DOCSLIB.ORG

A Separate and Restore Approach to Score-Informed Music Decomposition

Total Page：16

File Type：pdf, Size：1020Kb

Download full-text PDF Read full-text
A Separate and Restore Approach to Score-Informed Music Decomposition 2015 IEEE Workshop on Applications of Signal Processing to Audio and Acoustics October 18-21, 2015, New Paltz, NY A SEPARATE AND RESTORE APPROACH TO SCORE-INFORMED MUSIC DECOMPOSITION Christian Dittmar, Jonathan Driedger, Meinard Muller¨ International Audio Laboratories Erlangen∗, Am Wolfsmantel 33, 91058 Erlangen, Germany, {christian.dittmar;jonathan.driedger;meinard.mueller}@audiolabs-erlangen.de ABSTRACT the decomposition technique. In this paper, we follow the paradigm of score-informed Non-Negative Matrix Factorization (NMF) to Our goal is to improve the perceptual quality of signal components separate the mixture magnitude spectrogram as described in [3, 4]. extracted in the context of music source separation. Specifically, This procedure involves soft masks used to derive the targeted com- we focus on decomposing polyphonic, mono-timbral piano record- ponent magnitude spectrograms by Wiener filtering. The soft masks ings into the sound events that correspond to the individual notes of are usually obtained from multiplying suitable NMF templates and the underlying composition. Our separation technique is based on activations. In contrast to that, we propose to refine the soft masks score-informed Non-Negative Matrix Factorization (NMF) that has on the basis of a timbre-adapted dictionary of isolated tones. We re- been proposed in earlier works as an effective means to enforce a fer to this extension of the conventional method as restore approach musically meaningful decomposition of piano music. However, the and show that it can be beneficial for certain types of mixtures. method still has certain shortcomings for complex mixtures where The remainder of this paper is organized as follows: Section 2 pro- the tones strongly overlap in frequency and time. As the main con- vides a brief overview of related work, Section 3 reviews score- tribution of this paper, we propose a restoration stage based on re- informed NMF, Section 4 describes our proposed restore approach, fined Wiener filter masks to score-informed NMF. Our idea is to in- Section 5 discusses the experimental results and indicates directions troduce notewise soft masks created from a dictionary of perfectly for future work. isolated piano tones, which are then adapted to match the timbre of the target components. A basic experiment with mixtures of pi- ano tones shows improvements of our novel reconstruction method with regard to perceptually motivated separation quality metrics. A second experiment with more complex piano recordings shows that 2. RELATED WORK further investigations into the concept are necessary for real-world applicability. Music source separation using score information has first been in- Index Terms— Score-informed music processing, source sep- troduced in [5] and [6]. Related approaches used tensor factoriza- aration, music decomposition, signal reconstruction. tion [7] or synthesized music for component initialization [8]. An important starting point for our work is the procedure for score- informed music decomposition described in [3], where the authors 1. INTRODUCTION describe how to impose musically meaningful constraints on the components of a Non-Negative Matrix Factorization (NMF) with- The goal of music source separation is to decompose a music out the need for a dedicated component training. The same principle recording into its constituent signal components [1, 2]. We focus is extended to note-wise decomposition in [4]. Other authors de- on the special case of polyphonic, mono-timbral piano recordings, vised elaborate source-filter models to account for the time-varying where we aim to extract sound events that correspond to the com- spectral envelope of components with fixed pitch [9, 10]. In [11], a position’s individual notes. We assume that each sound event corre- semi-adaptive NMF variant, which allows to efficiently capture the sponding to a musical note can be characterized as a harmonic tone temporal evolution of component spectrograms, was proposed. with constant pitch as well as a sharp attack, a stable sustain, and a In [12], Ewert discussed the problems inherent to NMF decompo- release phase. Moreover, the mixture signal is assumed to be a lin- sition in case of overlapping partials of the targeted components. ear superposition of the isolated tones. This is, of course, a simpli- He proposed to use the activations and gains computed by a first fication since we neglect acoustic effects such as room impulse re- NMF stage to infer a time-frequency (TF) dependent weighting of sponses and sympathetic string resonances in the piano. With these the mixture magnitude spectrogram accounting for possible phase assumptions, the decomposition of piano music into isolated tones cancellations. He could show that this