Speech recognition method and device

A speech recognition device, comprising:

• one or more processors;
• a memory having instructions stored thereon that, when executed by the one or more processors, is configured to implement a speech recognition method comprising:
• receiving a speech signal;
• decoding the speech signal according to an acoustic model, a language model and a decoding network established in advance, and adding a blank unit dynamically during a decoding process to obtain an optimum decoding path with the added blank unit, wherein the acoustic model is obtained by training based on connectionist temporal classification, the acoustic model comprises basic pronunciation units and the blank unit, and the decoding network comprises a plurality of decoding paths consisting of the basic pronunciation units;
• and
• outputting the optimum decoding path as a recognition result of the speech signal;
• wherein decoding the speech signal according to an acoustic model, a language model and a decoding network established in advance, and adding a blank unit dynamically during a decoding process to obtain an optimum decoding path with the added blank unit comprises:
• extending each of current decoding paths according to a jumping path in the decoding network;
• adding the blank unit dynamically during an extending process to obtain at least one extending path with the added blank unit;
• obtaining a score of each of the at least one extending path on the acoustic model and the language model according to a current feature vector frame extracted from the speech signal;
• screening the at least one extending path according to the scores, and updating the current decoding path according to a screening result;
• and
• when the current feature vector frame is a final feature vector frame of the speech signal, selecting the optimum decoding path from the updated current decoding paths according to the scores.

A non-transitory computer-readable storage medium having stored therein one or more programs that, when executed by a device, causes the device to perform a speech recognition method, the speech recognition method comprising:

• receiving a speech signal;
• decoding the speech signal according to an acoustic model, a language model and a decoding network established in advance, and adding a blank unit dynamically during a decoding process to obtain an optimum decoding path with the added blank unit, wherein the acoustic model is obtained by training based on connectionist temporal classification, the acoustic model comprises basic pronunciation units and the blank unit, and the decoding network comprises a plurality of decoding paths consisting of the basic pronunciation units;
• and
• outputting the optimum decoding path as a recognition result of the speech signal;
• wherein decoding the speech signal according to an acoustic model, a language model and a decoding network established in advance and adding a blank unit dynamically during a decoding process to obtain an optimum decoding path with the added blank unit comprises:
• extending each of current decoding paths according to a jumping path in the decoding network, and adding the blank unit dynamically during an extending process to obtain at least one extending path with the added blank unit;
• obtaining a score of each of the at least one extending path on the acoustic model and the language model according to a current feature vector frame extracted from the speech signal;
• screening the at least one extending path according to the scores, and updating current decoding paths according to a screening result;
• and
• when the current feature vector frame is a final feature vector frame of the speech signal, selecting the optimum decoding path from the updated current decoding paths according to the scores.