New techniques for open-vocabulary spoken document retrieval

Abstract

This paper presents four novel techniques for open-vocabulary spoken document retrieval: a method to detect slots that possibly contain a query feature; a method to estimate occurrence probabilities; a technique that we call collection-wide probability re-estimation and a w eighting scheme which takes advantage of the fact that long query features are detected more reliably. These four techniques have been evaluated using the TREC-6 spoken document retrieval test collection to determine the improvements in retrieval e ectiveness with respect to a baseline retrieval method. Results show that the retrieval e ectiveness can be improved considerably despite the large number of speech recognition errors.

Experiments in spoken document retrieval using phoneme n-grams

“…For this collection, another 2580 words were added to the pronunciation dictionary to transform all words, including query words, to phoneme sequences. In practice, this is done via an automated process [23]. Here, four versions of the transcriptions were used for retrieval:…”

“…The rationale for this is that the incorrect transcription may be able to match relevant documents, which may also contain the incorrect transcription. This method is similar in concept to using a confusion matrix based approach by Wechsler [23] on the training collection, which can be used to determine which recognised phoneme is most likely to be recognised incorrectly as another. This technique, though not 100% accurate, is the only feasible approach for a larger collection.…”

“…Previous experiments by Wechsler [23,24], Ng and Zue [12,13,14] and Smeaton et al [19] on n-gram retrieval from phoneme transcriptions obtained directly from a phoneme recogniser showed that phoneme n-gram retrieval can be effective in practice. Other experiments have shown that phoneme retrieval can be used to complement word retrieval [25] when word recognition has failed, especially in situations where names and unknown words are mis-recognised.…”

Towards Learning Word Representation

“…Another work [24] suggests to guide word-embeddings with morphologically annotated data and shows achievement using German in a case study. Also, many papers study syllable-based n-gram methods to model language [25,26].…”

Speech-Driven Text Retrieval: Using Target IR Collections for Statistical Language Model Adaptation in Speech Recognition

“…-spoken document retrieval, in which written queries are used to search speech (e.g., broadcast news audio) archives for relevant speech information [5,6,15,16,17,19,20], -speech-driven (spoken query) retrieval, in which spoken queries are used to retrieve relevant textual information [2,3].…”