leads to more meaningful de- constitutes a limited, yet challenging source separation task, which compositions in a second NMF stage. The use of a weighted NMF may pave the way to more complex scenarios. had already been proposed by Virtanen in [13], where it was used as One of the challenges is to improve the perceptual quality of the sep- a means to fill gaps in the magnitude spectrogram that occurred due arated signals which may suffer from audible artifacts, depending to binary masking of predominant pitched signal components. In on the complexity of the music, the recording conditions, as well as [14], Cano et al. investigated the complex mutual influence of mag- ∗ nitude and phase on the quality of separated signals in source sep- This work has been supported by the German Research Foundation (DFG MU 2686/6-1). The International Audio Laboratories Erlangen aration. Cano proposed to soften the additivity constraint in source (AudioLabs) is a joint institution of the Friedrich-Alexander-Universitat¨ separation and suggested to use instrument specific resynthesis ap- Erlangen-Nurnberg¨ (FAU) and Fraunhofer IIS. proaches in [15]. 2015 IEEE Workshop on Applications of Signal Processing to Audio and Acoustics October 18-21, 2015, New Paltz, NY stood element-wise. Furthermore, J ∈ RK×M denotes an all-one matrix. 11840 3.2. Constraint Components via Score-Informed NMF 5920 2960 W(0) H(0) 1480 = Proper initialization of and is an effective means to con- 740 strain the degrees of freedom in the NMF iterations and enforces 392 Frequency (Hz) 0 1 2 3 4 convergence to a desired, musically meaningful solution. One pos- Time (Seconds) sibility is to impose constraints derived from a time-aligned, sym- 11840 G4 (392 Hz) A4 (440 Hz) bolic representation (i.e., machine readable score) of the recording 5920 ଵ ଶ 2960 [7]. Three constraints can be obtained from the musical score. First, 1480 + + the rank R of the decomposition is chosen according to the num- (0) 740 ber of unique musical pitches. Second, each column of W is 392 initialized with a prototype harmonic overtone series reflecting the 11840 5920 B4 (494 Hz) D5 (587 Hz) expected nature of a musical tone corresponding to the assigned ଷ ସ (0) 2960 musical pitch. Third, the rows of H are initialized as follows: 1480 + C ∈ RR×M 740 A binary constraint matrix s is constructed for each 392 s ∈ [1 : S], where Cs is 1 at entries that correspond to the pitch and 0 1 2 3 4 0 1 2 3 4 temporal position of the sth aligned note event and 0 otherwise. The (0) union (OR-sum) of all Cs is then used as initialization of H . With Figure 1: Artificial piano score used as illustrative example through- this initialization, each template obtained from iteration (1) typi- out the paper. Our target is the extraction of the isolated magnitude cally corresponds to an average spectrum (usually 1-normalized spectrograms corresponding to the individual notes G4, A4, B4, and [2]) of the corresponding musical pitch and each activation function D5 (ordered by pitch instead of onset time). obtained from (2) corresponds to the temporal amplitude envelope of all occurrences of that particular pitch throughout the recording. 3. SEPARATE 4. RESTORE In this section, we summarize the score-informed NMF approach as described in [3, 4]. In our signal model, we assume that the given Score-informed NMF as described in Section 3.2 yields a decom- V W(L) piano recording x is a linear mixture of notewise audio events xs, position of into musically meaningful templates and acti- (L) s ∈ [1 : S], where S ∈ N is the number of musical notes speci- vations H . In the following, we discuss the issues inherent to the V fied in the musical score (see our example score in Figure 1) such restoration of our targeted s from the components and introduce that x := s xs. Let X (m, k), m, k ∈ Z, be a complex-valued our extension to the conventional procedure. TF coefficient at the mth time frame and kth frequency bin of the Short-Time Fourier Transform (STFT) of our mixture signal x. Let 4.1. Component Magnitude Spectrogram Reconstruction T K×M V:=|X | ∈ R≥0 be a transposed version of the mixture sig- nal’s magnitude spectrogram. Our objective is to decompose V into As shown on the left hand side of Figure 2(a), we can use the component magnitude spectrograms Vs that correspond to the indi- results of score-informed NMF to reconstruct magnitude spec- VNMF ≈ vidual note events xs. Ignoring possible phase issues, we assume trograms corresponding to individual note objects as s (L) (L) that the additive relationship V:= s Vs is fulfilled (see Figure 1). W H Cs . In the conceptual illustration, the template and activation corresponding to the sth tone are indicated by a 3.1. Music Decomposition via NMF hatched column and row, respectively. The binary activation in the V corresponding Cs is visualized by a black box inside the hatched NMF can be used to decompose the magnitude spectrogram VNMF into spectral basis functions (also called templates) encoded by the row.
