TRAINING ACOUSTIC MODELS

• US 20130006612A1
• Filed: 06/29/2012
• Published: 01/03/2013
• Est. Priority Date: 06/30/2011
• Status: Active Grant

• Current Assignee: Google LLC (Alphabet Inc.)
• Original Assignee: Google Inc. (Alphabet Inc.)
• Inventors: Xu, Peng, Pereira, Fernando, Chelba, Ciprian I.
• Granted Patent: US 8,959,014 B2
• Time in Patent Office: Days
• US Class Current: 704/9
• CPC Class Codes: G10L 15/063 Training
• CPC Class Codes: G10L 15/14 using statistical models, e.g. Hidden Markov Models [HMMs] G10L15/18 takes precedence
• CPC Class Codes: G10L 15/187 Phonemic context, e.g. pronunciation rules, phonotactical constraints or phoneme n-grams
• CPC Class Codes: G10L 15/34 Adaptation of a single recogniser for parallel processing, e.g. by use of multiple processors or cloud computing
• CPC Class Codes: G10L 2015/0631 Creating reference templates; Clustering

Abstract

Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for training acoustic models. Speech data and data identifying a transcription for the speech data are received. A phonetic representation for the transcription is accessed. Training sequences are identified for a particular phone in the phonetic representation. Each of the training sequences includes a different set of contextual phones surrounding the particular phone. A partitioning key is identified based on a sequence of phones that occurs in each of the training sequences. A processing module to which the identified partitioning key is assigned is selected. Data identifying the training sequences and a portion of the speech data are transmitted to the selected processing module.

Claims

1. A system comprising:

one or more computers and one or more storage devices storing instructions that are operable, when executed by the one or more computers, to cause the one or more computers to perform operations comprising;

receiving speech data and data identifying a transcription for the speech data;

accessing a phonetic representation for the transcription;

extracting training sequences from the phonetic representation for a particular phone in the phonetic representation, each of the training sequences including a different set of contextual phones surrounding the particular phone;

identifying a partitioning key based on a sequence of phones that occurs in each of the training sequences;

selecting, from among a plurality of processing modules, a processing module to which the identified partitioning key is assigned, the processing module being designated to train a portion of an acoustic model that corresponds to the identified partitioning key; and

transmitting, to the selected processing module, (i) data identifying the training sequences and (ii) a portion of the speech data that corresponds to the training sequence that includes the most contextual phones.

2. The system of claim 1, wherein accessing the phonetic representation for the transcription comprises accessing a phonetic representation comprising context-independent phones.

3. The system of claim 1, wherein receiving speech data comprises receiving feature vectors that indicate speech characteristics.

4. The system of claim 1, wherein transmitting at least a portion of the speech data comprises transmitting a speech data instance for fewer than all of the training sequences in the set of training sequences.

5. The system of claim 1, wherein transmitting the at least a portion of the speech data comprises transmitting the speech data corresponding to the training sequence that includes the most contextual phones, without transmitting additional speech data for the other training sequences to the selected processing module.

6. The system of claim 1, wherein the operations further comprise:

receiving, at the selected processing module, the data identifying the training sequences and the portion of the speech data that corresponds to the training sequence that includes the most contextual phones; and

accessing, at the selected processing module, a different subset of the received speech data for each of the training sequences.

7. The system of claim 1, wherein identifying the partitioning key based on the sequence of phones that occurs in each of the training sequences comprises identifying the partitioning key based on a sequence of two or more consecutive phones that occurs in each of the training sequences.

8. The system of claim 1, wherein identifying the partitioning key based on the sequence of phones that occurs in each of the training sequences comprises identifying the partitioning key based on a sequence that includes one contextual phone before the particular phone and one contextual phone after the particular phone.

9. The system of claim 1, wherein identifying the partitioning key based on the sequence of phones that occurs in each of the training sequences comprises identifying a partitioning key for each of the training sequences, wherein the same partitioning key is identified for each of the training sequences.

