Acoustic-phonetic modeling for flexible vocabulary speech recognition