Primer on Audio Restoration

This primer describes the main principles associated with high-quality audio data restoration. Usually, this type of restoration is associated with music that has been recorded from an LP, tape, or a microphone. Topics include:


Audio restoration is a very broad term and is therefore difficult to define precisely. Informally, it can be described as follows. Each sound to which we listen includes the original, useful part of the sound, usually referred to as the “signal”, and an additional, unwanted part acquired in the process of transmitting or recording. This unwanted portion of the sound is usually referred to as “noise”. Noise can come from a number of sources. For example, when you make a cell phone call from your car, the signal is your voice. The associated noise can be composed of a number of other sources, including the traffic, the car engine, distortion from the telephone microphone, or distortion from the cellular network. The purpose of audio restoration is to minimize these unwanted noises while at the same time preserving the original signal to the greatest extent possible.

There are different methods for restoring audio. The method you select should be based upon the nature of the distortions found in the audio material and on the source of the audio signal.

This primer focuses on the audio restoration of recorded materials and primarily on recorded musical material. Just 15 years ago, high quality audio restoration activities were confined to a small circle of audio professionals. These activities required specialized labs of equipment and their costs were unreasonable for a typical consumer. When sound digitization and digital signal processing became widely available, these new technologies allowed for the replacement of costly analog equipment with specialized, but very complex, software applications. The recent proliferation of high-speed PCs equipped with quality sound cards has made audio restoration activities even more affordable. Today, digital audio restoration has almost become a mainstream technology. Its close cousin, digital image restoration, is already widely used by a variety of people whose only qualification is that they have access to a PC and a digital camera.

One of the major motivations in developing audio restoration technology has been the drive to digitize LPs and cassettes. Many people have a collection of music favorites stored on these older media. Quite often, they wish to preserve their collection and maintain or improve the quality of the audio recordings that, without intervention, would eventually degrade over time. Since not all old recordings will be released in quality digital format (CD, DVD, etc.), many people want to preserve the collection they already own.

This primer describes various software methods that are used to restore audio recorded from an LP, tape, or from a microphone. When such methods are applied correctly, they will not only help to preserve old recordings in an unaltered condition, but they can also significantly improve the quality of the recordings. Computer-based applications designed for audio restoration purposes are available from many audio software developers. For detailed information on the functionality of these programs, please refer to their associated manuals. This primer discusses only the basic principles of working with these applications — not the individual software products themselves.

Types of Distortion

The surface of most audio data mediums, such as vinyl or magnetic tape, is constantly degrading as a result of long-term use and storage. Any vinyl surface is subject to microfissures, scratches, and soiling. Magnetic tape has even less durability. The base of the tape can become fragile and hard, and the magnetic surface layer can wear out and peel off. Additionally, magnetic tape is subject to demagnetization. Over a period of time, the result is a constant deterioration in the quality of the audio. This deterioration shows itself in unwanted noise, clicks, and crackles.

To stop the deterioration process, audio data can be digitized and saved to a PC’s hard disk or to a CD or DVD. In digital form, audio data can be stored and used for a very long time without deterioration of its quality. The actual time depends on the type of digital medium and a number of other factors.

However, digitalization of audio does not remove any of the distortions or noise found in the audio. All of the distortions that were present when the audio was digitized remain in the resulting digital sound, and this process can actually add distortions to the sound. Audio restoration is the process used to suppress or eliminate such distortions by applying special restoration algorithms to process digital sound.

The first principle of audio restoration is expressed in the adage: “Do not harm”. The application of this principle requires quality restorations be done with great forethought and care. Often, it is better to leave some noise in the recording than to remove a significant portion of good material. Many times, such a process merely replaces an old distortion with a new one. If a high-quality audio restoration cannot be achieved, it is often better to leave the sound material in its original state.

Since human hearing easily adjusts to listening to quiet background noise or small clicks, these noises rarely prevent us from enjoying the music itself. At the same time, many listeners are annoyed by distortions in the audio material, such as the unnatural coloration of sound. The audio restoration methods described in this primer can be applied to obtain quality restorations of audio data, predominantly music.

Unfortunately, the very nature of audio restoration requires a trial and error approach. You cannot just tell the software to make your bad recording sound good. You have to carefully identify the nature of distortions, apply one of the restoration algorithms, and listen to the result. If you are not satisfied with the result, you must change one of the algorithm’s parameters and repeat the process. It may take many iterations to obtain the desired result.

The main types of audio distortions found in recordings made from vinyl, tape, or microphone, fall into one of several categories. To correctly apply an audio restoration algorithm, you must identify the category of audio distortions found in the recording to be restored. You must also recognize the types of distortions that are addressed poorly by software methods, and those distortions that cannot be eliminated by software methods at all.