Load more

Recommended publications
  • The Science of String Instruments
    The Science of String Instruments Thomas D. Rossing Editor The Science of String Instruments Editor Thomas D. Rossing Stanford University Center for Computer Research in Music and Acoustics (CCRMA) Stanford, CA 94302-8180, USA [email protected] ISBN 978-1-4419-7109-8 e-ISBN 978-1-4419-7110-4 DOI 10.1007/978-1-4419-7110-4 Springer New York Dordrecht Heidelberg London # Springer Science+Business Media, LLC 2010 All rights reserved. This work may not be translated or copied in whole or in part without the written permission of the publisher (Springer Science+Business Media, LLC, 233 Spring Street, New York, NY 10013, USA), except for brief excerpts in connection with reviews or scholarly analysis. Use in connection with any form of information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology now known or hereafter developed is forbidden. The use in this publication of trade names, trademarks, service marks, and similar terms, even if they are not identified as such, is not to be taken as an expression of opinion as to whether or not they are subject to proprietary rights. Printed on acid-free paper Springer is part of Springer ScienceþBusiness Media (www.springer.com) Contents 1 Introduction............................................................... 1 Thomas D. Rossing 2 Plucked Strings ........................................................... 11 Thomas D. Rossing 3 Guitars and Lutes ........................................................ 19 Thomas D. Rossing and Graham Caldersmith 4 Portuguese Guitar ........................................................ 47 Octavio Inacio 5 Banjo ...................................................................... 59 James Rae 6 Mandolin Family Instruments........................................... 77 David J. Cohen and Thomas D. Rossing 7 Psalteries and Zithers .................................................... 99 Andres Peekna and Thomas D.
    [Show full text]
  • Vapar Synth--A Variational Parametric Model for Audio Synthesis
    VAPAR SYNTH - A VARIATIONAL PARAMETRIC MODEL FOR AUDIO SYNTHESIS Krishna Subramani , Preeti Rao Alexandre D’Hooge Indian Institute of Technology Bombay ENS Paris-Saclay [email protected] [email protected] ABSTRACT control over the perceptual timbre of synthesized instruments. With the advent of data-driven statistical modeling and abun- With the similar aim of meaningful interpolation of timbre in dant computing power, researchers are turning increasingly to audio morphing, Engel et al. [4] replaced the basic spectral deep learning for audio synthesis. These methods try to model autoencoder by a WaveNet [5] autoencoder. audio signals directly in the time or frequency domain. In the In the present work, rather than generating new timbres, interest of more flexible control over the generated sound, it we consider the problem of synthesis of a given instrument’s could be more useful to work with a parametric representation sound with flexible control over the pitch. Wyse [6] had the of the signal which corresponds more directly to the musical similar goal in providing additional information like pitch, ve- attributes such as pitch, dynamics and timbre. We present Va- locity and instrument class to a Recurrent Neural Network to Par Synth - a Variational Parametric Synthesizer which utilizes predict waveform samples more accurately. A limitation of a conditional variational autoencoder (CVAE) trained on a suit- his model was the inability to generalize to notes with pitches able parametric representation. We demonstrate1 our proposed the network has not seen before. Defossez´ et al. [7] also model’s capabilities via the reconstruction and generation of approached the task in a similar fashion, but proposed frame- instrumental tones with flexible control over their pitch.
