ISSUANCE OF WORD TIMING BY END-TO-END MODEL

Title:

ISSUANCE OF WORD TIMING BY END-TO-END MODEL

Document Type and Number:

Japanese Patent JP2023109914

Kind Code:

A

Abstract:

To provide a method and system for 2-pass end-to-end speech recognition.SOLUTION: In a speech environment 100, a method to be mounted on a user device 110 receives a training example including audio data 202 representing oral speech 12 and ground truth transcription, inserts a placeholder symbol before a word about each word in speech, identifies a ground truth alignment for a start and an end of the word, and generates first and second constrained alignments for start and end word pieces. The first constrained alignment is matched to a ground truth alignment for a start of each word, and the second constrained alignment is matched to a ground truth alignment for an end of each word. The method constrains an attention head of a second path decoder by applying the first and second constrained alignments.SELECTED DRAWING: Figure 1A

More Like This:

JP2024023550	Assessing speaker recognition performance
WO/2023/060002	TRAINING FOR LONG-FORM SPEECH RECOGNITION
WO/2017/218492	NEURAL DECODING OF ATTENTIONAL SELECTION IN MULTI-SPEAKER ENVIRONMENTS

Inventors:

TARA N SAINATH
BASILIO GARCIA CASTILLO
DAVID RYBACH
TREVOR STROHMAN
PANG RUOMING

Application Number:

JP2023084811A

Publication Date:

August 08, 2023

Filing Date:

May 23, 2023

Export Citation:

Click for automatic bibliography generation Help

Assignee:

GOOGLE LLC

International Classes:

G10L15/16; G10L15/04

Attorney, Agent or Firm:

Yasuhiko Murayama
Shinya Mihiro
Tatsuhiko Abe

Previous Patent: Manufacturing method of reflective mask blank

Next Patent: Transdermal analyte sensor system and method