A grammar generation process generates a speech recognition grammar for interpreting search queries or a domain of items. The grammar comprises both single-term and multi-term utterances derived from the texts of the items (preferably the item titles). The utterances are derived in-part by expanding phrases selected from the item text into their individual terms plus all forward combinations of such terms. The forward combinations and individual terms that are deemed not useful to the search process are filtered out of the grammar. The process tends to produce a grammar containing the utterances that are most likely to occur within voice queries for the items, while maintaining a grammar size that is sufficiently small to provide reliable speech recognition.