    [Show full text]
  • A Comparison of Viola Strings with Harmonic Frequency Analysis
    University of Nebraska - Lincoln DigitalCommons@University of Nebraska - Lincoln Student Research, Creative Activity, and Performance - School of Music Music, School of 5-2011 A Comparison of Viola Strings with Harmonic Frequency Analysis Jonathan Paul Crosmer University of Nebraska-Lincoln, [email protected] Follow this and additional works at: https://digitalcommons.unl.edu/musicstudent Part of the Music Commons Crosmer, Jonathan Paul, "A Comparison of Viola Strings with Harmonic Frequency Analysis" (2011). Student Research, Creative Activity, and Performance - School of Music. 33. https://digitalcommons.unl.edu/musicstudent/33 This Article is brought to you for free and open access by the Music, School of at DigitalCommons@University of Nebraska - Lincoln. It has been accepted for inclusion in Student Research, Creative Activity, and Performance - School of Music by an authorized administrator of DigitalCommons@University of Nebraska - Lincoln. A COMPARISON OF VIOLA STRINGS WITH HARMONIC FREQUENCY ANALYSIS by Jonathan P. Crosmer A DOCTORAL DOCUMENT Presented to the Faculty of The Graduate College at the University of Nebraska In Partial Fulfillment of Requirements For the Degree of Doctor of Musical Arts Major: Music Under the Supervision of Professor Clark E. Potter Lincoln, Nebraska May, 2011 A COMPARISON OF VIOLA STRINGS WITH HARMONIC FREQUENCY ANALYSIS Jonathan P. Crosmer, D.M.A. University of Nebraska, 2011 Adviser: Clark E. Potter Many brands of viola strings are available today. Different materials used result in varying timbres. This study compares 12 popular brands of strings. Each set of strings was tested and recorded on four violas. We allowed two weeks after installation for each string set to settle, and we were careful to control as many factors as possible in the recording process.
    [Show full text]
  • Timbre Perception
    HST.725 Music Perception and Cognition, Spring 2009 Harvard-MIT Division of Health Sciences and Technology Course Director: Dr. Peter Cariani Timbre perception www.cariani.com Friday, March 13, 2009 overview Roadmap functions of music sound, ear loudness & pitch basic qualities of notes timbre consonance, scales & tuning interactions between notes melody & harmony patterns of pitches time, rhythm, and motion patterns of events grouping, expectation, meaning interpretations music & language Friday, March 13, 2009 Wikipedia on timbre In music, timbre (pronounced /ˈtæm-bər'/, /tɪm.bər/ like tamber, or / ˈtæm(brə)/,[1] from Fr. timbre tɛ̃bʁ) is the quality of a musical note or sound or tone that distinguishes different types of sound production, such as voices or musical instruments. The physical characteristics of sound that mediate the perception of timbre include spectrum and envelope. Timbre is also known in psychoacoustics as tone quality or tone color. For example, timbre is what, with a little practice, people use to distinguish the saxophone from the trumpet in a jazz group, even if both instruments are playing notes at the same pitch and loudness. Timbre has been called a "wastebasket" attribute[2] or category,[3] or "the psychoacoustician's multidimensional wastebasket category for everything that cannot be qualified as pitch or loudness."[4] 3 Friday, March 13, 2009 Timbre ~ sonic texture, tone color Paul Cezanne. "Apples, Peaches, Pears and Grapes." Courtesy of the iBilio.org WebMuseum. Paul Cezanne, Apples, Peaches, Pears, and Grapes c. 1879-80); Oil on canvas, 38.5 x 46.5 cm; The Hermitage, St. Petersburg Friday, March 13, 2009 Timbre ~ sonic texture, tone color "Stilleben" ("Still Life"), by Floris van Dyck, 1613.
    [Show full text]
  • Automatic Transcription of Bass Guitar Tracks Applied for Music Genre Classification and Sound Synthesis
    Automatic Transcription of Bass Guitar Tracks applied for Music Genre Classification and Sound Synthesis Dissertation zur Erlangung des akademischen Grades Doktoringenieur (Dr.-Ing.) vorlelegt der Fakultät für Elektrotechnik und Informationstechnik der Technischen Universität Ilmenau von Dipl.-Ing. Jakob Abeßer geboren am 3. Mai 1983 in Jena Gutachter: Prof. Dr.-Ing. Gerald Schuller Prof. Dr. Meinard Müller Dr. Tech. Anssi Klapuri Tag der Einreichung: 05.12.2013 Tag der wissenschaftlichen Aussprache: 18.09.2014 urn:nbn:de:gbv:ilm1-2014000294 ii Acknowledgments I am grateful to many people who supported me in the last five years during the preparation of this thesis. First of all, I would like to thank Prof. Dr.-Ing. Gerald Schuller for being my supervisor and for the inspiring discussions that improved my understanding of good scientific practice. My gratitude also goes to Prof. Dr. Meinard Müller and Dr. Anssi Klapuri for being available as reviewers. Thank you for the valuable comments that helped me to improve my thesis. I would like to thank my former and current colleagues and fellow PhD students at the Semantic Music Technologies Group at the Fraunhofer IDMT for the very pleasant and motivating working atmosphere. Thank you Christian, Holger, Sascha, Hanna, Patrick, Christof, Daniel, Anna, and especially Alex and Estefanía for all the tea-time conversations, discussions, last-minute proof readings, and assistance of any kind. Thank you Paul for providing your musicological expertise and perspective in the genre classification experiments. I also thank Prof. Petri Toiviainen, Dr. Olivier Lartillot, and all the collegues at the Finnish Centre of Excellence in Interdisciplinary Music Research at the University of Jyväskylä for a very inspiring research stay in 2010.
