Speech-to-Text Framework Using Distributed N-Gram Beam Search
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current speech-to-text conversion technologies face challenges in accuracy due to environmental noise, regional accents, dialects, idiomatic expressions, and individual pronunciation differences, leading to error-prone text inputs that result in false insights during content analysis.
Innovation Solution
A distributed processing system that employs a framework for speech-to-text conversion, utilizing a larger n-gram corpus and parallel processing across multiple node devices to improve accuracy by identifying likely sentence pauses, generating probability distributions for graphemes and n-grams, and performing beam searches to enhance transcript generation and text analytics.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional speech-to-text conversion is used, then processing speed is maintained, but accuracy deteriorates due to environmental noise, accents, and pronunciation variations
Solution Approach 1:
The patent divides the speech-to-text processing into distinct stages: acoustic model processing, n-gram language model processing, and beam search decoding. Each stage handles specific aspects of the conversion, with the n-gram corpus providing contextual probability information that complements the acoustic model's phoneme-level analysis. This segmentation allows the system to address accuracy issues at multiple levels without creating a monolithic complex system.
Solution Approach 2:
The n-gram language model serves as an intermediary between the acoustic model and the final text output. It provides probabilistic information about word sequences based on large corpora, acting as a bridge that translates acoustic probabilities into linguistically valid text. This intermediary layer helps resolve ambiguities caused by noise and accents by selecting sequences that are statistically more likely in natural language.
2Measurement precision
If larger n-gram corpora are used, then speech-to-text accuracy is improved, but computational resource requirements increase
Solution Approach 1:
The system uses beam search with a limited beam width to explore only the most probable word sequences rather than exhaustively searching all possible sequences. This partial exploration approach allows the system to leverage large n-gram corpora for accurate probability estimation while avoiding the computational burden of evaluating every possible word combination. The beam width parameter controls the trade-off between accuracy and computational resources.
Solution Approach 2:
The system dynamically adjusts processing parameters based on the confidence levels from the acoustic model and n-gram probabilities. When the acoustic model provides high-confidence predictions, the system can use smaller beam widths or skip certain n-gram lookups, reducing computational overhead. Conversely, when confidence is low, the system can allocate more computational resources to explore alternative sequences, optimizing the balance between accuracy and resource usage.
3Productivity
If parallel processing across multiple node devices is implemented, then processing efficiency is improved, but system coordination complexity increases
Solution Approach 1:
The patent divides the speech-to-text processing workload into independent segments that can be executed in parallel across multiple node devices. Each node can independently process different portions of the speech audio or different aspects of the language model queries, with results aggregated by a coordination layer. This segmentation enables efficient parallel processing while keeping individual node complexity manageable.
Solution Approach 2:
The system employs a universal n-gram corpus that can be shared and accessed by all node devices in the distributed system. This universal resource provides consistent probability information across all nodes, ensuring that parallel processing produces coherent results. The coordination mechanism leverages this universality by aggregating results from multiple nodes without requiring complex inter-node communication, as each node independently queries the same n-gram probabilities.
Data Source
AI summary
An apparatus includes processor(s) to: generate a set of candidate n-grams based on probability distributions from an acoustic model for candidate graphemes of a next word most likely spoken following at least one preceding word spoken within speech audio; provide the set of candidate n-grams to multiple devices; provide, to each node device, an indication of which candidate n-grams are to be searched for within the n-gram corpus by each node device to enable searches for multiple candidate n-grams to be performed, independently and at least partially in parallel, across the node devices; receive, from each node device, an indication of a probability of occurrence of at least one candidate n-gram within the speech audio; based on the received probabilities of occurrence, identify the next word most likely spoken within the speech audio; and add the next word most likely spoken to a transcript of the speech audio.