Difference between revisions of "SMC/SoC/2008"
→Speech recognition system for Malayalam
| Line 37: | Line 37: | ||
===Speech recognition system for Malayalam=== | ===Speech recognition system for Malayalam=== | ||
The aim is to develop a speech recognition system for Malayalam using the concepts of memory prediction framework. Memory prediction framework put forward by Jeff Hawkins in his book 'On Intelligence'(2004) is a theory of brain function, based on the hierarchical organization of human neocortex.It explains how the hierarchical structure enables brain to match sensory inputs to the stored memory patterns for predicting the future input sequences. According to this model, neocortex | The aim is to develop a speech recognition system for Malayalam using the concepts of memory prediction framework. Memory prediction framework put forward by Jeff Hawkins in his book 'On Intelligence'(2004) is a theory of brain function, based on the hierarchical organization of human neocortex.It explains how the hierarchical structure enables brain to match sensory inputs to the stored memory patterns for predicting the future input sequences. According to this model, neocortex has a layered structure with different layers storing constructs of varying complexity, with sensory inputs coming to the lowest layer. For example in case of vision, the lower layer receives retinal signals and layers up the hierarchy associates themselves with meaningful constructs like lines, two dimensional figures, and furthur up specific objects like faces etc. In speech the layers store different speech constructs from phonemes and syllables to phrases and sentences. The human speech perception and recognition can be understood using this hierarchical organization. | ||
If we mimic the way in which human brain recognizes speech, the resulting system will be more robust than the existing systems. The proposed system is trained with a carefully compiled database and different speech constructs are stored in different layers.When a speech segment to be recognized is given, a series of predictions start and signals will be passed upwards and downwards the layers, until the most probable speech construct is arrived at. For example if the most probable candidate for first word is 'how', predictions start as to what succeeding words can be. This continues until the last word is arrived at and the phrase giving maximum probability will chosen among these predictions. | If we mimic the way in which human brain recognizes speech, the resulting system will be more robust than the existing systems. The proposed system is trained with a carefully compiled database and different speech constructs are stored in different layers.When a speech segment to be recognized is given, a series of predictions start and signals will be passed upwards and downwards the layers, until the most probable speech construct is arrived at. For example if the most probable candidate for first word is 'how', predictions start as to what succeeding words can be. This continues until the last word is arrived at and the phrase giving maximum probability will chosen among these predictions. | ||