Parallel Speech Recognition Engines Merging Fast and Accurate Results

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Speech recognition systems face a trade-off between accuracy and latency, with long look-ahead buffers improving accuracy but causing sluggishness, and short buffers reducing accuracy, leading to user dissatisfaction.

Innovation Solution

Utilize a primary high-accuracy, high-latency speech recognition engine (SRE) and a secondary low-latency, low-accuracy SRE in parallel, merging their results to enhance accuracy while reducing perceived latency by replacing incorrect words with correct ones from the primary SRE.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If a long look-ahead buffer is used in speech recognition, then accuracy is improved, but latency increases causing sluggish response

Engineering Contradiction:
Improvespeech recognition accuracyVSAvoidlatency period
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The patent segments the speech recognition task into two parallel processes: a first speech recognition engine that processes audio data with lower latency to provide preliminary results, and a second speech recognition engine that processes the same audio data with higher accuracy. The results are then merged, with the second engine's corrections applied to improve accuracy without requiring the user to wait for complete processing of the entire recognition pipeline.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The first speech recognition engine performs preliminary speech recognition on audio data before the second engine completes its analysis. This preliminary action provides initial results to the user quickly, while the second engine works in parallel to provide corrections. The system displays preliminary results immediately and updates them with corrections from the second engine, ensuring both speed and accuracy.

Inventive Principle:
Principle #10Preliminary action

2Loss of time

If a short look-ahead buffer is used in speech recognition, then latency is reduced improving responsiveness, but accuracy deteriorates

Engineering Contradiction:
Improvelatency periodVSAvoidspeech recognition accuracy
Core Design Contradiction:
Loss of timeVSMeasurement precision

Solution Approach 1:

The patent merges the results from two parallel speech recognition engines: a first engine that provides low-latency preliminary recognition and a second engine that provides high-accuracy recognition. The merging process combines the speed advantage of the first engine with the accuracy advantage of the second engine, delivering both responsive performance and high recognition accuracy simultaneously.

Inventive Principle:
Principle #5Merging (Combining)

Solution Approach 2:

The patent introduces a result merging mechanism that acts as an intermediary between the two speech recognition engines and the user interface. This intermediary receives results from both engines, reconciles them by applying corrections from the second engine to the first engine's output, and presents the final merged result to the user, thereby coordinating the conflicting requirements of speed and accuracy.

Inventive Principle:
Principle #24Intermediary (Mediator)

Data Source

PatentEP4264602B1Speech recognition using a primary and a secondary speech recognition engine
Publication Date: 2025.08.20 MICROSOFT TECHNOLOGY LICENSING LLC
  • EP4264602B1 patent drawingFigure 1
  • EP4264602B1 patent drawingFigure 2
  • EP4264602B1 patent drawingFigure 3

AI summary

Disclosed speech recognition techniques improve user-perceived latency while maintaining accuracy by: receiving an audio stream, in parallel, by a primary (e.g., accurate) speech recognition engine (SRE) and a secondary (e.g., fast) SRE; generating, with the primary SRE, a primary result; generating, with the secondary SRE, a secondary result; appending the secondary result to a word list; and merging the primary result into the secondary result in the word list. Combining output from the primary and secondary SREs into a single decoder as described herein improves user-perceived latency while maintaining or improving accuracy, among other advantages.