Abstract

Audio data carries much more information than text and are widely used in many applications like broadcast news. But there is a lack of good techniques for retrieving some information from such data. We are trying to address this problem, called the Spoken Document Retrieval (SDR) for Indian languages, by matching speech-segments without actual recognition. The objective is to index the audio data in suitable format so that we can search in this indexed database for a given query. The output can be some audio file(s) or segment(s). At present, the experiments are done on labeled data and text query. The system is interactive and works based on probabilistic searching. Further work is being done on partially labeled as well as unlabeled speech data and speech query. The techniques being explored involve hypothesizing the word boundaries, developing phoneme-level Hidden Markov Models (HMMs), phoneme-boundary marking using segmental k-means algorithm, deriving syllable-level models, and Vector Quantization. We provide the outline and various issues related to both these scenarios: labeled and transcribed data, and unlabeled and nontranscribed data. The implementation details of the former case is given and the latter case is analyzed. We show that we could achieve reasonably well precision and recall for speech documents, too.

Citations