SFNet: A Computationally Efficient Source Filter Model Based Neural Speech Synthesis

Recently, neural speech synthesizers have achieved a high-quality synthesis for text-to-speech applications, but a real-time synthesis is possible only in the devices which have high memory and allow large computational complexity. In this work, we reduce the complexity of a speech synthesizer by reformulating the source-filter model of speech where the excitation signal is modeled as a sum of two signals. The first signal contains an impulse train that is computed from the pitch sequence. The second signal is modeled as white noise passed through a filter bank with frequency dependent gains. The parameters of the reformulated source-filter model are predicted using a neural network, referred to as SFNet. The network parameters are learnt by training the network using l(1)-error between the log Mel-spectrum of the predicted waveform and that of the ground-truth waveform. We demonstrate that there is a significant reduction in the memory and computational complexity compared to the state-of-the-art speaker independent neural speech synthesizer without any loss of the naturalness of the synthesized speech.