Applying neural network language models to weighted finite state transducers for automatic speech recognition

Patent metadata

Field	Value
Publication number	US-10049668-B2
Application number	US-201615156161-A
Country	US
Kind code	B2
Filing date	May 16, 2016
Priority date	Dec 2, 2015
Publication date	Aug 14, 2018
Grant date	Aug 14, 2018

Abstract

Systems and processes for converting speech-to-text are provided. In one example process, speech input can be received. A sequence of states and arcs of a weighted finite state transducer (WFST) can be traversed. A negating finite state transducer (FST) can be traversed. A virtual FST can be composed using a neural network language model and based on the sequence of states and arcs of the WFST. The one or more virtual states of the virtual FST can be traversed to determine a probability of a candidate word given one or more history candidate words. Text corresponding to the speech input can be determined based on the probability of the candidate word given the one or more history candidate words. An output can be provided based on the text corresponding to the speech input.

First claim

What is claimed is:

1. A non-transitory computer-readable medium having instructions stored thereon, the instructions, when executed by one or more processors, cause the one or more processors to:
receive speech input;
traverse, based on the speech input, a sequence of states and arcs of a weighted finite state transducer (WFST), wherein:
the sequence of states and arcs represents one or more history candidate words and a current candidate word; and
a first probability of the candidate word given the one or more history candidate words is determined by traversing the sequence of states and arcs of the WFST;

traverse a negating finite state transducer (FST), wherein traversing the negating FST negates the first probability of the candidate word given the one or more history candidate words;
compose a virtual FST using a neural network language model and based on the sequence of states and arcs of the WFST, wherein one or more virtual states of the virtual FST represent the current candidate word;
traverse the one or more virtual states of the virtual FST, wherein a second probability of the candidate word given the one or more history candidate words is determined by traversing the one or more virtual states of the virtual FST;
determine, based on the second probability of the candidate word given the one or more history candidate words, text corresponding to the speech input;
based on the determined text, perform one or more tasks to obtain a result; and
cause the result to be presented in spoken or visual form.

2. The non-transitory computer-readable medium of claim 1 , wherein the virtual FST is composed after traversing the sequence of states and arcs of the WFST.

3. The non-transitory computer-readable medium of claim 1 , wherein only one are transitions out of each virtual state of the one or more virtual states of the virtual FST.

4. The non-transitory computer-readable medium of claim 1 , wherein the instructions further cause the one or more processors to:
determine, using the neural network language model, a third probability of the candidate word given the one or more history candidate words, wherein the virtual FST is composed using the third probability of the candidate word given the one or more history candidate words.

5. The non-transitory computer-readable medium of claim 1 , wherein the instructions further cause the one or more processors to:
compose a second virtual FST using a second language model and based on the sequence of states and arcs, wherein one or more virtual states of the second virtual FST represents the current candidate word; and
traverse the one or more virtual states of the second virtual FST, wherein a fourth probability of the candidate word given the one or more history candidate words is determined by traversing the one or more virtual states of the second virtual FST, and wherein the text corresponding to the speech input is determined based on the fourth probability of the candidate word given the one or more history candidate words.

6. The non-transitory computer-readable medium of claim 5 , wherein the instructions further cause the one or more processors to:
interpolate the second probability of the candidate word given the one or more history candidate words and the fourth probability of the candidate word given the one or more history candidate words, and wherein:
a combined probability of the candidate word given the one or more history candidate words is determined by the interpolating; and
the text corresponding to the speech input is determined based on the combined probability of the candidate word given the one or more history candidate words.

7. The non-transitory computer-readable medium of claim 5 , wherein the second language model is an n-gram language model.

8. The non-transitory computer-readable medium of claim 1 , wherein the instructions further cause the one or more processors to:
compose the negating FST with the WFST prior to traversing the negating FST.

9. The non-transitory computer-readable medium of claim 8 , wherein the virtual FST is composed prior to traversing the one or more virtual states of the virtual FST.

10. The non-transitory computer-readable medium of claim 1 , wherein the WFST is a static finite state transducer built prior to receiving the speech input.

11. The non-transitory computer-readable medium of claim 1 , wherein the negating FST is a static finite state transducer built prior to receiving the speech input.

12. The non-transitory computer-readable medium of claim 1 , wherein the WFST is a single finite state transducer composed from a Hidden Markov Model (HMM) topology, a context dependent phonetic model, a lexicon, and a language model.

13. The non-transitory computer-readable medium of claim 12 , wherein the language model is a unigram language model or a bigram language model.

14. A non-transitory computer-readable medium having instructions stored thereon, the instructions, when executed by one or more processors, cause the one or more processors to:
receive speech input;
traverse, based on the speech input, a sequence of states and arcs of a weighted finite state transducer (WFST), wherein:
the sequence of states and arcs represents one or more history candidate words and a non-terminal class; and
a first probability of the non-terminal class given the one or more history candidate words is determined by traversing the sequence of states and arcs of the WFST;

traverse a negating finite state transducer (FST), wherein traversing the negating FST negates the first probability of the non-terminal class given the one or more history candidate words;
compose a virtual FST using a neural network language model and a user-specific language model FST, and based on the sequence of states and arcs of the WFST, wherein one or more virtual states of the virtual FST represent a current candidate word corresponding to the non-terminal class;
traverse the one or more virtual states of the virtual FST, wherein a probability of the current candidate word given the one or more history candidate words and the non-terminal class is determined by traversing the one or more virtual states of the virtual FST;
determine, based on the probability of the current candidate word given the one or more history candidate words and the non-terminal class, text corresponding to the speech input;
based on the determined text, perform one or more tasks to obtain a result; and
cause the result to be presented in spoken or visual form.

15. The non-transitory computer-readable medium of claim 14 , wherein the instructions further cause the one or more processors to:
determine, using the neural network language model, a second probability of the non-terminal class given the one or more history candidate words, wherein the virtual FST is composed using the second probability of the non-terminal class given the one or more history candidate words.

16. The non-transitory computer-readable medium of claim 14 , wherein the instructions further cause the one or more processors to:
traverse the user-specific language model FST, wherein a probability of the current candidate word among a plurality of candidate words represented in the user-specific language model FST is determined by traversing the user-specific language model FST, and wherein the virtual FST is composed using the probability of the current candidate word among the plurality of candidate words represented in the user-specific language model FST.

17. The non-transitory computer-readable medium of claim 14 , wherein the o

Assignees

• Apple Inc

Inventors

• Huang Rongqing
• OPARIN Ilya

Classifications

• G10L15/16

  using artificial neural networks · CPC title

• G10L15/193

  Formal grammars, e.g. finite state automata, context free grammars or word networks · CPC title

• G10L2015/085

  Methods for reducing search complexity, pruning · CPC title

• G10L15/285 Primary

  Memory allocation or algorithm optimisation to reduce hardware requirements · CPC title

• G10L15/197 Primary

  Probabilistic grammars, e.g. word n-grams · CPC title