Evaluating Sampling-based Filler Insertion with Spontaneous TTS

13th Edition of the Language Resources and Evaluation Conference (LREC 2022), Marseille.

Siyang Wang, Joakim Gustafson and Éva Székely

paper-pdf

Audio samples for Evalaution 1 and 2 in the paper are presented below.


--------------------------------------------------------------------------

Model comparison: TSGD dataset (in-data test sentences)


N-gram (n=3, KN smoothing), TSGD

LSTM-LM, TSGD

Modified GPT-2, TSGD

Ground truth TSGD vocoded

--------------------------------------------------------------------------

Model comparison: trained on TSGD, tested on chat-bot generated responses


N-gram (n=3, KN smoothing), TSGD

LSTM-LM, TSGD

Modified GPT-2, TSGD

No filler, TSGD

--------------------------------------------------------------------------

Model comparison: TCC dataset (in-data test sentences)


N-gram (n=3, KN smoothing), TCC

LSTM-LM, TCC

Modified GPT-2, TCC

Ground truth TCC vocoded

--------------------------------------------------------------------------

Model comparison: trained on TCC, tested on chat-bot generated responses


N-gram (n=3, KN smoothing), TCC

LSTM-LM, TCC

Modified GPT-2, TCC

No filler, TCC

--------------------------------------------------------------------------

No filler vs. filler inserted speech comparison: filler insertion model and TTS trained on TSGD)


no filler, TSGD

filler N-gram as above, TSGD

no filler, LJ

filler N-gram as above, LJ

--------------------------------------------------------------------------

No filler vs. filler inserted speech comparison: filler insertion model and TTS trained on TCC)


no filler, TCC

filler N-gram as above, TCC

no filler, LJ

filler N-gram as above, LJ

--------------------------------------------------------------------------

All rights reserved by authors of the paper "Evaluating Sampling-based Filler Insertion with Spontaneous TTS"(LREC 2022).

All audio samples are created by the authors.

These are for academic research purpose only.

Redistribution or reuse of any material shown on this website or in the paper is prohibited.