Speech Recognition Model Training for Mixed Short and Long Samples

Solution Overview

Problem

Existing speech recognition models struggle to accurately process speech content of varying time lengths, leading to poor recognition efficiency and accuracy, especially for long speech samples.

Innovation Solution

A speech recognition model is trained using a mixed set of speech samples with different time lengths, including short and long samples, to adjust parameters of the speech encoding and decoding units, enhancing the model's ability to handle diverse speech durations.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If the speech recognition model is trained only with short speech samples, then the model can process short speech content efficiently, but the model performs poorly on long speech samples

Engineering Contradiction:

Improverecognition efficiency for short speechVSAvoidrecognition accuracy for long speech

Core Design Contradiction:

ProductivityVSReliability

Solution Approach 1:

The training data is segmented into two distinct sets: a first set of short speech samples (time length < first threshold) and a second set of long speech samples (time length > second threshold). This segmentation allows the model to be trained separately on different time length categories, enabling it to master both short and long speech recognition without the performance degradation that would occur with mixed-length training alone.

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

The speech recognition model is designed to perform multiple functions by handling both short and long speech samples through a unified architecture. The model learns universal speech encoding and decoding capabilities that can adapt to varying time lengths, making it universally applicable to different speech duration scenarios rather than requiring separate specialized models.

Inventive Principle:

Principle #6Universality (Multi-functionality)

2Reliability

If the speech recognition model is trained only with long speech samples, then the model can handle long speech content, but the model is less efficient for short speech processing

Engineering Contradiction:

Improverecognition accuracy for long speechVSAvoidrecognition efficiency for short speech

Core Design Contradiction:

ReliabilityVSProductivity

Solution Approach 1:

The training data is segmented into two distinct sets: a first set of short speech samples (time length < first threshold) and a second set of long speech samples (time length > second threshold). This segmentation allows the model to be trained separately on different time length categories, enabling it to master both short and long speech recognition without the performance degradation that would occur with mixed-length training alone.

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

The model employs dynamic time length adaptation where the speech encoding unit and decoding unit can adjust their processing behavior based on the input speech duration. The model dynamically switches between optimized processing modes for short versus long speech samples, allowing it to maintain high efficiency across varying time lengths rather than being optimized for a single fixed duration.

Inventive Principle:

Principle #15Dynamics

3Ease of manufacture

If the speech recognition model uses a single training data set with uniform time length, then the training process is simple, but the model cannot adapt to diverse speech durations

Engineering Contradiction:

Improvetraining process simplicityVSAvoidability to handle diverse time lengths

Core Design Contradiction:

Ease of manufactureVSAdaptability or versatility

Solution Approach 1:

The training data is segmented into two distinct sets: a first set of short speech samples (time length < first threshold) and a second set of long speech samples (time length > second threshold). This segmentation allows the model to be trained separately on different time length categories, enabling it to master both short and long speech recognition without the performance degradation that would occur with mixed-length training alone.

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

The model utilizes parameter changes in the speech encoding unit and decoding unit to adapt to different time lengths. By adjusting model parameters based on the specific time length characteristics of the training data, the model achieves versatile performance across diverse speech durations while maintaining a relatively simple training process structure.

Inventive Principle:

Principle #35Parameter changes

Data Source

PatentUS20250378822A1Speech recognition model training

Publication Date: 2025.12.11 BEIJING ZITIAO NETWORK TECH CO LTD

AI summary

A method, an apparatus, a device, and a storage medium related to training a speech recognition model are provided. An example method provided here includes: obtaining a speech sample set, the speech sample set including a first set of speech samples and a second set of language samples, a time length of the first set of speech samples being less than a first threshold, and a time length of the second set of speech samples being greater than a second threshold; and training the speech recognition model with the speech sample set and corresponding text information, to at least adjust parameters of a speech encoding unit in the speech recognition model, the speech recognition model including the speech encoding unit configured to generate a speech encoded representation of speech content and a decoding unit configured to generate a speech recognition result based on the speech encoded representation.