A very low bit rate codec for wide band speech based on a long-term perceptual harmonic plus noise model
Faten Ben Ali, Sonia Djaziri-Larbi · 2016
The long-term (LT) modeling of speech is used to exploit the time correlation between successive speech parameters. A very low speech codec was developed based on this property, by applying the LT modeling on the harmonic plus noise model. This paper presents an additional speech compression by introducing a perceptual frequency masking. Inaudible components will not be considered in the LT modeling, reducing the data size to be coded. The resulting wideband speech codec (p-LT-HNM codec) is based on the perceptual LT harmonic plus noise. An important bit rate reduction is obtained when applying the p-LT-HNM codec to the TIMIT speech database, while preserving the same listening quality.