2019/06/03 by David Álvarez, Santiago Pascual, Álvarez, David +3
Computer Science · #Audio and Speech Processing (eess.AS) #FOS: Computer and information sciences #FOS: Electrical engineering #Music and Audio Processing #Sound (cs.SD) #Speech Recognition and Synthesis #Speech and Audio Processing #electronic engineering #information engineering
paper · pdf · doi:10.48550/arxiv.1906.00733
openalex publication_date 2019/06/03 · openalex created_date 2025/10/10 · openalex updated_date 2026/07/28
Text-to-speech (TTS) acoustic models map linguistic features into an acoustic\nrepresentation out of which an audible waveform is generated. The latest and\nmost natural TTS systems build a direct mapping between linguistic and waveform\ndomains, like SampleRNN. This way, possible signal naturalness losses are\navoided as intermediate acoustic representations are discarded. Another\nimportant dimension of study apart from naturalness is their adaptability to\ngenerate voice from new speakers that were unseen during training. In this\npaper we first propose the use of problem-agnostic speech embeddings in a\nmulti-speaker acoustic model for TTS based on SampleRNN. This way we feed the\nacoustic model with speaker acoustically dependent representations that enrich\nthe waveform generation more than discrete embeddings unrelated to these\nfactors. Our first results suggest that the proposed embeddings lead to better\nquality voices than those obtained with discrete embeddings. Furthermore, as we\ncan use any speech segment as an encoded representation during inference, the\nmodel is capable to generalize to new speaker identities without retraining the\nnetwork. We finally show that, with a small increase of speech duration in the\nembedding extractor, we dramatically reduce the spectral distortion to close\nthe gap towards the target identities.\n