Speech Recognition Decoding With Text-Set Constraint Filtering

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing speech recognition technologies face challenges in achieving high accuracy due to limited training data and slow iterative updates, especially in dynamic service environments, leading to poor recognition performance and incomplete task completion.

Innovation Solution

A method combining an acoustic model with a neural network language model (NNLM) and a constraint sub-model to guide the decoding process, using beam search to select candidate paths based on combined probabilities, ensuring the recognized text aligns with a predetermined text set.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If traditional speech recognition uses only acoustic models with limited training data, then the system complexity remains low, but the recognition accuracy deteriorates

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

Solution Approach 1:

The patent combines acoustic models with neural network language models to form a hybrid speech recognition system. The acoustic model processes speech signals to obtain acoustic probabilities, while the NNLM processes text sequences to obtain language probabilities. These two models are merged through a scoring function that combines acoustic probabilities and language probabilities, achieving higher recognition accuracy than either model alone while managing complexity through modular integration.

Inventive Principle:
Principle #5Merging (Combining)

2Measurement precision

If speech recognition systems are updated iteratively to improve accuracy, then recognition performance improves, but the update speed deteriorates due to slow iterative processes

Engineering Contradiction:
Improverecognition accuracyVSAvoidupdate speed
Core Design Contradiction:
Measurement precisionVSSpeed

Solution Approach 1:

The patent pre-trains the neural network language model on large-scale text corpora before deployment. This preliminary action allows the NNLM to learn linguistic patterns and constraints in advance, so that during the speech recognition process, the model can immediately apply these learned patterns without requiring slow iterative updates. The constraint sub-model is also pre-trained on text sets to establish beforehand which text sequences are valid, enabling fast constraint-based filtering during recognition.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The patent uses beam search decoding to generate multiple candidate transcription paths simultaneously, copying the decoding process across multiple hypotheses. This allows the system to explore multiple possible interpretations of the speech input in parallel, improving recognition accuracy without requiring slow sequential iterative updates. The beam search maintains multiple candidate paths with their respective scores, enabling efficient selection of the most likely transcription.

Inventive Principle:
Principle #26Copying

3Reliability

If the decoding process considers all possible candidate transcriptions, then the completeness of task completion improves, but the computational time deteriorates

Engineering Contradiction:
Improvetask completion completenessVSAvoiddecoding time
Core Design Contradiction:
ReliabilityVSLoss of time

Solution Approach 1:

The patent changes the parameter space by introducing language probabilities from the NNLM as an additional dimension for evaluating candidate transcriptions. Instead of relying solely on acoustic probabilities, the system evaluates candidates based on a combination of acoustic probabilities and language probabilities. The constraint sub-model further changes parameters by filtering candidates based on whether they match predefined text sets, reducing the effective search space while maintaining completeness for valid transcriptions.

Inventive Principle:
Principle #35Parameter changes

Solution Approach 2:

The patent applies different evaluation criteria to different parts of the decoding process. The beam search uses acoustic probabilities for initial candidate generation, while the constraint sub-model applies language probabilities and text set constraints to filter and rank candidates. This local differentiation of quality metrics allows efficient pruning of invalid candidates early in the process while maintaining thorough evaluation of promising candidates, reducing overall decoding time without sacrificing completeness.

Inventive Principle:
Principle #3Local quality

Data Source

PatentEP4407606B1Speech recognition method and apparatus, and device and medium
Publication Date: 2025.11.12 BEIJING BAIDU NETCOM SCI & TECH CO LTD
  • EP4407606B1 patent drawingFigure 1~2
  • EP4407606B1 patent drawingFigure 3
  • EP4407606B1 patent drawingFigure 4

AI summary

A method and an apparatus of recognizing a speech, a device, and a medium. The method includes: processing (S210), by using an acoustic model, speech data to be recognized and a first text segment obtained by recognition to obtain respective acoustic probabilities of a plurality of candidate text segments; processing (S220) the first text segment by using a first language sub-model to obtain respective initial language probabilities of the plurality of candidate text segments; processing (S230) the first text segment by using a constraint sub-model to obtain extendibility relationships of the plurality of candidate text segments with respect to the first text segment; adjusting (S240) the initial language probabilities of the candidate text segments according to the extendibility relationships to obtain respective first language probabilities of the plurality of candidate text segments; and determining (S250) a target text segment from the plurality of candidate text segments according to the first language probabilities and the acoustic probabilities.