语音识别系统结构

According to the auditory perception characteristics of human ear, the most important features of speech are extracted, and the speech signal is converted into feature vector.

Knowledge representation of acoustics, phonetics, environmental variables, as well as differences in gender and accent of the speaker.

It uses information such as acoustics, phonetics, environmental characteristics and the speaker’s gender accent.

The variable length of eigenvector sequences and the rich variability of audio signals

The richness and variability of audio signals is determined by the speaker’s various complex characteristics or speaking style and speed.

Caused by environmental noise, channel interference, dialect differences and other factors. Acoustic models need to be robust enough to

Modeling the probability distribution of speech feature vectors in this state

The most obvious property of the distribution of the mixed Gaussian model is its multimode, which makes the mixed Gaussian model can

Describe a lot of physical data that show multimodal properties, such as speech data, while single Gaussian distribution

It doesn’t fit. The multimodal nature of the data may come from a variety of potential factors, each of which determines the score.

A specific mixture of ingredients in cloth. If the factors are identified, the mixed distribution can be decomposed.

Become a set with independent distribution of multiple factors.

First of all, the first question is: how to know what sound each word should pronounce? That’s what it takes.

Dictionaries are generally regarded as modules juxtaposed with acoustic models and language models. You will encounter more than one word in the dictionary.

With the help of dictionaries, the acoustic model knows which sounds should be pronounced in turn for a given text string.

However, in order to calculate the degree of match between speech and phoneme strings, it is also necessary to know the beginning and end of each phoneme.

This is done through the dynamic programming algorithm, and the sound can be found efficiently by using the dynamic rule algorithm.

The demarcation point of the element, so that the product of the matching degree (expressed by probability) between each segment of speech and phoneme is maximum.

The degree of matching also takes into account the probability of conversion between phonemes; the latter is through the hidden horse

A posteriori probability of a speech feature vector sequence to a certain state

Probability graph model, which is used to represent the correlation between sequences

A weighted digraph in which each node is called a state.

There is a certain probability to transmit an observation symbol, and the probability of the jump is expressed by the weight on the edge.

That is, Markov assumes that the symbols emitted in each state are only related to the current state.

Off, which has nothing to do with other states and other symbols, that is, independent output hypothesis

Hidden Markov model can produce two random sequences, one is the state sequence, and the other is the state sequence.

It is a sequence of observation symbols, so it is a double random process, but it can only be observed by the outside world.

Observation symbol sequence, no state sequence can be observed.

The state sequence with the highest probability of occurrence

The transition probability and observation symbol emission probability of each state can be passed through Baum-Weir.

If the level of acoustic units is small, the number of acoustic units will be small, but they will be more sensitive to context.

Because of the phenomenon of co-pronunciation in continuous speech, it is necessary to share the front and back three phonemes.

With the sharp increase in the number of Markov models, the states are generally clustered.

The value of acoustic feature vector in speech recognition task is continuous, in order to eliminate the quantization process.

So consider using the continuous probability density function to pair the eigenvector to the state.

Deng Li and others introduced deep learning into the acoustic modeling of speech recognition.

Modeling the relationship between acoustic feature vector and state greatly improves the accuracy of speech recognition.

After that, the application of deep learning in acoustic modeling of speech recognition began to flourish.

Smaller units than phonemes are used in the actual system, but the principle is the same.

All need to know how to calculate the matching degree between a phoneme and a speech signal.

To do this, you need to find a suitable way to represent the voice signal. It is usually a voice message

The number is divided into many frames, and for each frame, it is converted into a series of operations such as Fourier transform.

From the training data, we can extract a large number of feature vectors and their corresponding sounds.

Using this data, we can train classifiers from features to phonemes.

The distribution of the feature vector of each phoneme, and then in the recognition stage, calculate the feature direction of each frame.

Knowledge representation of a set of character sequences

Indicates the probability of the occurrence of a word sequence

The purpose of the language model in speech recognition is to give the maximum probability according to the output of the acoustic model.

A language model is a probability model built for a language in order to create a model that can describe a given word.

The probability distribution of the occurrence of sequences in language

The language model will tell you that the probability of the first sentence is higher, more like a “human sentence.”

In the process of language modeling, including dictionary, corpus and model selection, it is very important to the performance of the product.

The modeling of language model needs to use complex model formula for simulation calculation.

The performance of the language model largely depends on the quality and volume of the corpus. That matches a specific task

Big corpus is always the most important. However, in practical application, such corpus is often difficult to find.

The model, that is, the probability of the current word, depends on only three historical words. Therefore, more distant historical words are being built.

In the model, it has no effect on the current word probability.

How much space is there. Under such a parameter scale, no matter how big the data is, it seems a little sparse.

The neural network language model technology proposed in recent years has solved the problem of large parameter space and long distance to some extent.

The problem of dependence. And for similar words, the probability estimation comes with a certain degree of smoothing from another.

One point of view solves the problem of sparse data. But the disadvantage of the neural network language model is the training time.

Long, the query speed is slow in practical application, so it needs to be accelerated with hardware.

Language models generally use the chain rule to disassemble the probability of a sentence into the probability product of each word in it.

Under the condition of knowing all the previous words, the probability of the current word

However, when the condition is too long, the probability is difficult to estimate, so the most common practice is to think

The probability distribution of each word depends only on the last few words in history. Such a language model is called

Then according to the Bayesian formula, we can get the outline of what happens under the condition of another word.

As a result, the probability of occurrence of adjacent words is calculated in a large number of corpus, and then a single word is calculated.

Because there are bound to be some obscure phrases that do not appear in the corpus, but they also occur.

Probability, so we need an algorithm to generate the probability of these obscure phrases, that is, smoothing.

Language model, which is the best estimate of training data, the formula is as follows: