An acoustic model training method for high-precision continuous speech recognition

Abstract

The invention belongs to the technical field of speech recognition, and discloses an acoustic model training method for high-precision continuous speech recognition, comprising the following steps: Step 1. Prepare training corpus and extract speech features; Step 2: Calculate the acoustic model; Step 3. Acoustic model Initialization, step 4. Iterate the initialization acoustic model in step 4 according to the number of training iterations determined in step 4. Step 5. After the training, select the acoustic model with the highest decoding accuracy before merging, average and merge the acoustic model parameters for the final acoustic model. The invention optimizes the voice model recognition sequence expression, and adjusts the model parameters while determining the correct marked text through pre-decoding, so as to reduce the overall difference between the easily confused correct result and the wrong result as much as possible, and improves the decoding accuracy of the acoustic model parameters.

Description

technical field [0001] The invention belongs to the technical field of speech recognition, in particular to an acoustic model training method for high-precision continuous speech recognition. Background technique [0002] The traditional acoustic modeling method is based on the Hidden Markov framework, and the Gaussian mixture model (GMM) is used to describe the probability distribution of speech acoustic features. Since the hidden Markov model belongs to a typical shallow learning structure, it is a kind of The simple structure of transforming the original input signal into the feature space is limited in its performance under massive data. In the later period, the academic community used the neural network and the hidden Markov model together, that is, the hybrid model models the output probability distribution to improve the overall effect, but the improvement is relatively still very limited. [0003] Continuous speech recognition is similar to natural speech. It is a p...

AI Technical Summary

Problems solved by technology

[0004] In order to solve the problem of poor long sentence recognition results, the usual practice in the industry is to use a large amount of speech corpus for training, which can improve the overall performance, but the training method of neural network and hidden Markov model and the final performance of the decoding model There is a bottleneck, and the root cause of the bottleneck in the final performance is that the gradient descent algorithm of deep learning represented by the neural network cannot infinitely fit all the characteristics of the training data, and the decoding model is only a mathematical modeling of the limited space of training data with a certain sample size , does not represent the possibility of unlimited data in reality, and it is difficult to improve the continuous speech recognition effect in the case of limited data training and decoding models

Images