Results of our new sound-restoration algorithms

Old movies often suffer from poor sound quality. It can be caused by pop-ups, nonlinear distortions and background noise. These imperfections disturb the artistic enjoyment and sometimes make the sound incomprehensible.

At the Department of Measurement and Information Systems at BUTE, we have developed digital sound-restoration algorithms to compensate these kind of imperfections. The results will be shown here.

(49 second, mono 44.1 kHz)

This is the original, corrupted signal, taken from the film "Sárga kaszinó". This is not a simulation: a VHS copy of a part from the film was given for us by the Hungarian National Film Archive.

The sound of the movie was recorded by variable density method. The development of the film was wrong and the sound became distorted. In addition to it, the film is seriously injured (split sometimes) and strong pop-ups can be heard in the sound during playing. The signal-to-noise ratio is small.

The restoration had three steps:

(49 second, mono 44.1 kHz)

Signal after pop-up elimination. The original signal contains long pop-ups up to 1000 sample. We have manually selected these pop-ups and eliminated them with our software. Unfortunately this is a not-too-fast algorithm.

(49 second, mono 44.1 kHz)

(7 second, mono 44.1 kHz)(/P)

(7 second, mono 44.1 kHz)(/P)

Signal after pop-up and nonlinear distortion elimination.

VHS is not too good format for sound restoration, due to the additional background noise of the video-tape and the strong low- and high-pass filtering of the video-recorder. However, we could develop a robust method, based on Tikhonov regularization, which method can provide a bit better sound in this case, too. It can work in real-time.

mod21 and mod22 are smaller parts from the film, where the differences are much more audible. mod21 is the distorted file and mod22 is the reconstructed one.

(49 second, mono 44.1 kHz)

Signal after pop-up reduction, distortion elimination and wide-band noise reduction. We used our noise-reduction program. It can work in real-time.

Here you can hear a simulation result with our nonlinear compensation algorithm:

(3 second, mono 44.1 kHz)

(3 second, mono 44.1 kHz)

(3 second, mono 44.1 kHz)

sim_d.wav is the distorted sound. It is distorted with a Gaussian error function and contaminated with white noise. The distortion is so high that the sound is almost incomprehensible.

sim_u.wav is a "restored" sound. Here the exact inverse function was used. During restoration, the noise is extremely amplified.

sim_r.wav is the restored sound that was handled with our algorithm. The algorithm can find a trade-off between the distorted and the noisy sound. The sound is not so distorted, but the noise level remained low.

You can hear, how our noise reduction algorithm works on a Tito Schipa recording:

(44 second, mono 44.1 kHz)

(44 second, mono 44.1 kHz)

schipa1.wav is the original recording from a 78 rpm shellac record.

schipa2.wav is the sound after real-time noise reduction.

Back to my homepage.

Last modification: 2002.01.15. 15:46