Architectures for deep neural network based acoustic models for automatic speech recognition
Bibliographic record
Abstract
In the recent years, Deep Neural Network-Hidden Markov Model (DNN-HMM) systems have overtaken the traditional Gaussian Mixture Model-Hidden Markov Model (GMM-HMM) systems as the state-of-the-art acoustic models in Automatic Speech Recognition (ASR). A lot of effort has been put in studying different deep learning architectures to improve ASR performance. However, most of these systems operate on the standard hand crafted spectral features which were used in the GMM-HMM systems. Recent research has shown that DNNs can operate directly on raw speech waveform input features. This thesismainly focuses on such network architectures which can operate directly on the speech waveform input features offering an alternative to standard signal processing. This thesis at first evaluates existing DNN based acoustic models trained on spectral features, analyzing various parameters affecting the performance of such networks. The ability of these DNN based systems to automatically acquire internal representation that are similar to mel-scale filter banks when fed with raw waveform input features is demonstrated. It is shown that increasing the size of the corpus helps in reducing the gap which exists between the Windowed Speech Waveform (WSW) DNNs and the Mel Frequency Spectral Coefficient (MFSC) DNNs performance. An investigation into efficient WSW DNN architectures is done and a proposed stacked bottleneck architecture is shown to reduce the gap that exists between the WSW DNN and the MFSC DNN by capturing improved spectral dynamic information. A combination of spectral features and waveformbased features is shown to improve the performance by providing additional information to the network. At last, redundancies associated with these systems are addressed and possible solutions are provided for reducing the size and complexity by using structured initialization and Singular Value Decomposition (SVD) based restructuring.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame machine prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. The Gemma side is a direct model label for every work in the frame, read from the title-only record. The Codex side is a classifier learned from the 10,348 direct Codex labels and calibrated to design-weighted sample rates; fields without enough sample support carry no Codex call. Candidate is the union of the two sides; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels.
Distilled classifier scores by category (both heads)
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.001 | 0.001 |
| Meta-epidemiology (narrow) | 0.001 | 0.000 |
| Meta-epidemiology (broad) | 0.000 | 0.001 |
| Bibliometrics | 0.000 | 0.000 |
| Science and technology studies | 0.000 | 0.000 |
| Scholarly communication | 0.001 | 0.001 |
| Open science | 0.001 | 0.001 |
| Research integrity | 0.001 | 0.002 |
| Insufficient payload (model declined to judge) | 0.005 | 0.002 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one source (direct Gemma or distilled Codex), not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".