RNN-T Audio Decoding with Diagonal Switching for Parallel Inference

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

RNN-T technologies in audio processing face significant computational and memory challenges, limiting their ability to perform parallel and fast calculations, which reduces their practical application.

Innovation Solution

A method for audio processing that involves diagonal switching of output positions in the decoding path of RNN-T models, reducing computation and memory usage by incrementing coordinates in the audio frame and text label sequence dimensions based on non-null or null character outputs, and adjusting network parameters through constrained training paths.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If RNN-T technologies are used for audio processing, then recognition accuracy is improved, but computational complexity and memory usage increase significantly

Engineering Contradiction:
Improverecognition accuracyVSAvoidcomputational complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent segments the decoding process into multiple parallel decoding paths, where each path processes a specific coordinate increment pattern. This divides the complex RNN-T computation into manageable segments that can be executed in parallel, reducing overall computational complexity while maintaining recognition accuracy.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent introduces a new dimension by adding coordinate increment constraints in the audio frame and text label sequence dimensions. This transforms the traditional single-path decoding into a multi-dimensional parallel decoding structure, enabling faster computation without sacrificing accuracy.

Inventive Principle:
Principle #17Another dimension (Dimensionality change)

2Measurement precision

If RNN-T technologies are used for audio processing, then recognition accuracy is improved, but memory occupation increases

Engineering Contradiction:
Improverecognition accuracyVSAvoidmemory occupation
Core Design Contradiction:
Measurement precisionVSQuantity of substance

Solution Approach 1:

By segmenting the decoding into multiple paths with different coordinate increment patterns, the patent reduces the memory footprint of each individual path. Each path processes a subset of the total computation, requiring less memory per path, while the parallel execution maintains overall accuracy.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent employs partial action by limiting each decoding path to specific coordinate increment patterns rather than exploring all possible paths. This partial exploration reduces memory requirements while still achieving accurate recognition through the combination of multiple constrained paths.

Inventive Principle:
Principle #16Partial or excessive action

3Reliability

If traditional RNN-T decoding is used, then comprehensive path exploration is achieved, but parallel and fast calculation is prevented

Engineering Contradiction:
Improvepath exploration completenessVSAvoidcalculation speed
Core Design Contradiction:
ReliabilityVSProductivity

Solution Approach 1:

The patent segments the complete path exploration into multiple parallel decoding paths, each handling specific coordinate increment patterns. This segmentation enables simultaneous execution of multiple paths, dramatically increasing calculation speed while collectively covering the comprehensive search space.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

By adding the dimension of parallel path execution with different coordinate increment constraints, the patent transforms sequential path exploration into concurrent multi-dimensional exploration. This enables both comprehensive coverage and high-speed calculation through parallel processing.

Inventive Principle:
Principle #17Another dimension (Dimensionality change)

Data Source

PatentEP4287180B1Method and apparatus for audio processing, electronic device and storage medium
Publication Date: 2025.12.03 BEIJING XIAOMI MOBILE SOFTWARE CO LTD
  • EP4287180B1 patent drawingFigure 1~2
  • EP4287180B1 patent drawingFigure 3~4
  • EP4287180B1 patent drawingFigure 5

AI summary

A method and apparatus for audio processing, an electronic device and a storage medium are provided. The method includes: obtaining (S101) an audio encoding result, wherein each element in the audio encoding result includes a coordinate in an audio frame number dimension and a coordinate in a text label sequence dimension; in response to an output result of an ith frame in a decoding path being a non-null character, respectively increasing (S102) the coordinate in the audio frame number dimension and the coordinate in the text label sequence dimension corresponding to an output position of the ith frame by 1 to obtain an output position of a (i+1)th frame in the decoding path; and determining (S103) an output result corresponding to the output position of the (i+1)th frame according to the output result of the ith frame and an element of the (i+1)th frame in the audio encoding result.