Method for conditioning a database for automatic speech processing

A method for training a neural network to assign graphemes to phonemes for automatic speech processing, comprising:

• conditioning a database of words formed of graphemes and associated phonemes comprising:
• assigning the graphemes to the phonemes to form grapheme/phoneme pairs for the words in the database having an identical number of graphemes and phonemes;
• assigning the graphemes to the phonemes for each word in the database having more graphemes than phonemes, to form grapheme/phoneme pairs until an assignment error is detected based on previously determined assignments, or no phonemes remain for the word, then combining a plurality of graphemes in the word to form a grapheme unit and assigning the grapheme unit to one of the phonemes;
• assigning the graphemes to the phonemes for each word in the database having fewer graphemes than phonemes, by assigning a single grapheme in the word to a plurality of phonemes combined to form a phoneme unit, with remaining graphemes and phonemes in the word corresponding to previously assigned grapheme/phoneme pairs;
• reexamining remaining words for previously determined grapheme/phoneme pairs and at least one of phoneme units and grapheme units determined previously;
• and
• erasing, at least after said assigning for the words in the database having an identical number of graphemes and phonemes, any assignment which contradicts another assignment determined to be valid;
• training the neural network by inputting the graphemes at input nodes and the associated phonemes at an output node of the neural network for at least some of the words in the database;
• and
• generating, from the database, individual phonemes and storing the individual phonemes in a phoneme file.

A computer-readable medium having stored thereon instructions for assigning graphemes to phonemes for synthesizing speech, comprising:

• conditioning a database of words formed of graphemes and associated phonemes comprising:
• assigning the graphemes to the phonemes to form grapheme/phoneme pairs for the words in the database having an identical number of graphemes and phonemes;
• assigning the graphemes to the phonemes for each word in the database having more graphemes than phonemes, to form grapheme/phoneme pairs until an assignment error is detected based on previously determined assignments, or no phonemes remain for the word, then combining a plurality of graphemes in the word to form a grapheme unit and assigning the grapheme unit to one of the phonemes;
• assigning the graphemes to the phonemes for each word in the database having fewer graphemes than phonemes, by assigning a single grapheme in the word to a plurality of phonemes combined to form a phoneme unit, with remaining graphemes and phonemes in the word corresponding to previously assigned grapheme/phoneme pairs;
• reexamining remaining words for previously determined grapheme/phoneme pairs and at least one of phoneme units and grapheme units determined previously;
• and
• erasing, at least after said assigning for the words in the database having an identical number of graphemes and phonemes, any assignment which contradicts another assignment determined to be valid;
• and
• training a neural network by inputting the graphemes at input nodes and the associated phonemes at an output node of the neural network for at least some of the words in the database;
• inputting, for each word to be synthesized, an input pattern formed of a letter to be assigned and, if present, at least one letter preceding the letter to be assigned and one subsequent letter in the word;
• and
• generating, from the neural network, an output pattern of one phoneme corresponding to the letter to be assigned.

A computer-readable medium having stored thereon instructions for assigning phonemes to graphemes for recognition of speech, comprising:

• conditioning a database of words formed of graphemes and associated phonemes comprising:
• assigning the graphemes to the phonemes to form grapheme/phoneme pairs for the words in the database having an identical number of graphemes and phonemes;
• assigning the graphemes to the phonemes for each word in the database having more graphemes than phonemes, to form grapheme/phoneme pairs until an assignment error is detected based on previously determined assignments, or no phonemes remain for the word, then combining a plurality of graphemes in the word to form a grapheme unit and assigning the grapheme unit to one of the phonemes;
• assigning the graphemes to the phonemes for each word in the database having fewer graphemes than phonemes, by assigning a single grapheme in the word to a plurality of phonemes combined to form a phoneme unit, with remaining graphemes and phonemes in the word corresponding to previously assigned grapheme/phoneme pairs;
• reexamining remaining words for previously determined grapheme/phoneme pairs and at least one of phoneme units and grapheme units determined previously;
• and
• erasing, at least after said assigning for the words in the database having an identical number of graphemes and phonemes, any assignment which contradicts another assignment determined to be valid;
• and
• training a neural network by inputting the phonemes at input nodes and the associated graphemes at an output node of the neural network for at least some of the words in the database;
• inputting, to the neural network, an input pattern formed of at least a phoneme to be assigned and, if present, a preceding phoneme and a subsequent phoneme;
• and
• generating, from the neural network, an output pattern having one letter corresponding to the phoneme to be assigned.

An apparatus comprising:

• a neural network causing a computer to execute a process comprising:
• conditioning a database of words formed of graphemes and associated phonemes comprising:
• assigning the graphemes to the phonemes to form grapheme/phoneme pairs for the words in the database having an identical number of graphemes and phonemes;
• assigning the graphemes to the phonemes for each word in the database having more graphemes than phonemes, to form grapheme/phoneme pairs until an assignment error is detected based on previously determined assignments, or no phonemes remain for the word, then combining a plurality of graphemes in the word to form a grapheme unit and assigning the grapheme unit to one of the phonemes;
• assigning the graphemes to the phonemes for each word in the database having fewer graphemes than phonemes, by assigning a single grapheme in the word to a plurality of phonemes combined to form a phoneme unit, with remaining graphemes and phonemes in the word corresponding to previously assigned grapheme/phoneme pairs;
• reexamining remaining words for previously determined grapheme/phoneme pairs and at least one of phoneme units and grapheme units determined previously;
• and
• erasing, at least after said assigning for the words in the database having an identical number of graphemes and phonemes, any assignment which contradicts another assignment determined to be valid;
• training the neural network by inputting the graphemes at input nodes and the associated phonemes at an output node of the neural network for at least some of the words in the database;
• and
• generating, from the neural network, an output pattern having one letter corresponding to the phoneme to be assigned.

An electronically readable data medium storing at least one computer program to control a processor to perform a method for conditioning a database of words formed of graphemes and phonemes for automatic speech processing using the database, said method comprising:

• automatically assigning the graphemes to the phonemes to form grapheme/phoneme pairs for the words in the database having an identical number of graphemes and phonemes;
• automatically assigning the graphemes to the phonemes for each word in the database having more graphemes than phonemes, to form grapheme/phoneme pairs until an assignment error is detected based on previously determined assignments, or no phonemes remain for the word, then combining a plurality of graphemes in the word to form a grapheme unit and assigning the grapheme unit to one of the phonemes;
• automatically assigning the graphemes to the phonemes for each word in the database having fewer graphemes than phonemes, by assigning a single grapheme in the word to a plurality of phonemes combined to form a phoneme unit, with remaining graphemes and phonemes in the word corresponding to previously assigned grapheme/phoneme pairs;
• automatically reexamining remaining words for previously determined grapheme/phoneme pairs and at least one of phoneme units and grapheme units determined previously;
• automatically erasing, at least after said assigning for the words in the database having an identical number of graphemes and phonemes, any assignment which contradicts another assignment determined to be valid;
• and
• generating, from the database, individual phonemes and storing the individual phonemes in a phoneme file.