The LF-model revisited. Transformations and frequency domain analysis

The main theme of our presentation is the parameterization of the human voice source. The LF-model provides an eflective representation of the inverse Jiltered speech wave form, but the four parameters are not easily handled in analysis and synthesis of connected speech. A data reduction scheme has recently been proposed (Fant et al., 1994) as a supplement to complete specifications. This system is now outlined in greater detail and is applied to descriptive analysis within a phonetic fiame. Specific emphasis is given the fiequency domain properties of the complete and the transformed systems. An analysis-by-synthesis spectrum matching scheme, which ensures source parameters optimal for speech synthesis, is proposed as an alternative or a complement to traditional inverse$ltering. A crucial requirement is the coherence of source data with the VT transfer finction and the synthesis architecture including higher pole corrections, vocal tract length and the specific samplingfiequency.