A practical spoken term detection for 2000 h of speech data using a pre-matching method for all phone trigrams
Ryo Mikami, Kazunori Kojima, Shi-wook Lee, Yoshiaki Itoh · The Journal of the Acoustical Society of America · 2025
In this paper, we propose a new spoken term detection method that achieves high accuracy, practical retrieval time, and practical memory requirements on 2000 h of speech data. The proposed method performs pre-matching of all phone trigrams against the speech data to construct indexes. When a phone trigram matches well with an utterance, the trigram is assumed to be spoken in that utterance. The utterance ID and its distance are stored in the indexes for each phone trigram. During retrieval, the index is used to approximate the distance between the query and the utterance, and detailed matching is performed for utterances with small distances. By limiting the number of stored utterances for each phone trigram, the proposed method reduces memory requirements compared to conventional methods that use a threshold for storing utterance counts. Evaluation experiments were conducted using approximately 2000 h of speech data from LaboroTVSpeech. The results show that the proposed method reduced memory usage by 14 points and retrieval time by 99 points while maintaining the same retrieval accuracy as conventional methods. While conventional methods involved detailed matching with approximately 110 000 utterances, the proposed method reduces this number to approximately 40. This led to a reduction in retrieval time.