10. The system of claim 1, wherein the training sequences are first training sequences that each comprise a central triphone;

wherein identifying the partitioning key based on the sequence of phones that occurs in each of the training sequences comprises identifying the partitioning key based on the central triphone included in the first training sequences; and

wherein the operations further comprise transmitting, to the processing module and not to any of the other processing modules in the plurality of processing modules, data identifying second training sequences comprising the same central triphone included in the first training sequences, the second training sequences being extracted from a phonetic representation for a transcription for second speech data.

11. The system of claim 1, wherein extracting the training sequences for the particular phone in the phonetic representation comprises identifying at least

a first sequence that includes one contextual phone before the particular phone and one contextual phone after the particular phone,

a second sequence that includes two contextual phones before the particular phone and two contextual phones after the particular phone, and

a third sequence that includes three contextual phones before the particular phone and three contextual phones after the particular phone.

12. The system of claim 1, wherein extracting the training sequences for the particular phone in the phonetic representation comprises extracting sequences of consecutive phones in the phonetic representation.

13. The system of claim 1, wherein the operations further comprise:

receiving, at the selected processing module, the data identifying the training sequences; and

aggregating, at the selected processing module, the portion of the speech data with speech data for other instances of the training sequences.

14. The system of claim 13, wherein the operations further comprise:

generating, at the selected processing module, a model for a first training sequence of the training sequences based on the aggregated speech data for first the training sequence; and

storing the generated model in a distributed associative array, the generated model being stored in a partition of the distributed associative array being associated with the identified partitioning key.

15. The system of claim 14, wherein generating the model for the first training sequence comprises generating a context-dependent Gaussian mixture model dependent on the sequence of contextual phones included in the first training sequence, the Gaussian mixture model representing the output distribution of a Hidden Markov Model state of a central phone of the first training sequence.

16. The system of claim 14, wherein storing the generated model in the distributed associative array comprises storing the generated model in the distributed associative array such that the generated model is associated with a key that uniquely corresponds to the first training sequence.

17. The system of claim 13, wherein the operations further comprise:

determining, at the selected processing module, that the aggregated speech data includes data for fewer than a threshold number of instances of a second training sequence of the training sequences; and

in response to determining that the aggregated speech data includes data for fewer than the threshold number of instances of the second training sequence, not generating a model for the second training sequence.

18. A computer-implemented method, comprising:

receiving speech data and data identifying a transcription for the speech data;

accessing a phonetic representation for the transcription;

extracting training sequences from the phonetic representation for a particular phone in the phonetic representation, each of the training sequences including a different set of contextual phones surrounding the particular phone;

identifying a partitioning key based on a sequence of phones that occurs in each of the training sequences;

selecting, from among a plurality of processing modules, a processing module to which the identified partitioning key is assigned, the processing module being designated to train a portion of an acoustic model that corresponds to the identified partitioning key; and

transmitting, to the selected processing module, (i) data identifying the training sequences and (ii) a portion of the speech data that corresponds to the training sequence that includes the most contextual phones.

19. The computer-implemented method of claim 18, wherein identifying the partitioning key based on the sequence of phones that occurs in each of the training sequences comprises identifying the partitioning key based on a sequence of two or more consecutive phones that occurs in each of the training sequences.

20. A computer storage medium encoded with a computer program, the program comprising instructions that when executed by one or more computers cause the one or more computers to perform operations comprising:

receiving speech data and data identifying a transcription for the speech data;

accessing a phonetic representation for the transcription;

extracting training sequences from the phonetic representation for a particular phone in the phonetic representation, each of the training sequences including a different set of contextual phones surrounding the particular phone;

identifying a partitioning key based on a sequence of phones that occurs in each of the training sequences;

selecting, from among a plurality of processing modules, a processing module to which the identified partitioning key is assigned, the processing module being designated to train a portion of an acoustic model that corresponds to the identified partitioning key; and

transmitting, to the selected processing module, (i) data identifying the training sequences and (ii) a portion of the speech data that corresponds to the training sequence that includes the most contextual phones.