End-to-end speech recognition model training method and system

Abstract

The invention discloses a training method for an end-to-end speech recognition model. The end-to-end speech recognition model includes an encoder and a decoder. The method includes: obtaining an acoustic model and a cross-entropy language model through pre-training; Model initialization of the encoder of the end-to-end speech recognition model; initialization of the decoder of the end-to-end speech recognition model according to the cross-entropy language model; training of the end-to-end speech recognition model after initialization. The multi-stage pre-training method avoids the long and slow learning phase in the early stage of model training, thereby greatly reducing the model training time. At the same time, this strategy has no hyperparameters to tune. Compared with the existing technology, it avoids a lot of tedious hyperparameter tuning.

Description

technical field [0001] The present invention relates to the technical field of speech recognition, in particular to a training method and system for an end-to-end speech recognition model. Background technique [0002] In order to train a good end-to-end system in the prior art, there are preheating and super-long learning rate scheduling strategies, but they are all equivalent to slowly increasing the learning intensity of the model. Continue training in the normal way. [0003] Essentially, warm-up and super-long learning rate scheduling are all about giving the model a good initial state. However, the disadvantage of these techniques is that a large number of hyperparameter tuning is required for a specific model, which is not universal, and requires a long time of training, which is very time-consuming. If these similar technologies want to achieve good results, users need to have rich experience in parameter adjustment, high professional requirements, and high labor c...

AI Technical Summary

Problems solved by technology

However, the disadvantage of these technologies is that a large number of hyperparameters need to be tuned for a specific model, which is not universal, and requires a long time of training, which is very time-consuming.

If these similar technologies want to achieve good results, users need to have rich experience in parameter adjustment, high professional requirements, and high labor costs.

Images