The Center for Studies in Music Technology at Yale: Research Abstract

Brahms at the Piano:

Denoising, Transcribing and Analyzing the 1889 Cylinder recording of Johannes Brahms

Jonathan Berger
Ronald Coifman
Maxim Goldberg
Charles Nichols

In 1889 Johannes Brahms recorded a segment of the first of his Ungarische Tanze in an arrangement for solo piano.
Using denoising procedures currently under development we attempted to analyze and reconstruct this historic recording in hopes of contributing to the understanding of performance practices in the late nineteenth century.

  • introduction

    On this web page we make available some of our

  • sound materials
  • analytical graphs
  • transcriptions
  • If you are interested in obtaining high quality recordings of our work you can send e-mail to Jonathan Berger.

    We also informally address a number of persistent questions regarding Brahms' recording.Specifically:

  • Why did Brahms choose to play this piece?
  • Why did he begin the recording at such an odd place in the work?
  • Who is really speaking in the introduction?
  • What is the history of the recording?
  • How does the denoising process work?
  • Pointers to more information.
  • Introduction

    On December 2 1889, Theo Wangemann, a representative of Thomas Edison recorded Johannes Brahms introducing himself and performing two segments of music at the piano. The recording took place at the house of the Fellinger family. The works recorded included measures thirteen through seventy two of a solo arrangement of Brahms' Ungarischen Tanz No. 1, and a segment of a paraphrase of Strauss' Libelle.

    A technical description of the denoising process can be found in Berger, Coifman and Goldberg. Removing Noise from Music Using Local Trigonometric Bases and Wavelet Packets, Journal of the Audio Engineering Society, 24:10, 1994. A complete description of our preliminary findings is published in Berger and Nichols. Brahms at the Piano: An Analysis of Data from the Brahms Cylinder, Leonardo Music Journal, vol 4., 1994.
    Return to index.

    Why did Brahms record the First Hungarian Dance?

    Features of Hungarian music became an integral part of his musical style. Brahms performed Hungarian gypsy music on concert tours with the violinist Remİnyi earlier in his career. Originally composed between 1852 and 1864 the four-hand set of twenty dances was transcribed by Brahms for two hands in 1872 and for orchestra in 1874. A violin and piano arrangement was prepared by Joachim. In addition to the the Hunagarian dances Brahms wrote the Variations on a Hungarian Song (op. 21 no. 1) in 1857 and the 11 Zigeunerlieder op. 103 in 1887. The Piano Quartet, op. 25 (1861) features a Rondo alla zingarese and a set of gypsy poems in translation that comprise four of the op. 112 quartets (1891).

    The Hungarian Dances were enormously popular during Brahms' lifetime. Whether or not this had bearing on Brahms' choice of the first dance for the recording is open to conjecture.
    Return to index.

    The History of the Recording (or as much as we could gather of it)

    In 1935 Dr. Fritz Bose attempted to copy the cylinder to LP disc by placing a microphone into the horn of the cylinder playing mechanism and directly cutting a disc. The cylinder was stored at the Preussische Staatsbibliothek. It was thought to be lost from the end of World War II until 1983 but was found by Kowar, Lechleitner and Schiller at what was then the Deutsche Staatsbibliothek, Berlin, GDR. This team attempted unsuccessfully to transfer the cylinder recording to more modern media.

    The British Library National Sound Archive owns acetates of the Brahms recording from the collection of Ludwig Koch. It is asumed that these are direct copies of Bose's attempt at an LP conversion of the recording.

    The team from the Vienna Phonogrammarchiv released two versions of the British Library recording, one with a band pass filtering between 90 and 5000 Hz and another which was highly processed. The latter resulted in a recording that is virtually free of noise but lacks both temporal resolution and significant musical information about the original performance.

    As a result of the restorative efforts by the Vienna Phonogrammarchiv an extensive performance analysis of this work was published by Gerda Lechleitner. Lechleitner attempted to compare data derived from the Brahms recording to other recordings of the work.

    Our analysis of the Brahms cylinder began as a test of an environment for signal analysis and reconstruction using wavelet packets. Although we were primarily concerned in evaluating the efficacy of this method of signal manipulation, we were intrigued by how radically our results differed from those of Lechleitner. In addition we were interested in generalizing the temporal data we gathered into performance attributes in order to determine how consistent Brahms was in his performance to the musical context and structural features of the work.
    Return to index.

    How does the denoising process work?

    In 1946 Denis Gabor, the inventor of the hologram, proposed that conventional representations of signals as continuous functions of time could be supplanted by a representation that combined time and frequency. Considering the finite information present in actual signals he divided this information plane into 'cells'. Each signal can be represented in many ways. Bell curves - like Gaussian curves are modulated in amplitude by sinusoids (called Gabor grains) and were later supplanted by more flexible variants of generalized wave representations. These variants spawned children that amplified characteristics of the signal being represented.

    Reconstruction of a signal is possible by using a paired highpass and lowpass filter formula called a qudrature mirror filter (QMF). Different QMFs produce mother wavelets with varying degrees of frequency resolution.

    WPLab is an application that performs signal analysis with orthogonal trigonometric and wavelet bases. The environment provides a visual interface between a signal waveform representation and the analysis of that signal, as well as a set of tools to reconstruct portions of the signal from the analysis window. WPLab reads ASCII collections of floating-point numbers and 8-bit mulaw or 16-bit linear monophonic sound files.

    We have been actively investigating the applicability of this environment to music, specifically the implications of reconstruction of selectively chosen portions of a musical signal in terms of reconstruction and high level analysis.

    Briefly described, the process of reconstruction is:

  • digitize the musical segment
  • choose QMF
  • perform wavelet analysis
  • selectively reconstruct the soundfile
  • generate a scorefile or MIDI file transcription of the analysis using timing data extracted from the reconstruction process.
  • Currently the selection of wavelet basis is done by trial and error. We are currently developing tools to automate the process of frequency detection and windowing boundaries. We are also exploring the use of linear predictive coding techniques to isolate certain attributes of the signal.
    Return to index.

  • sound examples

    Here are a few sound examples of our work:

  • - a portion of the original recording
  • - a segment of the denoised file (using 4 iterations)
  • - a segment of the denoised file (after averaging)
  • - an example of the overlay of midi reconstruction based on detected onsets If you are interested in obtaining high quality recordings of our work you can send e-mail to Jonathan Berger.
    Return to index.

    More Information

    Further information on denoising and musical applications of local trigonometric transforms and wavelet packets check out our denoising page . There is a wealth of material on wavelets on the WWW. A good place to start is: Wavelets and Time Frequency Analysis

    If you have any questions or comments please email them to: Jonathan Berger

    Back to CSMT Home Page

    Jonathan Berger
    Comments to author:

    All contents copyright (C) 1995.
    Yale Center for Studies in Music Technology.
    All rights reserved.
    Revised: June 1, 1995