The Voice Conversion Challenge 2016

Toda, Tomoki; Chen, Linghui; Saito, Daisuke; Villavicencio, Fernando; Wester, Mirjam; Wu, Zhizheng; Yamagishi, Junichi

doi:10.21437/interspeech.2016-1066

Cited by 169 publications

(136 citation statements)

References 29 publications

(39 reference statements)

Supporting

Mentioning

133

Contrasting

Unclassified

Order By: Relevance

“…The Voice Conversion Challenge (VCC) 2016 database [27] is a multi-speaker corpus consisting of 5 male and 5 female speakers. In this corpus, each speaker utters the same sentence set.…”

Section: Voice Conversion Challenge 2016 Databasementioning

confidence: 99%

Effects of Training Data Variety in Generating Glottal Pulses from Acoustic Features with DNNs

Airaksinen¹,

Alku²

2017

Interspeech 2017

View full text Add to dashboard Cite

Glottal volume velocity waveform, the acoustical excitation of voiced speech, cannot be acquired through direct measurements in normal production of continuous speech. Glottal inverse filtering (GIF), however, can be used to estimate the glottal flow from recorded speech signals. Unfortunately, the usefulness of GIF algorithms is limited since they are sensitive to noise and call for high-quality recordings. Recently, efforts have been taken to expand the use of GIF by training deep neural networks (DNNs) to learn a statistical mapping between framelevel acoustic features and glottal pulses estimated by GIF. This framework has been successfully utilized in statistical speech synthesis in the form of the GlottDNN vocoder which uses a DNN to generate glottal pulses to be used as the synthesizer's excitation waveform. In this study, we investigate how the DNN-based generation of glottal pulses is affected by training data variety. The evaluation is done using both objective measures as well as subjective listening tests of synthetic speech. The results suggest that the performance of the glottal pulse generation with DNNs is affected particularly by how well the training corpus suits GIF: processing low-pitched male speech and sustained phonations shows better performance than processing high-pitched female voices or continuous speech.

show abstract

“…The Voice Conversion Challenge (VCC) 2016 database [27] is a multi-speaker corpus consisting of 5 male and 5 female speakers. In this corpus, each speaker utters the same sentence set.…”

Section: Voice Conversion Challenge 2016 Databasementioning

confidence: 99%

Effects of Training Data Variety in Generating Glottal Pulses from Acoustic Features with DNNs

Airaksinen¹,

Alku²

2017

Interspeech 2017

View full text Add to dashboard Cite

show abstract

“…L'expérience a consisté dans le jugement par des auditeurs d'échantillons vocaux chantés, basés sur la similarité avec le chanteur cible et le caractère naturel du chanteur, comme utilisé pour la compétition de conversion d'identité vocale de 2016 (Toda et al, 2016). Pour ce faire, quatre chansons ontété sélectionnées pour l'expérience parmi les huit disponibles, et les deux premières phrases de ces quatre chansons ontété utilisées pour la conversion.…”

Section: Configurations Expérimentalesunclassified

“…Grâce aux avancées récentes, la conversion vocale a considérablement gagné en popularité et en qualité au cours des dernières années, menant notamment aux premières compétitions internationales sur la conversion d'identité vocale (Toda et al, 2016;Lorenzo-Trueba et al, 2018), et avec son extensionà la conversion vocale entre des langues différentes (Sündermann et al, 2006;Nakashika et al, 2016;Kinnunen et al, 2017) et la conversion de voix chantée (Villavicencio & Bonada, 2010;Villavicencio & Kenmochi, 2011;Doi et al, 2012;Kobayashi & Toda, 2014). La conversion vocale a un largeéventail d'applications : du divertissement (parler avec la voix d'une autre personne, par exemple via des applications mobiles), créative (reconstruire la voix de personnalités), et médicale (⌧réparation vocale pour les personnes présentant un handicap vocal).…”

Section: Introductionunclassified

Conversion d'Identité de la Voix Chantée par Sélection et Concaténation d'Unités Spectrales

