Transcription (music)

Since the advent of desktop publishing, musicians can acquire music notation software, which can receive the user's mental analysis of notes and then store and format those notes into standard music notation for personal printing or professional publishing of sheet music. Some notation software can accept a Standard MIDI File (SMF) or MIDI performance as input instead of manual note entry. These notation applications can export their scores in a variety of formats like EPS, PNG, and SVG. Often the software contains a sound library that allows the user's score to be played aloud by the application for verification.

Prior to the invention of digital transcription aids, musicians would slow down a record or a tape recording to be able to hear the melodic lines and chords at a slower, more digestible pace. The problem with this approach was that it also changed the pitches, so once a piece was transcribed, it would then have to be transposed into the correct key. Software designed to slow down the tempo of music without changing the pitch of the music can be very helpful for recognizing pitches, melodies, chords, rhythms and lyrics when transcribing music. However, unlike the slow-down effect of a record player, the pitch and original octave of the notes will stay the same, and not descend in pitch. This technology is simple enough that it is available in many free software applications.

The software generally goes through a two-step process to accomplish this. First, the audio file is played back at a lower sample rate than that of the original file. This has the same effect as playing a tape or vinyl record at slower speed – the pitch is lowered meaning the music can sound like it is in a different key. The second step is to use Digital Signal Processing (or DSP) to shift the pitch back up to the original pitch level or musical key.

As mentioned in the Automatic music transcription section, some commercial software can roughly track the pitch of dominant melodies in polyphonic musical recordings. The note scans are not exact, and often need to be manually edited by the user before saving to file in either a proprietary file format or in Standard MIDI File Format. Some pitch tracking software also allows the scanned note lists to be animated during audio playback.

Pitch detection is often the detection of individual notes that might make up a melody in music, or the notes in a chord. When a single key is pressed upon a piano, what we hear is not just one frequency of sound vibration, but a composite of multiple sound vibrations occurring at different mathematically related frequencies. The elements of this composite of vibrations at differing frequencies are referred to as harmonics or partials.

For instance, if we press the Middle C key on the piano, the individual frequencies of the composite's harmonics will start at 261.6 Hz as the fundamental frequency, 523 Hz would be the 2nd Harmonic, 785 Hz would be the 3rd Harmonic, 1046 Hz would be the 4th Harmonic, etc. The later harmonics are integer multiples of the fundamental frequency, 261.6 Hz ( ex: 2 x 261.6 = 523, 3 x 261.6 = 785, 4 x 261.6 = 1046 ). While only about eight harmonics are really needed to audibly recreate the note, the total number of harmonics in this mathematical series can be large, although the higher the harmonic's numeral the weaker the magnitude and contribution of that harmonic. Contrary to intuition, a musical recording at its lowest physical level is not a collection of individual notes, but is really a collection of individual harmonics. That is why very similar-sounding recordings can be created with differing collections of instruments and their assigned notes. As long as the total harmonics of the recording are recreated to some degree, it does not really matter which instruments or which notes were used.

A first step in the detection of notes is the transformation of the sound file's digital data from the time domain into the frequency domain, which enables the measurement of various frequencies over time. The graphic image of an audio recording in the frequency domain is called a spectrogram or sonogram. A musical note, as a composite of various harmonics, appears in a spectrogram like a vertically placed comb, with the individual teeth of the comb representing the various harmonics and their differing frequency values. A Fourier Transform is the mathematical procedure that is used to create the spectrogram from the sound file’s digital data.

The task of many note detection algorithms is to search the spectrogram for the occurrence of such comb patterns (a composite of harmonics) caused by individual notes. Once the pattern of a note's particular comb shape of harmonics is detected, the note's pitch can be measured by the vertical position of the comb pattern upon the spectrogram.

There are basically two different types of music which create very different demands for a pitch detection algorithm: monophonic music and polyphonic music. Monophonic music is a passage with only one instrument playing one note at a time, while polyphonic music can have multiple instruments and vocals playing at once. Pitch detection upon a monophonic recording was a relatively simple task, and its technology enabled the invention of guitar tuners in the 1970s. However, pitch detection upon polyphonic music becomes a much more difficult task because the image of its spectrogram now appears as a vague cloud due to a multitude of overlapping comb patterns, caused by each note's multiple harmonics.

