BE2M31DSP seminar
Adaptive methods of frequency-domain additive-noise cancellation
(suppression of stationary noise in a non-stationary signal)
- Processed signals (raw binary format, to load the signal into MATLAB use the function loadbin.m)
- Clean non-stationary signals s[n]
- short frame of voiced speech sound vm0.bin (possibility to observe a details in time waveform),
- longer utterance SA001S01.CS0 (possible listening).
- Stationary additive noises n[n]
- low-frequency noise - nc1.bin,
- high-frequency noise - nc2.bin,
- band-noise - nc3.bin.
- Mixture (noisy signal) x[n] should be created by addition of clean signal and noise scaled by the constant k = 0.1, tj. s[n] = s[n] + k*n[n]. Try also a variant with lower level of noise (k = 0.05) or with higher level of noise (k = 0.5).
- Modification of frequency-domain noise cancellation techniques for non-stationary signals
- Algorithm of spectral subtraction does not require any modification. It is already based on subtraction of magnitude noise spectrum from the current short-time spectrum of the noisy speech signal.
- Define frequency response of standard WF for each short-time frame (i.e. work with adaptive WF), similarly define the gain for noise suppression basedon DCT.
- Required spectral characteristics of noise should be estimated from available additive noise (in the case of the processing the signal vm0.bin) or from the first 30 short-time frames of noisy speech signal without speech activity (when the signal SA001S01.CS0 is processed).
- Estimations of power spectral densities of clean speech computed using available clean speech signal (non-realistic limit case).
- To realize realistic solution using noise speech signal only, estimate power spectral densities of clean speech on the basis of spectral subtraction at the level of power DFT or DCT spectra.
- Result :
- Observe waveforms and spectrograms of input and output signals for the case of processing of clean signals
vm0.bin and SA001S01.CS0 with additive noise nc2.bin using Wiener filtering, DCT-based noise cancellation, and spectral subtraction.
- Estimate always SNR of input and output signals (and estimate SNRE as well) and compare obtained results
for all three above mentioned techniques. To compute SNR use available reference clean speech signal and compute the power of speech only from the part with speech activity.
- Compare obtained results also for additive noises nc1.bin and nc3.bin.
- Analyze the impact of smoothing of clean speech PSD estimation using exponential forgeting with the parameter q = 0.5.