This study investigates the relationship between the production and perception of word-initial gemination in stops and fricatives in Tashlhiyt Berber. Gemination in this language is primarily implemented through longer duration, even for utterance-initial voiceless stops. This timing information is sufficient for native listeners to identify geminate fricatives and voiced stops and distinguish them from their singleton counterparts. For voiceless stops, however, native listeners’ discrimination performance is only slightly above chance level. Native speakers can thus encode a phonemic contrast at the articulatory level and yet be unable to fully decode it at the perceptual level. Implications of these results for the general issue of phonological representation of gemination are briefly discussed.