Recent Advances in Continuous Speech Recognition Using the Time-Sliced Paradigm

We developed a method called Time-Slicing [1] for the analysis of the speech signal. It enables a neural network to recognize connected speech as it comes, without having to fit the input signal into a fixed time-format, nor label or segment it phoneme by phoneme. The neural network produces an immediate hypothesis of the recognized phoneme and its size is small enough to run even on a PC. To implement the time-slicing paradigm two different neural network architectures were tried: the TimeSliced Recurrent Recognizer (TSRR)[2] using an Elman style recurrent network, and the Time-Sliced Recurrent Cascade Correlation Network (TS-RCCN)[3] using the Recurrent Cascade-Correlation Architecture (RCC) [4]. The RCC is a powerful feature-learner, and manipulating its original structure some improvements were gained in its ability to generalize and recognize phonemes. This new structure was called the Parallel-RCC [5]. This paper presents the latest recognition results obtained with the Parallel-RCC and describes some of the attempts to further analyze the networks output with an errorrecovery method to obtain the final result. Recent Advances in Continuous Speech Recognition Using the Time-Sliced Paradigm Ingrid Kirschning, Hideto Tomabechi and Jun-Ichi Aoe Dept. of Information Science and Intelligent Systems Faculty of Engineering, University of Tokushima, 2-1 Minami Josanjima-Cho, Tokushima Shi, 770 Japan. E-mail : [email protected] Abstract : We developed a method called Time-Slicing [1] for the analysis of the speech signal. It enables a neural network to recognize connected speech as it comes, without having to fit the input signal into a fixed time-format, nor label or segment it phoneme by phoneme. The neural network produces an immediate hypothesis of the recognized phoneme and its size is small enough to run even on a PC. To implement the time-slicing paradigm two different neural network architectures were tried: the Time-Sliced Recurrent Recognizer (TSRR)[2] using an Elman style recurrent network, and the Time-Sliced Recurrent Cascade Correlation Network (TSRCCN)[3] using the Recurrent Cascade-Correlation Architecture (RCC) [4]. The RCC is a powerful feature-learner, and manipulating its original structure some improvements were gained in its ability to generalize and recognize phonemes. This new structure was called the Parallel-RCC [5]. This paper presents the latest recognition results obtained with the Parallel-RCC and describes some of the attempts to further analyze the networks output with an error-recovery method to obtain the final result. We developed a method called Time-Slicing [1] for the analysis of the speech signal. It enables a neural network to recognize connected speech as it comes, without having to fit the input signal into a fixed time-format, nor label or segment it phoneme by phoneme. The neural network produces an immediate hypothesis of the recognized phoneme and its size is small enough to run even on a PC. To implement the time-slicing paradigm two different neural network architectures were tried: the Time-Sliced Recurrent Recognizer (TSRR)[2] using an Elman style recurrent network, and the Time-Sliced Recurrent Cascade Correlation Network (TSRCCN)[3] using the Recurrent Cascade-Correlation Architecture (RCC) [4]. The RCC is a powerful feature-learner, and manipulating its original structure some improvements were gained in its ability to generalize and recognize phonemes. This new structure was called the Parallel-RCC [5]. This paper presents the latest recognition results obtained with the Parallel-RCC and describes some of the attempts to further analyze the networks output with an error-recovery method to obtain the final result.