I found these allophone recordings: http://little-scale.blogspot.it/2009/02/sp0256-al2-creative-commons-sample-pack.html whose lengths (excluding the silence at the beginning and at end) are very close to the datasheet values. The guy who made the recordings got the chip to play all of the sounds, and then hechopped them up automatically using thresholds. For each sounds, he got the IC to play silence after each one has finished. Are you sure that the length values in the datasheet are wrong?
[gerald] I can only comment on what I measured myself, but the length in this page are close to the one I measured, and indeed shorter than the datasheet value. The allophone recording I did can be found in this post : http://www.cpcwiki.eu/forum/amstrad-cpc-hardware/amstrad-ssa-1-speech-synthesizer-rom-needed/msg5056/#msg5056 . Recording was done using a logic analyser capturing the digital output of the SPS256, then postprocessed from PWM to analogue (8 bit signed).