Speech recognition as a retrieval problem
This master's thesis deals with ASR systems that are training independent and do not employ any knowledge about language characteristics. These systems utilize a retrieval-based approach, recognizing a spoken word string by finding the most similar word utterance in a reference set for each word utterance contained in the test string. This method is based on the fundamental assumption that each possibly occurring word in a test string is contained in the set of references, as well. The problem is referred to as connected word recognition. If the best matching utterance for each word contained in the test string is known, it can be identified. This approach has the advantage to be mainly independent of the language and is more flexible than many common approaches. A connected word recognizer can be used, for example, to enable human-computer interactions by means of voice or to automatically transcribe a speech database.
Bonn, Univ., Master Thesis, 2013