The contribution of aroma to the flavour of Cheddar cheese has been in
vestigated using 40 samples of diverse origin, including cheese made f
rom raw milk. A panel of 15 trained assessors rated the cheese for 9 a
roma attributes and for 10 flavour descriptors. The attributes chosen
discriminated between samples. The prediction, by linear regression (L
R), principal component regression (PCR) and partial least squares reg
ression (PLS1), of individual flavour ratings from the aroma of the ch
eese was compared. PCR and PLS1 predicted flavour with slightly greate
r certainty than did LR. A further improvement in predictive power was
achieved by post-processing of PCR and PLS1 scores using an artificia
l neural network. Despite the success of this strategy, flavour respon
ses which depend on taste - bitter, salt, sweet and sour - or trigemin
al stimuli (e.g. creamy) were poorly predicted. As a result, caution s
hould be exercised when the prediction of cheese flavour is based only
on assessment of volatile components of the head space above the chee
se. This caveat applies whether the characterisation is made by tradit
ional head-space analysis, by a newly-developed 'electronic nose' or b
y a human assessor.