    [Show full text]
  • Advances in Perceptual Stereo Audio Coding Using Linear Prediction Techniques
    Advances in Perceptual Stereo Audio Coding Using Linear Prediction Techniques PROEFSCHRIFT ter verkrijging van de graad van doctor aan de Technische Universiteit Eindhoven, op gezag van de Rector Magnificus, prof.dr.ir. C.J. van Duijn, voor een commissie aangewezen door het College voor Promoties in het openbaar te verdedigen op dinsdag 15 mei 2007 om 16.00 uur door Arijit Biswas geboren te Calcutta, India Dit proefschrift is goedgekeurd door de promotoren: prof.dr. R.J. Sluijter en prof.Dr. A.G. Kohlrausch Copromotor: dr.ir. A.C. den Brinker °c Arijit Biswas, 2007. All rights are reserved. Reproduction in whole or in part is prohibited without the written consent of the copyright owner. The work described in this thesis has been carried out under the auspices of Philips Research Europe - Eindhoven, The Netherlands. CIP-DATA LIBRARY TECHNISCHE UNIVERSITEIT EINDHOVEN Biswas, Arijit Advances in perceptual stereo audio coding using linear prediction techniques / by Arijit Biswas. - Eindhoven : Technische Universiteit Eindhoven, 2007. Proefschrift. - ISBN 978-90-386-2023-7 NUR 959 Trefw.: signaalcodering / signaalverwerking / digitale geluidstechniek / spraakcodering. Subject headings: audio coding / linear predictive coding / signal processing / speech coding. Samenstelling promotiecommissie: prof.dr. R.J. Sluijter, promotor Technische Universiteit Eindhoven, The Netherlands prof.Dr. A.G. Kohlrausch, promotor Technische Universiteit Eindhoven, The Netherlands dr.ir. A.C. den Brinker, copromotor Philips Research Europe - Eindhoven, The Netherlands prof.dr. S.H. Jensen, extern lid Aalborg Universitet, Denmark Dr.-Ing. G.D.T. Schuller, extern lid Fraunhofer Institute for Digital Media Technology, Ilmenau, Germany dr.ir. S.J.L. van Eijndhoven, lid TU/e Technische Universiteit Eindhoven, The Netherlands prof.dr.ir.
    [Show full text]
  • Expressive Sampling Synthesis
    i THESE` DE DOCTORAT DE l’UNIVERSITE´ PIERRE ET MARIE CURIE Ecole´ doctorale Informatique, T´el´ecommunications et Electronique´ (EDITE) Expressive Sampling Synthesis Learning Extended Source–Filter Models from Instrument Sound Databases for Expressive Sample Manipulations Presented by Henrik Hahn September 2015 Submitted in partial fulfillment of the requirements for the degree of DOCTEUR de l’UNIVERSITE´ PIERRE ET MARIE CURIE Supervisor Axel R¨obel Analysis/Synthesis group, IRCAM, Paris, France Reviewer Xavier Serra MTG, Universitat Pompeu Fabra, Barcelona, Spain Vesa V¨alim¨aki Aalto University, Espoo, Finland Examiner Sylvain Marchand Universit´ede La Rochelle, France Bertrand David T´el´ecom ParisTech, Paris, France Jean-Luc Zarader UPMC Paris VI, Paris, France This page is intentionally left blank. Abstract This thesis addresses imitative digital sound synthesis of acoustically viable in- struments with support of expressive, high-level control parameters. A general model is provided for quasi-harmonic instruments that reacts coherently with its acoustical equivalent when control parameters are varied. The approach builds upon recording-based methods and uses signal trans- formation techniques to manipulate instrument sound signals in a manner that resembles the behavior of their acoustical equivalents using the fundamental control parameters intensity and pitch. The method preserves the inherent quality of discretized recordings of a sound of acoustic instruments and in- troduces a transformation method that retains the coherency with its timbral variations when control parameters are modified. It is thus meant to introduce parametric control for sampling sound synthesis. The objective of this thesis is to introduce a new general model repre- senting the timbre variations of quasi-harmonic music instruments regarding a parameter space determined by the control parameters pitch as well as global and instantaneous intensity.
