Keyword searching technology solves the problem of finding the right word or phrase in the speech flow. Keyword spotting systems make it possible to analyze daily conversations between people and point locations of keywords in the speech flow. They allow extracting necessary information from the large flow of telephone conversations, despite the spontaneous organization of the dialogue, unlimited vocabulary, the number of interlocutors, and noise interference. The search process can run either in real time mode, or in generated sound files.
Keyword searching is closely linked with the speech recognition technology. Until recently it was part of it. In fact, all this is very relative. With good recognition of continuous speech and conversion it into a print there is no need in search technology of keywords: the search will be carried out in the generated text. Nevertheless, because of insufficient quality of modern continuous speech recognition systems, this approach is not an ideal solution.
Approaches to the selection process of keywords:
- Pattern search - several versions of keyword pronouncing by various speakers are recorded for each keyword, on which basis a pattern is created that is used for searching a given word in the speech by means of dynamic programming algorithms. The main drawback of this approach is that creating a pattern for each word is a fairly expensive process.
- Search for phonemic lattice - a large network of possible sounds is constructed for the speech signal with the help of phonemic recognition system, in which the search for possible pronunciations of words or phrases occurs. This technology is similar to text searching in the output of the recognition system, but operates at sounds’ level, rather than whole words.
- Search based on Hidden Markov Models (HMM) - HMM corresponding to keyword’s pronunciation is created for each typed keyword, which is used to search in audio signal. This is one of the most promising techniques at the moment; it provides a sufficiently high percentage of correctly identified keywords.
The problem of keyword searching in speech can be divided into several areas:
- Identification and determination of the location of keyword or phrase (search in a speech audio/sound record, speech databases, real-time search)
- Recognition of commands in continuous speech flow, menu-drive
- Understanding of speech meaning by keywords or phrases searching for dialogue systems
Keyword searching technologies are widely used in the security services, telecommunication companies, radio stations, broadcasting companies, call-centers and other organizations that use a large stream or archive of speech information. They are needed for rapid search in large data sets. Keyword searching systems can be used not only in telephone conversations, but also in video, audio streams, that greatly accelerates the process of data tracking.
In call-centers searching systems are used to monitor the operators’ work. In case of disputable issues or conflict situations supervisor can always connect to the conversation. Also, this technology can be used very effectively to analyze customer needs, demands, and interests in new products and will help to conclude on the perspective of company’s development.