Obin¹,

Pham²,

Röebel³

2018

XXXIIe Journées D’Études Sur La Parole

View full text Add to dashboard Cite

RÉSUMÉCet article présente un algorithme de sélection d'unités spectrales pour la conversion de l'identité de la voix chantéeà partir de bases de données non parallèles. Les algorithmes de conversion basés sur des unités de parole présentent des avantages importants pour la conversion de l'identité vocale : la conversion vocale par sélection d'unités permet la préservation des caractéristiques originales de la voix cible, en utilisant des unités réelles ; et la segmentation en unités linguistiques permet d'apprendre la conversionà partir d'enregistrements de la voix cible non nécessairement alignés avec ceux de la voix source. La contribution principale de cet article est de réaliser la sélection des unités spectrales de la voix cible en fonction de plusieurs facteurs : acoustique, linguistique (phonèmes) et musicaux (hauteur, intensité et durée). Pour ce faire, la sélection de la séquence d'unités d'enveloppe spectrale estétablie comme un problème d'optimisationà partir d'une fonction de coût multiple qui comprend la distorsion spectrale des chanteurs source et cible ainsi que les différences de hauteur, d'intensité et de durée des unités spectrales correspondantes. L'objectif est de guider la sélection vers des enveloppes spectrales du chanteur cible partageant un contexte musical similaire avec celles du chanteur source. Il est montré lors d'une expérience perceptive que l'algorithme proposé améliore le naturel de la conversion et la similarité avec la voix cible. ABSTRACTThis paper presents a unit-selection algorithm for non-parallel singing voice conversion. Unit-based algorithms presents important advantages for voice conversion : the speech segmentation into linguistic units allows the possibility to learn the conversion from on-the-fly databases of the target voice not necessarily aligned to the source voice, and unit-selection voice conversion allows the preservation of the original characteristics of the target voice, by using real units. The main idea of this paper is that the spectral envelopes of a speaker vary according to multiple factors : linguistics (phonemes), and musical (pitch, intensity, and duration). Accordingly, the selection of the sequence of spectral envelope units is established as a multi-target optimization problem, including the spectral distortion of the source and target singers, and the pitch, intensity, and duration differences of the corresponding spectral envelopes. The objective is to guide the selection towards spectral envelopes of the source and target singers sharing a similar musical context. It is shown that the proposed algorithm improves conversion naturalness and target similarity.MOTS-CLÉS : conversion de l'identité vocale, voix chantée, conversion non-parallèle, sélection d'unités, optimisation multi-cible.

show abstract

“…The DIFFVC framework first estimates a sequence of spectral differential from the F0 transformed signal with a trained differential GMM (DIFFGMM), The converted speech is then obtained by directly filtering the F0 transformed speech with the spectral differential. In the Voice Conversion Challenge (VCC) 2016 [26], a predecessor version [25] was evaluated as one of the best systems achieving the best conversion accuracy on speaker identity and high speech quality. It was then adopted as the baseline system of the VCC2018, and achieved the second highest sound quality for the same-gender speaker conversion pairs [27].…”

Section: Introductionmentioning

confidence: 99%

Generalization of Spectrum Differential based Direct Waveform Modification for Voice Conversion

Huang¹,

Wu²,

Kobayashi³

et al. 2019

10th ISCA Workshop on Speech Synthesis (SSW 10)

Self Cite

View full text Add to dashboard Cite

We present a modification to the spectrum differential based direct waveform modification for voice conversion (DIFFVC) so that it can be directly applied as a waveform generation module to voice conversion models. The recently proposed DIFFVC avoids the use of a vocoder, meanwhile preserves rich spectral details hence capable of generating high quality converted voice. To apply the DIFFVC framework, a model that can estimate the spectral differential from the F0 transformed input speech needs to be trained beforehand. This requirement imposes several constraints, including a limitation on the estimation model to parallel training and the need of extra training on each conversion pair, which make DIFFVC inflexible. Based on the above motivations, we propose a new DIFFVC framework based on an F0 transformation in the residual domain. By performing inverse filtering on the input signal followed by synthesis filtering on the F0 transformed residual signal using the converted spectral features directly, the spectral conversion model does not need to be retrained or capable of predicting the spectral differential. We describe several details that need to be taken care of under this modification, and by applying our proposed method to a non-parallel, variational autoencoder (VAE)-based spectral conversion model, we demonstrate that this framework can be generalized to any spectral conversion model, and experimental evaluations show that it can outperform a baseline framework whose waveform generation process is carried out by a vocoder.

show abstract

The Voice Conversion Challenge 2016

Cited by 169 publications

References 29 publications

Effects of Training Data Variety in Generating Glottal Pulses from Acoustic Features with DNNs

Effects of Training Data Variety in Generating Glottal Pulses from Acoustic Features with DNNs

Conversion d'Identité de la Voix Chantée par Sélection et Concaténation d'Unités Spectrales

Generalization of Spectrum Differential based Direct Waveform Modification for Voice Conversion

Contact Info

Product

Resources

About