Improving phoneme-based spoken document retrieval with phonetic context expansion

In this paper, we investigate the feasibility of a phoneme-based approach of spoken document retrieval. We propose improvements in the detection of keywords by expanding the phonetic context around the requests. The evaluation is done using the French ESTER corpus with 193 country names and it shows that expanding the phonetic contexts improves significantly the precision of the baseline system without affecting the recall. Finally, the improved system can achieve, in noisy transcriptions (a phoneme error rate of 23%), approximately 56.5% recall and 47% precision. These results are obtained with an exact matching search which enables fast access to the information in O(n) for a request of n phonemes.