Pitch Estimation and Voiced–Unvoiced Classification of Speech - Digital Speech: Coding for Low Bit Rate Communication Systems - page 179

Pre-emphasized Energy Ratio
Voiced and unvoiced speech can be discriminated by normalized pre-
emphasized energy.

N

1 |

− − |

s(i) s(i 1 )

i =

Pr = (6.45)

N

1 |

s(i) |

i =

The variance of the difference between adjacent samples is usually much
lower in voiced regions than in unvoiced regions. The first-order correlation
of voiced samples is around 0.85 but that of unvoiced samples is nearly
zero, which is a clear indication of the voiced - unvoiced discriminatory
characteristic of this parameter. A speech waveform and its corresponding
normalized pre-emphasized energy is shown in Figure 6.26.

Figure 6.26 Speech waveform and its normalized pre-emphasized energy with a
possible voicing threshold of 0.9 (shown by the dashed line)