Systems and methods for training statistical speech translation systems from speech utilizing a universal speech recognizer

A method comprising:

• recognizing a spoken language in a source language from a first audible source;
• receiving language in a target language spoken by a human translator, the received language being a translation of the spoken language from the first audible source by the human translator;
• recognizing, by a universal speech recognition component adapted to recognize a plurality of spoken languages from the human translator, the received language in the target language spoken by the human translator translating simultaneous with the spoken language from the first audible source;
• creating a source language hypothesis of the spoken language from the first audible source using a first language model;
• creating a target language hypothesis of the language spoken by the human translator using a second language model;
• performing an iterative translation cycle, the cycle comprising:
• translating, using a first machine translation model, the source language hypothesis into a target language result;
• translating, using a second machine translation model, the target language hypothesis into a source language result;
• training, by a processor, the first language model using the source language result;
• training the universal speech recognition component using the translation from the source language hypothesis to the target language result;
• and
• training the first machine translation model and the second machine translation model using the source language hypothesis and the target language hypothesis.

A system comprising:

• first recognition means for:
• recognizing a spoken language in a source language from a first audible source;
• and
• creating a source language hypothesis of the spoken language from the first audible source using a first language model;
• second recognition means for:
• receiving language in a target language spoken by a human translator, the received language being a translation of the spoken language from the first audible source by the human translator;
• recognizing the received language in the target language spoken by the human translator translating simultaneous with the spoken language from the first audible source, wherein the second recognition means is a universal speech recognition component adapted to recognize a plurality of spoken languages from the human translator;
• and
• creating a target language hypothesis of the language spoken by the human translator using a second language model;
• first translation means for translating the source language hypothesis into a target language result;
• second translation means for translating the target language hypothesis into a source language result;
• means for training the first language model using the source language result;
• means for training the universal speech recognition component using the translation from the source language hypothesis to the target language result;
• means for training the first translation means and the second translation means using the source language hypothesis
• the target language hypothesis from;
• and
• wherein the means for training the language models and translation means collectively perform an iterative translation cycle.

A non-transitory computer readable medium having stored thereon instructions which, when executed by a processor, cause the processor to:

• recognize, a spoken language in a source language from a first audible source;
• receive, at a universal speech recognition component adapted to recognize a plurality of spoken languages from the human translator, language in a target language spoken by a human translator, the received language being a translation of the spoken language from the first audible source by the human translator;
• recognize the received language in the target language spoken by the human translator translating simultaneous with the spoken language from the first audible source;
• create a source language hypothesis of the spoken language from the first audible source using a first language model;
• create, a target language hypothesis of the language spoken by the human translator using a second language model;
• perform an iterative translation cycle, the cycle causing the processor to:
• translate, using a first machine translation model, the source language hypothesis into a target language result;
• translate, using a second machine translation model, the target language hypothesis into a source language result;
• train the first language model using the source language result;
• train the universal speech recognition component using the translation from the source language hypothesis to the target language result;
• and
• train the first machine translation model and the second machine translation model using the source language hypothesis and the target language hypothesis.