    [Show full text]
  • Uncontrolled Manifolds and Short-Delay Reflexes in Speech Motor Control : a Modeling Study Andrew Szabados
    Uncontrolled manifolds and short-delay reflexes in speech motor control : a modeling study Andrew Szabados To cite this version: Andrew Szabados. Uncontrolled manifolds and short-delay reflexes in speech motor control : a mod- eling study. Neurons and Cognition [q-bio.NC]. Université Grenoble Alpes, 2017. English. NNT : 2017GREAS039. tel-01731536 HAL Id: tel-01731536 https://tel.archives-ouvertes.fr/tel-01731536 Submitted on 14 Mar 2018 HAL is a multi-disciplinary open access L’archive ouverte pluridisciplinaire HAL, est archive for the deposit and dissemination of sci- destinée au dépôt et à la diffusion de documents entific research documents, whether they are pub- scientifiques de niveau recherche, publiés ou non, lished or not. The documents may come from émanant des établissements d’enseignement et de teaching and research institutions in France or recherche français ou étrangers, des laboratoires abroad, or from public or private research centers. publics ou privés. THÈSE Pour obtenir le grade de DOCTEUR DE LA COMMUNAUTÉ UNIVERSITÉ GRENOBLE ALPES Spécialité : PCN - Sciences cognitives, psychologie et neurocognition Arrêté ministériel : 25 mai 2016 Présentée par Andrew SZABADOS Thèse dirigée par Pascal PERRIER (EDISCE), Professeur, G-INP préparée au sein du Laboratoire Grenoble Images Parole Signal Automatique dans l'École Doctorale Ingénierie pour la santé la Cognition et l'Environnement Uncontrolled manifolds et réflexes à courte latence dans le contrôle moteur de la parole : une étude de modélisation Uncontrolled manifolds and
    [Show full text]
  • Parsing the Spectral Envelope
    PARSING THE SPECTRAL ENVELOPE: TOWARD A GENERAL THEORY OF VOCAL TONE COLOR Ian Howell Doctor of Musical Arts Thesis The New England Conservatory of Music Submitted: 18 May 2016 Advisors: Katarina Markovic & Alan Karass DMA Committee Readers: Matthias Truniger & Thomas Novak Howell: Parsing the Spectral Envelope 2 Contents Abstract ..................................................................................................................................................................... 4 Acknowledgements ............................................................................................................................................... 5 Foreword .................................................................................................................................................................. 6 1. How We Draw Vowels: An Introduction to Current Models .............................................................. 8 2. What are Timbre and Tone Color? ........................................................................................................... 20 3. Exploring the Special Psychoacoustics of Sung Vowels .................................................................... 27 Absolute Spectral Tone Color ............................................................................................................................................... 29 The Multiple Missing Fundamentals ................................................................................................................................
