DurIAN replaces end-to-end attention with a duration model in a Tacotron-style decoder, matching Tacotron 2 naturalness (MOS 4.26 vs 4.28) with 0% word skip/repeat errors, and adds a multi-band WaveRNN speedup and style scaling.
Speech parameter generation algorithms for hmm-based speech synthesis,
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
citation-role summary
background 1
citation-polarity summary
fields
cs.CL 1years
2019 1verdicts
CONDITIONAL 1roles
background 1polarities
background 1representative citing papers
citing papers explorer
-
DurIAN: Duration Informed Attention Network For Multimodal Synthesis
DurIAN replaces end-to-end attention with a duration model in a Tacotron-style decoder, matching Tacotron 2 naturalness (MOS 4.26 vs 4.28) with 0% word skip/repeat errors, and adds a multi-band WaveRNN speedup and style scaling.