A system for voice conversion based on adaptive filtering and line spectral frequency distance optimization for text-to-speech synthesis

This paper proposes a new voice conversion algorithm that modifies the source speaker’s speech to sound as if produced by a target speaker. To date, most approaches for speaker transformation are based on mapping functions or codebooks. We propose a linear filtering based approach to the problem of mapping the spectral parameters of one speaker to those of the other. In the proposed method, the transformation is performed by filtering the source speaker’s Line Spectral Pair (LSP) frequencies to obtain the LSP frequency estimates of the target speaker. Speech signal is timealigned into a sequence of HMM states. The filters are designed for each HMM state using the aligned data. We consider two methods for spectral conversion. A linear transformation for the LSP’s was obtained using the adaptive steepest gradient descent approach. Mean values of LSP’s are adjusted to match those of the target speaker. In order to prevent the LSP vectors from resulting in unstable vocal tract filters, weighted least square estimation is used. This approach optimizes differences between source and target LSP’s. Weights are inverses of the source LSP variances. This approach is integrated into a Time Domain Pitch Synchronous Overlap and Add (TD-PSOLA) analysis-synthesis framework. The algorithm is objectively evaluated using a distance measure based on the log-likelihood ratio of observing the input speech, given Gaussian mixture speaker models for both the source and the target voice. Results using the Gaussian mixture model formulated criteria demonstrate consistent transformation using a 5 speaker database. The algorithm offers promise for rapidly adapting textto-speech systems to new voices.