Unsupervised Acoustic Model Training via Self-Generated Transcriptions
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current speech recognition systems face challenges in acquiring and utilizing large amounts of accurate training data, making it difficult and costly to adapt acoustic models to specific speakers or environments, and requiring manual transcription and speaker involvement.
Innovation Solution
An unsupervised acoustic modeling service that continuously monitors and updates acoustic models using confidence levels from speech recognition results, allowing for adaptation to individual speakers or groups without requiring known correct transcriptions, and can be implemented in distributed or local computing environments.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If supervised acoustic model training is used, then speech recognition accuracy can be improved through training data, but it becomes difficult, time-consuming, and expensive to acquire training data and correct transcriptions
Solution Approach 1:
The system performs self-training by automatically generating transcriptions from acoustic model hypotheses and using these to update the acoustic model without human intervention. The acoustic model generates hypotheses about speech content, these hypotheses become training transcriptions, and the model iteratively improves itself using its own outputs, eliminating the need for manual transcription by human speakers.
Solution Approach 2:
The system implements a feedback loop where the acoustic model generates speech recognition hypotheses, these hypotheses are used as training data to update the acoustic model, and the updated model produces improved hypotheses. This closed-loop feedback mechanism allows continuous self-improvement of the acoustic model using automatically generated training signals rather than manually collected data.
2Measurement precision
If supervised acoustic model training is used, then speech recognition accuracy can be improved, but the process becomes expensive requiring human speakers to both speak and transcribe
Solution Approach 1:
The system eliminates the need for expensive human transcription services by enabling the acoustic model to generate its own training transcriptions automatically. The model processes speech signals, generates hypotheses about the spoken content, and uses these hypotheses as training data, completely removing the requirement for human speakers to perform both speaking and transcription tasks.
Solution Approach 2:
The acoustic model hypotheses serve as an intermediary between the speech signal and the training data. Instead of requiring direct human transcription of speech, the system uses the acoustic model's hypothesis generation capability as a mediator to create training transcriptions, which are then used to update the model, replacing the expensive human transcription process.
3Measurement precision
If acoustic models are trained with extensive training data, then speech recognition accuracy improves, but the complexity of data acquisition and processing increases
Solution Approach 1:
The system eliminates complex data acquisition infrastructure by enabling the acoustic model to generate its own training data internally. Instead of requiring external systems for collecting, storing, and managing large datasets of speech and manual transcriptions, the model performs self-training using speech signals and its own hypothesis generation, dramatically simplifying the data acquisition architecture.
Solution Approach 2:
The invention extracts the transcription generation function from the external human operator and embeds it within the acoustic model system itself. By taking out the need for external transcription services and data management infrastructure, the system achieves continuous training with simplified architecture, where the acoustic model generates and uses its own training data internally.
Data Source
AI summary
An unsupervised acoustic modeling service for speech recognition is disclosed. A computing device may be present in a listening zone, such as a household or office, and may receive an audio signal that may include speech. In some instances, the speech is not directed to the computing device. Speech recognition results may be generated from the audio signal, using an acoustic model, and the results used to update the acoustic model. For example, the acoustic model may be updated to reflect a particular speaker's pronunciation of certain sound units, such as phonemes. By using speech that is not necessarily directed to the computing device, more data may be available for updating an acoustic model.