| 摘要 |
PURPOSE:To synthesize a voiceless sound and a voiced sound at a proper mixing rate at all times and to synthesize a signal close to a natural voice by determining the mixing rate of the voiceless sound and voiced sound by using a neural network. CONSTITUTION:A formant extracting circuit 3 finds the power of each frequency band by fast Fourier analysis from a voice signal segmented by an analytic section segmenting circuit 2 and extracts the feature of its strong/weak pattern, and its output is inputted to a voiced sound synthesizer 5 and a voiceless sound synthesizer 6. The neural network 9 has a hierarchical structure and detects the mixing ratio of the voiced sound and voiceless sound and detects a no-voice parts. The voiced sound mixing ratio B' and voiceless sound mixing ratio A' are applied to multiplying circuits 10 and 11 and the output level (mixing ratio) is adjusted according to the mixing ratios A' and B'. The adjusted output is inputted to an adding circuit 12, which adds the inputted voiced sound and voiceless sound signals to synthesize the voice signal. Consequently, the natural voice where the voiceless sound and voiced sound are smoothly connected can be synthesize with the simple constitution. |
