Speech recognition device and computer program

1. A speech recognition device, comprising:

a processor configured to:

implement an acoustic model based on an End-to-End neural network responsive to an observed sequence formed of prescribed acoustic features obtained from a speech signal, wherein the End-to-End neural network calculates probability of the observed sequence being a certain symbol sequence;

in response to a symbol sequence candidate, calculate a posterior probability of a word sequence corresponding to the symbol sequence candidate;

decode said speech signal based on probabilities calculated by said acoustic model for a plurality of symbol sequences selected based on an input observed sequence, and a posterior probability calculated for each of said plurality of symbol sequences;

in response to a symbol sequence candidate, calculate a posterior probability of a word sequence for the symbol sequence candidate, using an occurrence probability of a word sequence by a word level language model, a posterior probability of said symbol sequence candidate for each word sequence, and an occurrence probability of the symbol sequence candidate by a symbol level language model.

2. The speech recognition device according to claim 1 , wherein said acoustic model is an acoustic model based on CTC.

3. The speech recognition device according to claim 1 , wherein the processor is further configured to calculate a posterior probability of a word sequence for each symbol candidate by dividing a product of said occurrence probability of a word sequence and the posterior probability of a symbol sequence candidate by the occurrence probability of the symbol sequence.

4. A computer program stored on a non-transitory, machine readable medium, when executed causing a computer to function as the processor described in claim 1 .

5. The speech recognition device according to claim 2 , wherein said processor calculates a posterior probability of a word sequence for each symbol candidate by dividing a product of said occurrence probability of a word sequence and the posterior probability of a symbol sequence candidate by the occurrence probability of the symbol sequence.

6. The speech recognition device according to claim 2 , wherein said processor is further configured to:

convert a sub-word sequence s to and output a word sequence W having a sub-word posterior probability P(s|W) added;

convert the word sequence W to and output a word sequence W having an occurrence probability P(W) based on a language model of word sequence level added;

output a sub-word sequence s having an occurrence probability P(s) −β based on a sub-word-level language model added; and

output, as a result of decoding, a hypothesis having the highest score among hypotheses of word sequences received by said WFST, wherein said β is a prescribed scaling factor.

7. The speech recognition device according to claim 1 , wherein said processor is further configured to:

convert a sub-word sequence s to and output a word sequence W having a sub-word posterior probability P(s|W) added;

convert the word sequence W to and output a word sequence W having an occurrence probability P(W) based on a language model of word sequence level added;

output a sub-word sequence s having an occurrence probability P(s) −β based on a sub-word-level language model added; and

output, as a result of decoding, a hypothesis having the highest score among hypotheses of word sequences received by said WFST, wherein said β is a prescribed scaling factor.

8. The speech recognition device according to claim 3 , wherein said processor is additionally configured to:

convert a sub-word sequence s to and output a word sequence W having a sub-word posterior probability P(s|W) added;

convert the word sequence W to and output a word sequence W having an occurrence probability P(W) based on a language model of word sequence level added;

output a sub-word sequence s having an occurrence probability P(s) −β based on a sub-word-level language model added; and

output, as a result of decoding, a hypothesis having the highest score among hypotheses of word sequences received by said WFST, wherein said β is a prescribed scaling factor.

9. A speech recognition device, comprising:

a processor configured to:

implement an acoustic model based on an End-to-End neural network responsive to an observed sequence formed of prescribed acoustic features obtained from a speech signal, wherein the End-to-End neural network calculates probability of the observed sequence being a certain symbol sequence;

in response to a symbol sequence candidate, calculate a posterior probability of a word sequence corresponding to the symbol sequence candidate;

decode said speech signal based on probabilities calculated by said acoustic model for a plurality of symbol sequences selected based on an input observed sequence, and a posterior probability calculated for each of said plurality of symbol sequences;

convert a sub-word sequence s to and output a word sequence W having a sub-word posterior probability P(s|W) added;

convert the word sequence W to and output a word sequence W having an occurrence probability P(W) based on a language model of word sequence level added;

output a sub-word sequence s having an occurrence probability P(s) −β based on a sub-word-level language model added; and

output, as a result of decoding, a hypothesis having the highest score among hypotheses of word sequences received by said processor, wherein said β is a prescribed scaling factor.