Managed ASR Service with Pre-Trained Acoustic Models

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Implementing machine learning techniques, particularly automatic speech recognition (ASR), is challenging due to algorithmic complexities, resource intensity, and the need for frequent model regeneration when environments or use cases change, differing significantly from traditional software engineering practices.

Innovation Solution

A fully managed and continuously trained ASR service that uses a provider network to upload audio files, analyze them using acoustic and language models, and automatically generate transcripts, with features like punctuation and formatting, enabling efficient transcription and analysis of audio and video content across various applications.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If machine learning models are developed and deployed using traditional approaches, then model accuracy can be achieved, but the time and resources required for model generation and updates become extremely consumptive

Engineering Contradiction:
Improvemodel accuracyVSAvoidmodel generation time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The system pre-trains acoustic models using large corpora of audio data in advance, so that when deployment is needed, the model is already prepared and can be quickly adapted to specific use cases rather than training from scratch each time

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The system uses pre-trained models as a base and performs targeted fine-tuning by adjusting only the necessary parameters for specific applications, rather than retraining the entire model, thus maintaining accuracy while reducing time and resource consumption

Inventive Principle:
Principle #35Parameter changes

2Reliability

If machine learning models are highly specialized for particular use cases, then model performance improves, but any change to the environment or use case requires complete regeneration of a new model

Engineering Contradiction:
Improvemodel performanceVSAvoidmodel adaptability
Core Design Contradiction:
ReliabilityVSAdaptability or versatility

Solution Approach 1:

The system separates the acoustic model training into hierarchical layers: a general base model trained on diverse data, and specialized adapters or fine-tuned components for specific domains. This allows the core model to remain while only specific components are adjusted for new use cases

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The system creates a universal base acoustic model that can serve multiple different speech recognition tasks and domains. The model is designed with modular components that can be adapted to different languages, accents, and application contexts without requiring complete regeneration

Inventive Principle:
Principle #6Universality (Multi-functionality)

3Measurement precision

If machine learning techniques are implemented with high accuracy, then speech recognition quality improves, but the computational resources and complexity required become tremendously difficult

Engineering Contradiction:
Improvespeech recognition accuracyVSAvoidsystem complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The system introduces pre-processing and post-processing layers that act as intermediaries between the acoustic model and the final transcription output. These intermediaries handle tasks like noise reduction, speaker verification, and result formatting, allowing the core acoustic model to remain relatively simple while achieving high overall system accuracy

Inventive Principle:
Principle #24Intermediary (Mediator)

Data Source

PatentEP3698357B1Automatic speech recognition service
Publication Date: 2022.04.06 AMAZON TECH INC
  • EP3698357B1 patent drawingFigure 1
  • EP3698357B1 patent drawingFigure 2
  • EP3698357B1 patent drawingFigure 3

AI summary

Techniques for automated speech recognition (ASR) are described. A user can upload an audio file to a storage location. The user then provides the ASR service with a reference to the audio file. An ASR engine analyzes the audio file, using an acoustic model to divide the audio data into words, and a language model to identify the words spoken in the audio file. The acoustic model can be trained using audio sentence data, enabling the transcription service to accurately transcribe lengthy audio data. The results are punctuated and normalized, and the resulting transcript is returned to the user.