Note that the following list refers to the most typical distortions that are peculiar to vinyl or magnetic tape. It is not meant to be a definitive list of distortion categories.

The first category of distortions is called Background Noise. This group can be further divided into the following subgroups:

  • Stationary Background Noise - noise with a constant level and frequency, such as tape or microphone hiss, power-line noise, drive motor hum, etc.
  • Non-stationary Background Noise - noise with a variable level and frequency response such as quiet background speech, sound of water, wind, car engine, etc.

A second category of distortions is called Impulsive Noise. This group can be further divided into the following subgroups:

  • Short-duration Noise Pulse - noise that consists of short-duration (up to 3.0 ms) pulses of random amplitude and random duration. This distortion is most typical for vinyl and is usually described using words like pops, clicks, or crackles. Small scratches can also be included in this category.
  • Transient Noise pulse - noise that consists of relatively long-duration noise pulses. The most typical example is long scratches.

A third category of distortions is called Pitch Variation Defects. These are tonal distortions that are usually related to uneven movements of the audio medium, or with deformation of the medium itself. Most often these distortions occur in magnetic tape, but they can also occur with vinyl records that have become warped. These distortions usually occur because of a malfunction of the equipment or damage to the medium, not because of a natural deterioration of recordings over time. Sometimes, such distortions can be eliminated using a software method, but this requires a customized approach; generic restoration applications are not appropriate. This primer does not address the elimination of such distortions.

There are other types of distortion, for example fading of various natures, extraneous sounds and tearing of the signal due to the severe damage of the medium or malfunction of audio equipment (jammed magnetic tape, cracked LP, self-triggering of microphone etc). Such distortions are virtually impossible to eliminate and therefore they are not considered further in this primer.

Thus, we have identified five basic types of distortion that are peculiar to vinyl, magnetic tape and microphone recordings. In the rest of this primer we will mostly discuss Stationary Background Noise and Short-duration Noise Pulses. These two types of distortion are the most typical and can be effectively suppressed by software methods. Other types of distortions are less common. In many cases it’s not possible to remove such distortions and sometimes it’s possible to suppress them, but the process requires a customized approach and is often accompanied by notable loss of quality in the audio material. In most cases, the complexity of the process is not balanced by the resulting improvement of quality.

One exception for Non-stationary Background Noise is a distortion that is typical for vinyl, called Low-Frequency Rumble. This distortion can be suppressed effectively by removing frequencies below 30-40 Hz from a recording. In the case of a vinyl recording, removing these frequencies does not result in a significant loss of useful musical material, especially since these frequencies are usually completely inaudible. These frequencies can sometimes be removed using an equalizer, however, the use of a special high-pass or low-cut filter is recommended. Such filters (high-pass, low-pass, band-pass, high-cut, low-cut, band-cut, notch-filter, etc.) remove specific frequency bands from the sound recording. Users can adjust the boundaries of these bands and set other parameters. These filters are included in many computer applications that have audio restoration capabilities, and they may also be available in sound-editing programs.

The rest of this primer focuses on Stationary Background Noise and Short-duration Noise Pulses. These two types of distortion are the most typical and can be effectively suppressed by software methods.

Removal of Short Duration Noise Pulses

This type of distortion is most often encountered in recordings made from LPs. It can be heard as individual clicks or light crackles, and is caused by microfissures, such as dust and dirt, on the LP’s surface. If your recording contains this type of distortion, then you must address their removal first, before attempting to address other types of distortions. If left untouched, these clicks and crackles can negatively affect the performance of other background-noise removal algorithms.

There are several types of click-removing algorithms. These algorithms generally involve two steps. The first step is the detection of click-type distortions. Algorithms identify these distortions by watching for an abrupt increase in the recording level, also known as an attack on the signal. A parameter within the algorithm identifies exactly how great in size the level increase should be for any point to be identified as a distortion. Most often this parameter is called the Sensitivity or Threshold (refer to your algorithm’s manual for the exact name). To determine the appropriate value for the sensitivity setting, you may need to change this parameter many times and review the results. Setting this parameter incorrectly will either result in many clicks left unaddressed or the algorithm will also begin to modify quick attacks of sound itself (e.g. a snare drum).

After the algorithm has detected a distortion, it attempts to correct the problem. Depending on the type of algorithm, it either tries to replace the short distortion with another piece of sound of similar characteristics, or it will interpolate across the distortion using data from the adjacent (good) pieces of sound. Most algorithms can properly restore a distorted piece if its length does not exceed 3.0 ms.

Some algorithms allow you to set a parameter to define the maximum length of a fragment to be repaired. Such a parameter should be set in the range from 1.5 to 3.0 ms for quality restoration of vinyl recordings. The precise value you use should be based on experimentation with the actual audio data to be restored. Some algorithms have other settings that can affect the quality of the resulting audio. To adjust these settings, refer to the algorithm’s manual.

Removal of Stationary Background Noise

When recordings are made from vinyl, magnetic tape, or from a microphone, Background Noise distortions are usually heard as tape or microphone hiss, power-line noise, and/or drive motor hum.

With speech recordings, the easiest way to suppress stationary background noise is to remove all the frequencies that are not in the normal speech range. These include frequencies below 100-300 Hz and above 4000-5000 Hz. To some extent, this procedure can be performed using an equalizer, however the use of a special band-pass filter that leaves only the applicable frequency ranges in the recording is recommended.

This method, however, is of little use in restoring music recordings. The frequency range in such recordings is very broad and is usually tightly mixed with the frequency bands of stationary background noise. In musical recordings, the most effective method to remove such distortions is the use of algorithms based on FFT (Fast Fourier Transform). These algorithms make changes directly to the frequency spectrum of the recorded sound. To perform correctly, these algorithms require an isolated sample of the noise, without any other sounds, that you wish to remove from your recording.

Again, use of such algorithms involves two steps. The first step is the identification of the noise fragment contained in the recording, which does not contain any other sounds. You must look for a part of the recording that contains the distortion, but does not contain music or speech. The recording that you wish to restore must contain such a noise fragment. Otherwise, you will not be able to achieve a high-quality result using this method. Therefore, you must first identify the location of such fragment in your recording. The algorithm analyzes the fragment and saves its frequency characteristics.

During the second step, the algorithm “subtracts” the frequency characteristics of the noise sample from the frequency content of the entire recording. This significantly decreases the presence of stationary background noise in the recording. The word “subtracts” is given here in quotes because this process is much more complicated than a simple mathematical operation. Usually, you can adjust the amount of “subtracted” noise using the appropriate setting in the application. Again, several iterations using different settings may be required until a satisfactory result is achieved.

It is important to note that FFT-based algorithms are not ideal restoration tools. They all add some amount of distortion to the recording. As a rule, such distortions are audible as either so-called musical noise (resembling quiet babbling of water or dragging of a wire yarn), or unwanted coloration of sound - metal sound. The good news is that modern FFT-based algorithms almost completely moderate such distortions. However, you need to take these added distortions into account. Do not try to remove too much noise from your recording since it will lead to a greater risk of introducing new distortions. It is usually preferred that you leave some noise in the recording while at the same time keeping its vividness and natural sound.

There are other algorithms and methods for stationary background noise removal. Generally, they are effective only for certain types of stationary background noise. For example, a common distortion in recordings made from an old LP or tape is called white noise, which occurs because of the random nature of audio medium degradation. This noise is audible as uniform background hissing in the entire frequency range, and is most commonly referred to as hiss.

A distinctive feature of white noise is that it has an equal representation across all frequencies that are contained in the recording’s spectrum. To remove such noise, you can use FFT-based algorithms and some other algorithms based on the Autoregressive (AR) model of sound. AR algorithms can perform uniform smoothing of the sound and therefore diminish small sound variations that make up the white noise. The advantage of AR algorithms is that they do not add the distortions that usually accompany FFT-based algorithms and they do not require a sample fragment of the distortion. However, AR algorithms can noticeably clip the high-frequency components of a recording and sometimes adds other distortions. Therefore, you need to carefully adjust the settings of these algorithms and be careful not to try and remove too much of the noise component of the recording. These algorithms only relate to the removal of white noise or hiss.

Another common example of specific background noise is power-line noise. This is a uniform hum at 50 or 60 Hz and across multiple frequencies. If your recording has a fragment that contains only such a noise, it can be removed using an FFT-based algorithm. However, it can also be removed using a “notch filter”. This type of filter removes only very narrow frequency bands. Because the bands to be removed are very narrow and specific, the intervention in the sound is quite minimal. A “notch filter” can be used to remove power-line noise without the addition of any significant new distortions.


Unfortunately, there are no specific methods or techniques that will consistently guarantee high-quality restoration of an audio recording. In most cases, you will have to rely on your own hearing, taste, and experience to find the best restoration methods and optimal algorithm settings. If you wish to restore your recordings with the maximum possible quality, you should carefully review the documentation supplied with your audio restoration application and be prepared for some intensive experimentation with different algorithm settings.

You must also take into account that the application of audio restoration algorithms is not reversible, once you save your result. Therefore, it is strongly recommended that you make backup copies of your recordings before embarking on the audio restoration process.