Anne Cutler and Dennis Norris propose a model of speech segmentation in stress languages, suggesting that strong syllables trigger segmentation of the speech signal, while weak syllables do not. Experiments show that words embedded in bisyllables with two strong syllables are detected more slowly than those in bisyllables with a strong and a weak syllable. This is attributed to segmentation, requiring assembly of speech material across a segmentation point. Speech recognition models based on phonemic or syllabic recoding or left-to-right processes do not predict this result. Segmentation at strong syllables helps detect efficient lexical access points. Speech recognition involves identifying meanings from acoustic signals. Recognizers store discrete meanings and locate them in memory. Due to the infinite number of potential utterances, recognizers store discrete units, not complete utterances. Speech signals are continuous, and determining word boundaries is challenging without reliable cues. Machine systems often use acoustic templates, while psychological models preprocess signals for prelexical classification. Syllabic classification is more efficient than phonemic, as it reduces unnecessary access attempts. However, in English, syllables are not reliable segmentation units due to variable structures and unclear boundaries. English has strong and weak syllables, with strong containing full vowels and weak containing reduced vowels. A hypothesis suggests speech is classified into feet, with each foot containing one strong syllable and possibly weak syllables. However, experiments show that English listeners do not classify speech into feet. Experiments with bisyllabic nonwords showed that words embedded in strong syllables were detected more slowly, indicating segmentation effects. In Experiment 1, words in bisyllables with two strong syllables were detected more slowly than those with a strong and a weak syllable. Experiment 2, without the second syllable, showed no difference, supporting the segmentation hypothesis. Experiment 3 with CVC words confirmed that segmentation effects were specific to words spanning syllables. The results support the segmentation model, where strong syllables trigger segmentation, delaying word detection when words span syllables. Alternative explanations based on syllable length or intensity were refuted. The findings challenge models of speech recognition that do not account for segmentation. Segmentation at strong syllables is efficient for lexical access, and the results suggest that speech recognition involves segmentation triggered by strong syllables. This process is compatible with phonetic classification models and can be incorporated into broader frameworks for speech recognition.Anne Cutler and Dennis Norris propose a model of speech segmentation in stress languages, suggesting that strong syllables trigger segmentation of the speech signal, while weak syllables do not. Experiments show that words embedded in bisyllables with two strong syllables are detected more slowly than those in bisyllables with a strong and a weak syllable. This is attributed to segmentation, requiring assembly of speech material across a segmentation point. Speech recognition models based on phonemic or syllabic recoding or left-to-right processes do not predict this result. Segmentation at strong syllables helps detect efficient lexical access points. Speech recognition involves identifying meanings from acoustic signals. Recognizers store discrete meanings and locate them in memory. Due to the infinite number of potential utterances, recognizers store discrete units, not complete utterances. Speech signals are continuous, and determining word boundaries is challenging without reliable cues. Machine systems often use acoustic templates, while psychological models preprocess signals for prelexical classification. Syllabic classification is more efficient than phonemic, as it reduces unnecessary access attempts. However, in English, syllables are not reliable segmentation units due to variable structures and unclear boundaries. English has strong and weak syllables, with strong containing full vowels and weak containing reduced vowels. A hypothesis suggests speech is classified into feet, with each foot containing one strong syllable and possibly weak syllables. However, experiments show that English listeners do not classify speech into feet. Experiments with bisyllabic nonwords showed that words embedded in strong syllables were detected more slowly, indicating segmentation effects. In Experiment 1, words in bisyllables with two strong syllables were detected more slowly than those with a strong and a weak syllable. Experiment 2, without the second syllable, showed no difference, supporting the segmentation hypothesis. Experiment 3 with CVC words confirmed that segmentation effects were specific to words spanning syllables. The results support the segmentation model, where strong syllables trigger segmentation, delaying word detection when words span syllables. Alternative explanations based on syllable length or intensity were refuted. The findings challenge models of speech recognition that do not account for segmentation. Segmentation at strong syllables is efficient for lexical access, and the results suggest that speech recognition involves segmentation triggered by strong syllables. This process is compatible with phonetic classification models and can be incorporated into broader frameworks for speech recognition.