An Efficient Subband Linear Prediction for LPCNet-Based Neural Synthesis
By Yang Cui, Lei He, Frank K. Soong. 2020
10 citations recorded by the source corpus.
LPCNet neural vocoder and its variants have shown the ability to synthesize high-quality speech in small footprint by exploiting domain knowledge in speech. In this paper, we introduce subband linear prediction in LPCNet for producing high fidelity speech more efficiently with consideration of subband correlation. Speech is decomposed into multiple subband signals with linear prediction to reduce the complexity of neural vocoder. A novel subband-based autoregressive model is proposed to learn the joint distribution of the subband sequences by introducing a reasonable assumption, which keeps the dependence between subbands while accelerating the inference speed. Based upon the human auditory perception sensitivity to the harmonic speech components in the baseband, we allocate more computational resources to model the low-frequency sub-band to synthesize natural phase and magnitude of the synthesized speech. Both objective and subjective tests show the proposed subband LPCNet neural vocoder can synthesize higher quality speech than the original fullband one (MOS 4.62 vs. 4.54), at a rate nearly three times faster.
Subfields: Artificial Intelligence
How researchers see this paper
The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.
In Artificial Intelligence
Citations in the corpus
- A Survey on Neural Speech Synthesis (2021, methodology): “Some following works further improve LPCNet from different perspectives, such as reducing complexity for speedup [364, 269, 148], and improving stability for better quality [126].”