Shulou(Shulou.com)06/01 Report--

This article is to share with you the content of a sample analysis of Python speech recognition. The editor thinks it is very practical, so share it with you as a reference and follow the editor to have a look.

Overview

A new deep learning chapter will be opened to talk about the application of deep learning in speech recognition (Speech Recognition). Speech recognition technology can convert speech into computer-readable input, let the computer understand what we want to express, and achieve real human-computer interaction. Hope that through the study of this column, you can have a basic understanding of the field of speech recognition.

RNN

RNN (Recurrent Neural Network) is a cyclic neural network, which is used to deal with tasks with correlated inputs. The RNN network consists of an input layer, a hidden layer, and an output layer, as shown in the figure:

Calculate

The hidden layer (Hidden Layer) defines the state of the entire network. The calculation process of the RNN network is as follows:

Calculation status (State)

Calculated output:

Problems in RNN

Gradient vanishing problem (Vanishing gradient problem). If the derivative is less than 1, as the number of layers of the network increases, the gradient and the new will move in the direction of exponential decay, which is the disappearance of the gradient, as shown in the figure:

We can see that with the increase of time, the perception of the deep network to the shallow layer becomes weaker and weaker, and the gradient is close to 0. 5%.

Gradient explosion problem (Exploding gradient problem). If the derivative is greater than 1, the gradient will increase exponentially with the increase of the number of network layers, which is called gradient explosion. When the derivative of RNN network is greater than 1, timing dependence will occur, resulting in gradient explosion.

LSTM

LSTM (Long Short Term Memory), that is, long-term and short-term memory model. LSTM is a special RNN model, which solves the problem of gradient disappearance and gradient explosion in the process of long sequence training. Compared with ordinary RNN, LSTM can perform better in longer sequences. Compared to RNN with only one transitive state ht, LSTM has two transitive states: ct (cell state) and ht (hidden state).

LSTM adds three control units: input gate, output gate and forgetting gate. The cell of LSTM determines which information is left behind and which is forgotten, thus solving the problem of long sequence dependence in neural networks.

GRU

GRU (Gate Recurrent Unit) is similar to LSTM, but easier to calculate. GRU consists of reset door, update door, and output door. The reset door has the same function as the amnesia text of LSTM, which is used to determine the departure or retention of information. By the same token, the function of the update door is similar to that of the LSTM input gate.

Seq2seq

Seq2seq consists of two RNN, Encoder and Decoder. Encoder outputs the variable length sequence, encodes it into encoderstate, and then outputs the variable length sequence by Decoder.

Attention model

Attention is a mechanism for improving the effectiveness of RNN's Encoder and Decoder models. It is widely used in many fields such as machine translation, speech recognition, image annotation and so on. The attention mechanism in deep learning is essentially similar to the human selective visual attention mechanism. The core goal is to select the information that is more critical to the current task goal from many information.

Attention is essentially a mechanism of content-based addressing. That is, the states with similar given states are selected from some state sets in the network, and then the subsequent information extraction is done.

Firstly, the weights are calculated according to the characteristics of Encoder and Decoder, and then the features of Encoder are weighted and summed as the input of Decoder. Its function is to present the characteristics of Encoder to Decoder in a better way. Not all context have an impact on the generation of the next state, Attention is to choose the appropriate context to generate the next state.

Teacher Forcing mechanism

The prediction ability of the early RNN in the training process is very weak, if a unit prediction is wrong, it is difficult for the later unit to get the right results. For example, we translate a sentence:

Life is like a box of chocolates.You never know what you're going to get

Life is like a box of chocolates, you never know what you're going to get.

If we translate life into "Siberia", the possibility of later translation is almost 0. 5%.

Teacher Forcing is a network training method that uses the previous label as the input to the next state. Or use the above example to illustrate: when using the Teacher Forcing mechanism, even if we translate life into "Siberia", the next Decoder input we will use the previous label as the state, that is, "life" rather than "Siberia". In this way, the prediction ability of RNN network is greatly improved.

Thank you for reading! This is the end of this article on "sample Analysis of Python speech recognition". I hope the above content can be of some help to you, so that you can learn more knowledge. if you think the article is good, you can share it for more people to see!

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.