Automatic transcription of singing, examples

Author: Matti Ryynänen, homepage
Last modified: Wed May 5 13:57:00 EEST 2004

Transcription method

All the melodies are transcribed with the method described in [1]. The method takes an acoustic input file (in .wav format) and produces a MIDI file, including notes with velocity values, a tempo estimate, and a key signature estimate. The tempo is estimated with the method proposed in [2], and the key estimation method was originally proposed in [3], but simplified in [1]. The transcription method uses

Three-state hidden Markov model to describe note events
Pitch, voicing, and accent as musical features
Note bigrams (with key estimation) for between-note transitions

In addition, vibrato in sung notes is detected and coded as pitch bend messages in MIDI files. The note velocities are determined by the average of RMS energies (mapped to range 0-1) of the voiced frames during notes (variable E) and compressed with mu-law mapping, log(1 + mu * E) / log( 1 + mu ), where mu = 10.

The following figures of the scores were generated by importing the transcribed MIDI files to the Emagic Logic Audio software and then using the graphic export function of the Logic Audio score editor. Notice that the transcriptions were not manually processed in any way, except that the MIDI sequences were shifted in time (with eighth-note shifts) to match the bar lines with the performance, e.g., in cases of upbeat starts in melodies. The score editor quantises note starts and interprets note lengths by given accuracy. In most cases, an eighth-note and an eighth-note triplet score quantiser was set on. Tempo, key signature, or notes were not modified with the sequencer, i.e., Logic Audio was used only for creating the figures, not the synthesised MIDI files in the following.

References

[1] M. Ryynänen, "Probabilistic Modelling of Note Events in the Transcription of Monophonic Melodies," Master's thesis, Tampere University of Technology, March 2004. Available: .pdf .ps
[2] A. P. Klapuri, A. J. Eronen, and J. T. Astola, "Analysis of the meter of acoustic music signals," IEEE Transactions on Speech and Audio Processing, to appear.
[3] T. Viitaniemi, A. Klapuri and A. Eronen. "A probabilistic model for the transcription of single-voice melodies," in Proceedings of the 2003 Finnish Signal Processing Symposium, FINSIG'03, May 2003.

Examples

In the following examples, acoustic inputs and the corresponding transcriptions (synthesised directly from the MIDI files) are provided in mp3 format.

Song: En etsi valtaa loistoa

Singing technique: syllable
Key signature: F major (correctly estimated)
Acoustic input: mp3
Transcription output (synthesised MIDI): mp3
Score:

Comments: an easy transcription case, since only one syllable used. Vibrato found in measures 5 and 9.

Song: Lännen lokari

Singing technique: syllable
Key signature: G minor (correctly estimated)
Acoustic input: mp3
Transcription output (synthesised MIDI): mp3
Score:

Comments: quite easy, although notes are short. Vibrato found in measure 9. However, some adjacent eighth-notes are interpreted as fourth-notes.

Song: Brother, can you spare me a dime

Singing technique: syllable
Key signature: A minor (correctly estimated)
Acoustic input: mp3
Transcription output (synthesised MIDI): mp3
Score:

Comments: some good vibratos and also correctly transcribed notes with accidentals (also a few wrong ones). Usually, the first note is hard to transcribe correctly.

Song: Pieni tytön tylleröinen

Singing technique: lyrics
Key signature: D minor (correctly estimated)
Acoustic input: mp3
Transcription output (synthesised MIDI): mp3
Score:

Comments: in the fifth measure, singing with the lyrics results in additional note boundaries.

Song: Over The Rainbow

Singing technique: syllable
Key signature: C major (correctly estimated)
Acoustic input: mp3
Transcription output (synthesised MIDI): mp3
Score:

Comments: some good vibratos.

Song: Oravan pesä

Singing technique: syllable
Key signature: F major (correctly estimated)
Acoustic input: mp3
Transcription output (synthesised MIDI): mp3
Score:

Comments: The key signature was manually changed to 3/8. Too high octave errors occur in measures 8 and 11, caused by the front-end pitch estimator.

Song: Yesterday

Singing technique: humming
Key signature: F major (correctly estimated)
Acoustic input: mp3
Transcription output (synthesised MIDI): mp3
Score:

Comments: a very good humming transcription.

Song: Puhelinlangat laulaa

Singing technique: lyrics
Key signature: C minor (correctly estimated)
Acoustic input: mp3
Transcription output (synthesised MIDI): mp3
Score:

Comments: In the beginning of the singing performance (measures 1 and 2), there are some inaccuracies which are hard to transcribe correctly.

Song: Hopeinen kuu (Guarda che luna)

Singing technique: lyrics
Key signature: D minor (correctly estimated)
Acoustic input: mp3
Transcription output (synthesised MIDI): mp3
Score:

Comments: The song includes triplets and difficult melodic phrases.

Song: Sävel rakkauden

Singing technique: syllable
Key signature: F major (correctly estimated)
Acoustic input: mp3
Transcription output (synthesised MIDI): mp3
Score:

Comments: A good transcription with vibratos. The tempo estimate was slightly too slow which moved the notes an eighth-note backwards in measures 13-16 (a long melody).