Seq2Seq Speech Translation Model for Multi-Language Accuracy

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Current speech translation methods are limited to one-to-one language translation and suffer from reduced accuracy due to multiple steps involved, leading to error amplification and longer processing times.

Innovation Solution

A speech translation method and apparatus that utilizes a pre-trained speech translation model to directly translate speech from one language to any other language within a set, employing a sequence-to-sequence model with an encoder and decoder to simplify the process and reduce errors, allowing for multi-language translation.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If a three-step process (speech-to-text, text translation, text-to-speech) is used for speech translation, then translation functionality is achieved, but translation accuracy deteriorates due to error accumulation across multiple steps

Engineering Contradiction:

Improvetranslation accuracyVSAvoidprocess complexity

Core Design Contradiction:

ReliabilityVSDevice complexity

Solution Approach 1:

The patent merges the three separate steps (speech-to-text, text translation, text-to-speech) into a single integrated speech translation model. This unified model directly translates source speech to target speech without intermediate text conversion steps, thereby eliminating error accumulation and improving translation accuracy while reducing overall system complexity.

Inventive Principle:

Principle #5Merging (Combining)

Solution Approach 2:

The patent extracts and removes the intermediate text representation steps from the translation process. By eliminating the speech-to-text and text-to-speech conversion steps, the system directly processes speech-to-speech translation, preventing error propagation through intermediate representations and simplifying the overall process.

Inventive Principle:

Principle #2Taking out (Extraction)

2Productivity

If a three-step process is used for speech translation, then translation functionality is achieved, but processing time increases due to multiple sequential steps

Engineering Contradiction:

Improvetranslation efficiencyVSAvoidprocessing time

Core Design Contradiction:

ProductivityVSLoss of time

Solution Approach 1:

The patent combines three sequential processing steps into a single unified speech translation model. This integration allows the system to perform speech-to-speech translation in one continuous process rather than three separate sequential steps, significantly reducing total processing time and improving translation efficiency.

Inventive Principle:

Principle #5Merging (Combining)

Solution Approach 2:

The patent skips the intermediate text conversion steps (speech-to-text and text-to-speech) and directly performs speech-to-speech translation. By rushing through the translation process in a single step without pausing for intermediate conversions, the system eliminates unnecessary processing delays and accelerates overall translation speed.

Inventive Principle:

Principle #21Skipping (Rushing through)

3Adaptability or versatility

If a three-step process is used for speech translation, then translation functionality is achieved, but the system is limited to one-to-one language translation

Engineering Contradiction:

Improvelanguage translation flexibilityVSAvoidsystem structure

Core Design Contradiction:

Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The patent creates a universal speech translation model that can handle multiple language pairs simultaneously. The unified model is designed to translate between any source language and any target language within a supported language set, making the system highly adaptable and versatile. This multi-functional capability allows a single system to serve multiple translation needs without requiring separate specialized systems for each language pair.

Inventive Principle:

Principle #6Universality (Multi-functionality)

Data Source

PatentUS11586831B2Speech translation method electronic device and computer-readable storage medium using SEQ2SEQ for determining alternative translated speech segments

Publication Date: 2023.02.21 DOUYIN VISION CO LTD

AI summary

Provided are a speech translation method and apparatus, an electronic device and a storage medium. The method includes: acquiring a source speech corresponding to a to-be-translated language; acquiring a specified target language; inputting the source speech and indication information matched with the target language into a pre-trained speech translation model, where the speech translation model is configured to translate a language in a first language set into a language in a second language set, the first language set includes a plurality of languages, the first language set includes the to-be-translated language, the second language set includes a plurality of languages, and the second language set includes the target language; and acquiring a translated speech corresponding to the target language and output by the speech translation model; where the to-be-translated language is different from the target language.