Bottom-up acoustic-phonetic decoding for the selection of word cohorts from a large vocabulary