Editing Large Vocabulary Continuous Speech Recognition (section)

=== Development of End-to-End Models ===
All the time, LVCSR is evolving towards more [[Development of End-to-End Models|End-to-End models]]. Traditional LVCSR typically includes various modules, such as acoustic models, language models, and pronunciation dictionaries. However, in recent decades, to simplify the process and enhance LVCSR efficiency, experts have been focusing on how to use a neural network to integrate all these modules into a unified model.

This approach employs an RNNs model for phoneme recognition, which is the process of converting spoken sounds into text. A distinctive feature of this novel method is its ability to perform this conversion without the need for complex alignment operations, enabling the direct generation of the desired textual output. Additionally, it can operate in real-time scenarios with stringent timing requirements.

The foundation of this method is an extension of a neural machine translation model, which shares similarities with certain existing speech recognition techniques. However, what sets it apart is its capability to calculate scores for all positions within both the input and output sequences, subsequently using these scores to assist in the recognition process. An innovative aspect of this approach is the explicit utilization of these scores for alignment purposes, facilitating the generation of accurate textual representations. Furthermore, the decoder state in this model incorporates information regarding prior alignment choices, enhancing the precision of speech understanding.

A key advantage of this method is its ability to perform decoding in an almost deterministic manner, making it suitable for real-time speech recognition without the need for computationally intensive procedures. This performance characteristic suggests the feasibility of extending the method to large-vocabulary speech recognition systems. It also opens up the possibility of directly searching for the most probable word sequences, as opposed to conducting searches at the phoneme or frame level, as traditionally done in Hidden Markov Model (HMM)-based hybrid systems. This feature significantly contributes to improving the speed and accuracy of speech recognition.<ref>Chorowski, J., Bahdanau, D., Cho, K., & Bengio, Y. (2014). ''End-to-end Continuous Speech Recognition using Attention-based Recurrent NN: First Results'' (arXiv:1412.1602;). arXiv. <nowiki>http://arxiv.org/abs/1412.1602</nowiki></ref>