Real-time Voice Recognition with Conformer Buffer Mechanism

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Real-time voice recognition systems face significant delays due to the high computational load during the encoding process, which hinders their effectiveness in providing immediate recognition results.

Innovation Solution

The implementation of a real-time voice recognition method that utilizes a buffer region to store intermediate processing results, allowing for the encoding of audio feature sequences in chunks with overlapping frames, reducing redundant computations and enhancing processing speed by using a Conformer network with a buffer mechanism to encode and decode voice data efficiently.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If the encoding process processes the entire audio feature sequence at once, then the recognition accuracy is improved, but the recognition delay increases due to high computational load

Engineering Contradiction:
Improverecognition accuracyVSAvoidrecognition delay
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The patent divides the audio feature sequence into multiple chunks, where each chunk contains multiple audio frames. The encoding process processes these chunks sequentially rather than the entire sequence at once, reducing the computational load at each step and enabling faster recognition while maintaining accuracy through proper chunk design with overlapping frames

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent performs preliminary encoding on historical chunks and stores their intermediate processing results in a buffer region before they are needed for decoding. This allows the decoding process to retrieve pre-computed results immediately, reducing the overall recognition delay while maintaining accuracy

Inventive Principle:
Principle #10Preliminary action

2Loss of time

If the encoding process processes smaller chunks of audio frames, then the recognition delay is reduced, but the computational redundancy increases due to repeated processing of overlapping frames

Engineering Contradiction:
Improverecognition delayVSAvoidcomputational redundancy
Core Design Contradiction:
Loss of timeVSLoss of energy

Solution Approach 1:

The patent performs preliminary encoding on historical chunks and stores their intermediate processing results in a buffer region. When subsequent chunks need to process overlapping frames, the system retrieves the pre-computed intermediate results from the buffer instead of re-computing them, significantly reducing computational redundancy while enabling fast processing of smaller chunks

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The patent copies the intermediate processing results from historical chunks to the buffer region, making them readily available for reuse. This copying mechanism allows the system to avoid re-processing overlapping frames in subsequent chunks, reducing computational energy consumption while maintaining the ability to process smaller chunks quickly

Inventive Principle:
Principle #26Copying

Data Source

PatentUS20240062744A1Real-time voice recognition method, model training method, apparatuses, device, and storage medium
Publication Date: 2024.02.22 TENCENT TECHNOLOGY (SHENZHEN) CO LTD
  • US20240062744A1 patent drawing
  • US20240062744A1 patent drawing
  • US20240062744A1 patent drawing

AI summary

A real-time voice recognition method and a real-time voice recognition model training method are provided. The model training method includes: obtaining an audio feature sequence of sample voice data, the audio feature sequence comprising audio features of a plurality of audio frames of the sample voice data; inputting the audio feature sequence to an encoder of the real-time voice recognition model; chunking the audio feature sequence into a plurality of chunks by the encoder according to a mask matrix; encoding each of the chunks to obtain a hidden layer feature sequence of the sample voice data; decoding the hidden layer feature sequence by a decoder of the real-time voice recognition model to obtain a predicted recognition result for the sample voice data; and training the real-time voice recognition model based on the predicted recognition result and a real recognition result of the sample voice data.