This paper presents a fast word spotting approach based on phoneme acoustic models. This approach simplifies the decoding algorithm of speech recognition process using a generic model and a model of the word to be detected. The ratio between the word model likelihood and the generic model likelihood is given to a classifier which was trained to detect relevant peaks values that indicate word occurrences. Experiments were performed on continuous speech utterances of the Portuguese TECNOVOZ database obtaining an accuracy of 94.8% against 1.0% false alarm. This result prompted the application of this technique in thousands of words or multi-word expressions to TV broadcast news programs which is expected to enhance our acoustic models.