    [Show full text]
  • Analysing Deep Learning-Spectral Envelope Prediction Methods For
    Analysing Deep Learning-Spectral Envelope Prediction Methods for Singing Synthesis Frederik Bous Axel Roebel UMR STMS - IRCAM UMR STMS - IRCAM CNRS, Sorbonne University CNRS, Sorbonne University Paris, France Paris, France [email protected] [email protected] Abstract—We conduct an investigation on various hyper- The system of [7] uses recurrent neural networks to model parameters regarding neural networks used to generate spectral the statistic properties needed for their concatenative synthesis. envelopes for singing synthesis. Two perceptive tests, where WaveNet [8] goes further and models the raw audio, rather the first compares two models directly and the other ranks models with a mean opinion score, are performed. With these than concatenating existing audio or using a vocoder. Shortly tests we show that when learning to predict spectral en- after that, end-to-end systems like Tacotron [9] and Deep Voice velopes, 2d-convolutions are superior over previously proposed [10] were developed which create raw audio from input on 1d-convolutions and that predicting multiple frames in an iterated phoneme level or even character level. The authors of [11] fashion during training is superior over injecting noise to the used the architecture of [8] to learn input data for a parametric input data. An experimental investigation whether learning to predict a probability distribution vs. single samples was per- singing synthesizer. formed but turned out to be inconclusive. A network architecture While WaveNet processes data that is inherently one di- is proposed that incorporates the improvements which we found mensional (i.e., raw audio), spectral envelopes are generated to be useful and we show in our experiments that this network in [11].
    [Show full text]
  • A Comparison of Spectral Smoothing Methods for Segment Concatenation Based Speech Synthesis Q David T
    Speech Communication 36 (2002) 343–374 www.elsevier.com/locate/specom A comparison of spectral smoothing methods for segment concatenation based speech synthesis q David T. Chappell b, John H.L. Hansen a,b,* a Robust Speech Processing Laboratory (RSPL), Center for Spoken Language Research (CSLR), Room E265, University of Colorado, 3215 Marine St., P.O. Box 594, Boulder, CO 80309-0594, USA b Department of Electrical Engineering, P.O. Box 90291, Duke University, Durham, NC 27708-0291, USA Received 21 April 1999; received in revised form 24 May 2000; accepted 15 December 2000 Abstract There are many scenarios in both speech synthesis and coding in which adjacent time-frames of speech are spectrally discontinuous. This paper addresses the topic of improving concatenative speech synthesis with a limited database by proposing methods to smooth, adjust, or interpolate the spectral transitions between speech segments. The objective is to produce natural-sounding speech via segment concatenation when formants and other spectral features do not align properly. We consider several methods for adjusting the spectra at the boundaries between waveform segments. Tech- niques examined include optimal coupling, waveform interpolation (WI), linear predictive parameter interpolation, and psychoacoustic closure. Several of these algorithms have been previously developed for either coding or synthesis, while others are enhanced. We also consider the connection between speech science and articulation in determining the type of smoothing appropriate for given phoneme–phoneme transitions. Moreover, this work incorporates the use of a recently- proposed auditory-neural based distance measure (ANBM), which employs a computational model of the auditory system to assess perceived spectral discontinuities.
    [Show full text]
  • Spectral Envelope Coding in Cat Primary Auditory Cortex: Linear and Non-Linear Effects of Stimulus Characteristics
    European Journal of Neuroscience, Vol. 10, pp. 926–940, 1998 © European Neuroscience Association Spectral envelope coding in cat primary auditory cortex: linear and non-linear effects of stimulus characteristics Barbara M. Calhoun and Christoph E. Schreiner UCSF/UCB Bioengineering Graduate Group, the Coleman Lab, and the Keck Center, University of California, San Francisco, CA 94143-0732, USA Keywords: complex stimulus, ripple stimuli, transfer functions Abstract Electrophysiological studies in mammal primary auditory cortex have demonstrated neuronal tuning and cortical spatial organization based upon spectral and temporal qualities of the stimulus including: its frequency, intensity, amplitude modulation and frequency modulation. Although communication and other behaviourally relevant sounds are usually complex, most response characterizations have used tonal stimuli. To better understand the mechanisms necessary to process complex sounds, we investigated neuronal responses to a specific class of broadband stimuli, auditory gratings or ripple stimuli, and compared the responses with single tone responses. Ripple stimuli consisted of 150–200 frequency components with the intensity of each component adjusted such that the envelope of the frequency spectrum is sinusoidal. It has been demonstrated that neurons are tuned to specific characteristics of those ripple stimulus including the intensity, the spacing of the peaks, and the location of the peaks and valleys (C. E. Schreiner and B. M. Calhoun, Auditory Neurosci., 1994; 1: 39–61). Although previous results showed that neuronal response strength varied with the intensity and the fundamental frequency of the stimulus, it is shown here that the relative response to different ripple spacings remains essentially constant with changes in the intensity and the fundamental frequency.
    [Show full text]