One sentence voice adaptation using GMM-based frequency-warping and shift with a sub-band basis spectrum model

Abstract

This paper presents a rapid voice adaptation algorithm using GMM-based frequency warping and shift with parameters of a subband basis spectrum model (SBM)[1]. The SBM parameter represents a shape of a spectrum of speech. It is calculated by fitting a sub-band basis to the log-spectrum. Since the parameter is the frequency domain representation, frequency warping can be directly applied to the SBM parameter. A frequency warping function that minimize the distance between source and target SBM parameter pairs in each mixture component of a GMM is derived using a DP (Dynamic programming) algorithm. The proposed method is evaluated in an unit-selection based voice adaptation framework applied to a unit-fusion based text-to-speech synthesizer. The experimental results show that the proposed adaptation method is effective for rapid voice adaptation using just one sentence, compared to the conventional GMM.-based linear transformation of mel-cepstra.Index Terms-voice adaptation, frequency warping, subband basis spectrum model, unit fusion speech synthesis

Voice adaptation by color-encoded frame matching as a multi-objective optimization problem for future games

“…Sekii et al [26] sees good MOS results for their DNN approach, but execution time is not ideal for real-time use as it requires time to extract, convert and restore features. Kotani et al [27] while using larger data and time spent training sees typical VA performance and [28] with a route of frequency warping by considering many warping factors achieves decent results compared to the then stateof-the-art statistical methods. Many speaker data sets, layers in NN and long training times across various methods all fall towards the average mark, while [29] (DNN) see mixed results regarding speaker likeness in relation to training volume.…”

ML Parameter Generation with a Reformulated MGE Training Criterion — Participation in the Voice Conversion Challenge 2016

“…Such functions are normally trained from constant-dimension acoustic feature vectors provided by a vocoder. In many other solutions, the conversion function can be applied only to a specific speech signal representation: modification of formant frequencies and bandwidths [17,18], frequency warping (FW) [19,20,21], FW followed by amplitude scaling (AS) [22,23], etc. When footprint is not an issue, the system can keep some training data from the target speaker and then perform VC through frame selection [24], feature trajectory selection [25], or exemplar-driven transforma- tions [26].…”