Noise Reduction – Examples
In this section we present spectrograms and audio files demonstrating some noise reduction algorithms of the previous section. The files are recorded at 16 kHz sampling rate and stored in .wav-format.
A short speech sentence is mixed with car noise such that the signal-to-noise ratio is about 5 dB. Please click on the spectrograms below to listen to the audio files.
Speech Signal
The spectrogram of the clean speech signal clearly shows the formants and short speech pauses.
Microphone Signal
The microphone signal is a mix of speech and car noise. The noise is relatively stationary and has its energy concentrated to the lower frequencies.
Spectral Subtraction
Compared to the microphone signal, the non-speech areas are darker, which indicates a lower noise level. However, the residual noise differs very much from the original one and sounds very annoying. In the spectrogram the short spectral peaks and valleys of the noise ('musical noise') can be seen clearly.
Ephraim and Malah
The spectrogram shows the result after processing with one of the algorithms of Ephraim and Malah. As can be seen, the residual noise looks more similar to the original one.
Psychoacoustical Weighting Rule
The spectrogram shows the result after processing with the psychoacoustical weighting rule of [Gustafsson et al.-02]. The residual noise now looks very similar to the original one. Compared to the result of the Ephraim and Malah algorithm a higher residual noise power remains due to the exploitation of the masking effect, resulting in a lower speech distortion.
Audio Files
|
127 K |
|
|
127 K |
|
|
127 K |
|
|
127 K |
|
|
127 K |