speech spectral coefficients. Audio examples which underline the improvement relative to traditional phase-blind approaches can be found here.

For the following audio examples we however simply combined the reconstructed phase with the noisy amplitude to study the effects of using the reconstruction phase in isolation. Even without any explicit amplitude estimation scheme, modification of the spectral phase alone results in noise reduction between the spectral harmonics during voiced speech. As a reference, we also include the cases where the oracle clean speech phase is used in combination with the noisy amplitude, and a traditional amplitude enhancement scheme.

Babble noise at 0 dB SNR:

Clean Speech	Noisy Speech
Using reconstructed phase up to 2 kHz	Using reconstructed phase up to 4 kHz
Using oracle clean phase	Traditional amplitude enhancement (LSA)

Traffic noise at 5 dB SNR:

Clean Speech	Noisy Speech
Using reconstructed phase up to 2 kHz	Using reconstructed phase up to 4 kHz
Using oracle clean phase	Traditional amplitude enhancement (LSA)

Speech shaped noise at 0 dB SNR:

Clean Speech	Noisy Speech
Using reconstructed phase up to 2 kHz	Using reconstructed phase up to 4 kHz
Using oracle clean phase	Traditional amplitude enhancement (LSA)