Speech Recognition With Augmented Acoustic Score Integration
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing speech recognition technologies face performance degradation due to environmental noise, varying speech speeds, speaker features, and amplitude differences, which are not effectively addressed by single or multiple acoustic models, leading to increased time, cost, and memory consumption.
Innovation Solution
A speech recognition apparatus generates multiple augmented speech data through speed, volume, and quality conversions, integrates acoustic scores, and generates an integrated lattice using a single acoustic model to improve recognition performance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If multiple acoustic models are used to address different factors (noise, speech speed, speaker features, amplitude), then recognition performance is improved, but training time and computational cost increase significantly
Solution Approach 1:
The patent segments the input speech data into multiple augmented versions by applying different transformations (speed conversion, volume conversion, quality conversion) to create diverse training samples from a single original speech, effectively replacing the need for multiple acoustic models with one model trained on transformed data
Solution Approach 2:
The patent changes parameters of the input speech data (speed, volume, quality) through conversion processes to generate augmented speech data that covers various speaking conditions, allowing a single acoustic model to handle multiple factors that would otherwise require separate models
2Reliability
If multiple acoustic models are used to address different factors, then recognition performance is improved, but memory capacity requirements increase significantly
Solution Approach 1:
The patent merges multiple acoustic models into a single acoustic model by training it on augmented speech data that combines various speech characteristics (different speeds, volumes, qualities) through data transformation, thereby reducing memory requirements while maintaining comprehensive recognition capability
Solution Approach 2:
The single acoustic model is designed to be universal by training it on diverse augmented speech data that represents multiple speaking conditions and factors, enabling it to perform functions that would traditionally require multiple specialized models
3Object-affected harmful factors
If noise-robust speech recognition is executed by inputting both input signal and speech-emphasized signal to a single acoustic model, then environmental noise is addressed, but speech speed, speaker features, and amplitude differences cannot be resolved
Solution Approach 1:
The patent segments the noise suppression task from the speech enhancement task by applying noise suppression to create speech-emphasized signals, then separately applying speed, volume, and quality conversions to create augmented data, allowing the acoustic model to handle both noise and speech variability independently
Solution Approach 2:
The patent performs preliminary processing steps (noise suppression, speed conversion, volume conversion, quality conversion) on the input speech data before feeding it to the acoustic model, preparing the data in advance to cover various conditions so the model can focus on accurate recognition
Data Source
AI summary
According to one embodiment, a speech recognition apparatus includes processing circuitry. The processing circuitry generates a plurality of augmented speech data, based on input speech data, generates a plurality of acoustic scores, based on the plurality of augmented speech data and an acoustic model, generates a plurality of adjusted acoustic scores by resampling the acoustic scores, generates an integrated acoustic score by integrating the adjusted acoustic scores, generates an integrated lattice, based on the integrated acoustic score, a pronunciation dictionary, and a language model, and searches a speech recognition result with a highest likelihood from the integrated lattice.