CTC Acoustic Model Training Without Fixed Alignments
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional speech recognition systems require lengthy training procedures involving fixed alignment targets and prior models, making them inefficient for updating and refreshing acoustic models, especially when pronunciations or training data change.
Innovation Solution
The use of connectionist temporal classification (CTC) techniques for training acoustic models without fixed alignment targets, allowing for the generation of context-dependent phone inventories using approximate alignments, and enabling the training of CTC models directly from audio and written-domain transcriptions, which reduces training time and facilitates model updates.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If conventional training procedures with fixed alignment targets and prior models are used, then manufacturing precision of acoustic models is improved, but productivity deteriorates due to lengthy training time
Solution Approach 1:
The patent applies preliminary action by pre-determining context-dependent phone inventories using approximate alignments from a first CTC model before training the second CTC model. This preliminary preparation of training data structures and phone inventories enables the second model to be trained more efficiently without requiring lengthy conventional alignment procedures, thus improving training speed while maintaining accuracy.
Solution Approach 2:
The patent uses copying by generating context-dependent phone inventories from approximate alignments of a first CTC model, then using these copied structures as training targets for a second CTC model. This copying approach allows the system to leverage previously computed approximate alignments without repeating the lengthy fixed alignment target generation process, thereby improving productivity while preserving manufacturing precision.
2Manufacturing precision
If fixed alignment targets are used for training, then manufacturing precision of acoustic models is improved, but loss of time increases due to lengthy training procedures
Solution Approach 1:
The patent applies this principle by using a first CTC model to generate approximate alignments that are then discarded after being used to determine context-dependent phone inventories for training the second model. These approximate alignments serve as temporary, disposable training aids that enable efficient training without requiring the system to maintain or recompute expensive fixed alignment targets throughout the entire training process.
Solution Approach 2:
The patent changes the parameter of alignment precision by transitioning from exact fixed alignment targets in conventional methods to approximate alignments generated by CTC models. This parameter change allows the system to accept lower precision intermediate alignments that are sufficient for determining context-dependent phone inventories, thereby significantly reducing training time while maintaining final model accuracy through the two-stage CTC training approach.
3Manufacturing precision
If GMM and DNN models are used in the training pipeline, then manufacturing precision is improved, but device complexity increases
Solution Approach 1:
The patent extracts and removes GMM and DNN models from the traditional acoustic model training pipeline, replacing them with a two-stage CTC model approach. By taking out these complex intermediate models, the system simplifies the device architecture while maintaining accuracy through direct CTC training that learns temporal alignments end-to-end without requiring separate GMM or DNN components.
Solution Approach 2:
The patent applies universality by using CTC models to perform multiple functions that previously required separate GMM and DNN models. The first CTC model generates approximate alignments and context-dependent phone inventories, while the second CTC model performs final acoustic modeling, making the CTC framework a universal solution that replaces multiple specialized models with a unified approach.
Data Source
AI summary
Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for training acoustic models and using the trained acoustic models. A connectionist temporal classification (CTC) acoustic model is accessed, the CTC acoustic model having been trained using a context-dependent state inventory generated from approximate phonetic alignments determined by another CTC acoustic model trained without fixed alignment targets. Audio data for a portion of an utterance is received. Input data corresponding to the received audio data is provided to the accessed CTC acoustic model. Data indicating a transcription for the utterance is generated based on output that the accessed CTC acoustic model produced in response to the input data. The data indicating the transcription is provided as output of an automated speech recognition service.