Speech recognition modeling method and speech recognition modeling device

Abstract

The invention provides a speech recognition modeling method and a speech recognition modeling device. The above speech recognition modeling method comprises steps: speech signals are converted into a feature vector sequence and an annotated text corresponding to the speech signals is converted into a modeling unit sequence, wherein each modeling unit in the modeling unit sequence is a complete initial consonant or vowel producing unit or an initial consonant or vowel producing unit related to a context; a blank label is added in front of or behind any modeling unit in the modeling unit sequence; and based on connectionist temporal classification (CTC), the feature vector sequence and the modeling unit sequence after the blank label is added are trained, and a speech recognition model is built. According to the CTC-based deep neural network initial consonant and vowel modeling of the invention, the recognition speed and the recognition accuracy for the built speech recognition model can be improved.

Description

technical field [0001] The invention relates to the technical field of speech recognition, in particular to a modeling method and device for speech recognition. Background technique [0002] Existing large vocabulary Chinese speech recognition methods are mainly based on hybrid methods, such as: Gaussian Mixture Model (GaussianMixtureModel; hereinafter referred to as: GMM) + Hidden Markov Model (HiddenMarkovModel; hereinafter referred to as: HMM), deep neural network (DeepNeuralNetwork; hereinafter Abbreviation: DNN)+HMM, etc. Specifically, in speech recognition based on statistical mixed method modeling, methods for estimating the state posterior probability of hidden Markov models include: Gaussian mixture model, deep neural network (specifically, deep multi-layer perceptron) , deep convolutional neural network and deep recurrent neural network, etc., as well as a combination model of several. [0003] Speech signals are typical time-series signals with short-term statio...

AI Technical Summary

Problems solved by technology

[0003] Speech signals are typical time-series signals with short-term stationary characteristics, but due to factors such as background noise, channel, speaker (ie gender, age, speech rate and / or accent, etc.), the speech signal is very complex

[0004] However, existing speech recognition methods are all based on hybrid methods, and the modeling units in speech recognition are all based on traditional state modeling units. State modeling greatly restricts all existing modeling methods. Using the above The speech recognition model established by the state modeling unit has poor recognition performance for speech recognition

Images