ML - Speech - advanced speech model

Reference resources

Huawei cloud College

RNN

Cyclic neural network （Recurrent Neural Networks,RNN） It is a periodic connection through hidden layer nodes , A neural network to capture dynamic information in serialized data , The serialized data can be classified .
No connection with other networks ,RNN You can save the state of a context , Can store in any long context window 、 Study 、 Express relevant information . And it is no longer limited to the boundary of traditional neural network in space , It can be extended in time series .
RNN It is widely used in scenes related to sequences , Such as a video composed of frames of images , A piece of audio , And a sentence of words .

RNN Network structure

RNN Structural expansion

x Is an input that is being read at the current time , And output a value h;s Is the state of one of the sequences , Which has been processed by the corresponding activation function .

standard RNN

BPTT

RNN Forward propagation of is ： Calculate once in chronological order ,BPTT Is to pass the accumulated residual back from the last , This is similar to ordinary neural network training , The difference is that we add the gradients at each moment .

LSTM

Long and short term memory network （Long Short-Term Memmory,LSTM）, It is a kind of time recurrent neural network , It is suitable for processing and predicting events with long intervals and delays in time series .
LSTM And RNN The difference is this , It adds a method to judge whether the information is useful “ processor ”, The structure of this processor becomes cell.cell Three doors are placed in , They are called input gates , Forgetting gate and output gate . A message enters the network , You can judge whether it is useful according to the rules . Only the information that conforms to the algorithm authentication will be left .

LSTM And speech recognition

DNN Each frame of voice and several frames before and after it are spliced together as the input of the network , So as to use the context information in the speech sequence .DNN The number of frames per input is fixed , Different window lengths will affect the results .
RNN To a certain extent, customer service DNN The shortcomings of , however RNN It's easy to see the gradient disappear , Unable to remember long-term information .
LSTM Through a specific gating unit, the error of the current time can be saved and selectively transmitted to a specific unit , So as to avoid the problem of gradient disappearance . Suitable for long relevant information and location intervals . It is applicable to the need to connect the previous long-term information to the current task

LSTM： The initial state

LSTM： Oblivion gate

LSTM： Input gate

LSTM： to update