Speech recognition system based on acoustic model of binary neural network

Abstract

The invention provides a speech recognition system based on the acoustic model of a binary neural network. The observation probability distribution of a hidden markov model is modeled based on the binary neural network, and extracted speech features are trained. In this way, the acoustic model is obtained. Binary data are adopted during the runaway process instead of traditional 32-bit floating-point numbers, so that the storage and memory usage of the model is greatly reduced. By adopting the binary neural network, hardware instructions are fully used during the calculation process for realizing the accelerated operation. In the prior art, a conventional model is provided with a plurality of GPUs only applied to a server for calculation, while the calculation can be conducted on the CPU of mobile equipment. Meanwhile, the model training is conducted based on the acceleration of the binary neural network, so that the model training time can be greatly shortened.

Description

technical field [0001] The invention relates to a technology in the field of information processing, in particular to a speech recognition system based on a binary neural network acoustic model. Background technique [0002] Existing neural networks for acoustic model modeling (including but not limited to DNN, CNN, RNN) have millions or even hundreds of millions of network weights, and each weight is stored as a 32-bit floating point number, Requires a lot of storage and memory to run. For the usual neural network acoustic model, a large number of parameters means a large amount of calculation, which requires high computing power of the device. For the usual neural network acoustic model, it is necessary to use a large amount of speech data to train the model, even with a large amount of computing resources, the training time is very long. [0003] Using 32-bit floating-point numbers as the data type of network weights, existing hardware (such as CPU, Central Processing U...

AI Technical Summary

Problems solved by technology

[0004] The present invention aims at the shortcomings of the prior art that the model training speed is relatively slow, and it is impossible to accelerate bit operations by using CPU or GPU, and proposes a speech recognition system based on a binary neural network acoustic model, which uses a binary value to replace the traditional 32 1-bit floating-point numbers, which greatly reduces the storage and memory usage of the model; the binary neural network used can make full use of hardware instructions to accelerate calculations, and the model that could only be calculated using multiple GPUs on the server before can now be used on mobile run on the CPU of the device; and the present invention benefits from the acceleration of the binary neural network when performing model training, and the model training time can also be greatly shortened

Images