Another method of pitch detection was invented by Martin Piszczalski in conjunction with Bernard Galler in the 1970s[2] and has since been widely followed.[3] It targets monophonic music. Central to this method is how pitch is determined by the human ear.[4] The process attempts to roughly mimic the biology of the human inner ear by finding only but a few of the loudest harmonics at a given instant. That small set of found harmonics are in turn compared against all the possible resultant pitches' harmonic-sets, to hypothesize what the most probable pitch could be given that particular set of harmonics.

To date, the complete note detection of polyphonic recordings remains a mystery to audio engineers, although they continue to make progress by inventing algorithms which can partially detect some of the notes of a polyphonic recording, such as a melody or bass line.

Beat tracking is the determination of a repeating time interval between perceived pulses in music. Beat can also be described as 'foot tapping' or 'hand clapping' in time with the music. The beat is often a predictable basic unit in time for the musical piece, and may only vary slightly during the performance. Songs are frequently measured for their Beats Per Minute (BPM) in determining the tempo of the music, whether it be fast or slow.

Since notes frequently begin on a beat, or a simple subdivision of the beat's time interval, beat tracking software has the potential to better resolve note onsets that may have been detected in a crude fashion. Beat tracking is often the first step in the detection of percussion instruments.

Despite the intuitive nature of 'foot tapping' of which most humans are capable, developing an algorithm to detect those beats is difficult. Most of the current software algorithms for beat detection use a group competing hypothesis for beats-per-minute, as the algorithm progressively finds and resolves local peaks in volume, roughly corresponding to the foot-taps of the music.

To transcribe music automatically, several problems must be solved:

1. Notes must be recognized – this is typically done by changing from the time domain into the frequency domain. This can be accomplished through the Fourier transform. Computer algorithms for doing this are common. The fast Fourier transform algorithm computes the frequency content of a signal, and is useful in processing musical excerpts.

2. A beat and tempo need to be detected (Beat detection)- this is a difficult, many-faceted problem.[5]

The method proposed in Costantini et al. 2009[6] focuses on note events and their main characteristics: the attack instant, the pitch and the final instant. Onset detection exploits a binary time-frequency representation of the audio signal. Note classification and offset detection are based on constant Q transform (CQT) and support vector machines (SVMs). An audio example can be found here.

This in turn leads to a “pitch contour” namely a continuously time-varying line that corresponds to what humans refer to as melody. The next step is to segment this continuous melodic stream to identify the beginning and end of each note. After that, each “note unit” is expressed in physical terms (e.g., 442 Hz, .52 seconds). The final step is then to map this physical information into familiar music-notation-like terms for each note (e.g., an A4, quarter note).

In terms of actual computer processing, the principal steps are to 1) digitize the performed, analog music, 2) do successive short-term, fast Fourier transform (FFTs) to obtain the time-varying spectra, 3) identify the peaks in each spectrum, 4) analyze the spectral peaks to get pitch candidates, 5) connect the strongest individual pitch candidates to get the most likely time-varying, pitch contour, 6) map this physical data into the closest music-notation terms. These fundamental steps, originated by Piszczalski in the 1970s, became the foundation of automatic music transcription. [2]

The most controversial and difficult step in this process is detecting pitch .[7] The most successful pitch methods operate in the frequency domain, not the time domain. While time-domain methods have been proposed, they can break down for real-world musical instruments played in typically reverberant rooms.

The pitch-detection method invented by Piszczalski[4] again mimics human hearing. It follows how only certain sets of partials “fuse” together in human listening. These are the sets that create the perception of a single pitch only. Fusion occurs only when two partials are within 1.5% of being a perfect, harmonic pair (i.e., their frequencies approximate a low-integer pair set such as 1:2, 5:8, etc.) This near harmonic match is required of all the partials in order for a human to hear them as only a single pitch.