Two-Pass Speech Recognition With Deliberation for Low Latency

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing end-to-end speech recognition systems face challenges in achieving high accuracy and low latency, particularly in mobile environments, due to the limitations of recurrent neural network transducer (RNN-T) models and non-streaming listen-attend-spell (LAS) models, which hinder real-time voice transcription and on-device deployment.

Innovation Solution

A two-pass speech recognition system is implemented, comprising a first-pass RNN-T network followed by a second-pass LAS network, enhanced with a deliberation network that attends to both acoustics and first-pass text hypotheses, utilizing shared encoders and joint training to improve accuracy and reduce latency.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Loss of time

If a single-pass RNN-T model is used for speech recognition, then the system achieves low latency and streaming capability, but the Word Error Rate is high

Engineering Contradiction:
ImprovelatencyVSAvoidWord Error Rate
Core Design Contradiction:
Loss of timeVSMeasurement precision

Solution Approach 1:

The speech recognition system is divided into two distinct passes: a first-pass RNN-T decoder that performs streaming recognition with low latency, and a second-pass LAS decoder that performs accurate but non-streaming recognition. The system segments the recognition task temporally, allowing each pass to optimize for its specific function while collectively achieving both low latency and high accuracy through the merging of their output hypotheses.

Inventive Principle:
Principle #1Segmentation

2Measurement precision

If a large conventional model is used to achieve high accuracy, then the Word Error Rate is reduced, but the model size and computational complexity increase, preventing on-device deployment

Engineering Contradiction:
ImproveWord Error RateVSAvoidmodel size
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The system merges the outputs of two separate, relatively small models (RNN-T and LAS decoders) to achieve the accuracy of a large conventional model. By combining the hypothesis outputs from both decoders and selecting the best result, the system attains high recognition accuracy without requiring a single large model, thus enabling deployment on mobile devices with limited computational resources.

Inventive Principle:
Principle #5Merging (Combining)

3Measurement precision

If a two-pass system with separate encoders is used to improve accuracy, then the Word Error Rate is reduced, but the computational overhead and latency increase

Engineering Contradiction:
ImproveWord Error RateVSAvoidcomputational overhead
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The encoder is designed as a universal, shared component that serves both the RNN-T decoder and the LAS decoder in the two-pass system. This multi-functional encoder processes acoustic features and generates representations that are utilized by both decoders, eliminating the need for separate encoders and thereby reducing computational overhead while maintaining the accuracy benefits of the two-pass architecture.

Inventive Principle:
Principle #6Universality (Multi-functionality)

Data Source

PatentEP4414896B1Deliberation model-based two-pass end-to-end speech recognition
Publication Date: 2025.10.22 GOOGLE LLC
  • EP4414896B1 patent drawingFigure 1A
  • EP4414896B1 patent drawingFigure 1B
  • EP4414896B1 patent drawingFigure 2A

AI summary

Disclosed herein is a computer-implemented method when executed on data processing hardware causes the data processing hardware to perform operations comprising: receiving a sequence of audio features characterizing an utterance; based on the sequence of audio features, generating, using a first-pass decoder model, a plurality of first-pass speech recognition hypotheses, each first-pass speech recognition hypothesis corresponding to a candidate transcription of the utterance; generating, using a long short-term memory (LSTM) encoder, a first-pass encoding of the plurality of first-pass speech recognition hypotheses; and based on the sequence of audio features and the first-pass encoding, generating, using a second-pass decoder model, a second-pass hypothesis that rescores the plurality of first-pass speech recognition hypotheses.