Bibliographic record
Abstract
Speech signals encode information across multiple, often overlapping, scales in time. I argue that, by designing acoustic models faithful to that multi-scale encoding, speech recognition accuracy can sometimes be improved. I do so by partitioning the acoustic model into a three-component pipeline and design interventions for each. In the first stage, feature extraction, the acoustic model transforms an audio signal into a filter-bank representation. We modify the traditional recipe, experimenting with different filters with better resolution in both frequency and scale. We also rearrange the spectral windowing equation to capture overlapping temporal information with higher resolution. With these modifications, we find significant gains in phone recognition performance on already-competitive models. For the second stage in which machine learning takes place, we propose a novel neural layer which is scale equivariant, similar to how a convolutional layer is shift equivariant. We pit those layers against one another in a phone recognition task and find that, although the former does outperform the latter, it uses twice the parameters to do so. When fixing the number of parameters, performance is commensurate between models. In the final stage, the acoustic model must transduce the learned representation into a transcription. We design a statistical framework for transducing time series events compatible with multi-scale phenomena by modelling event locations and types separately. Determining that marginalizing out locations as a latent variable would be combinatorically infeasible in its full generality, we propose either making simplifying assumptions until computation is tractable or performing stochastic estimation. We offer algorithms for stochastic estimation with nicer theoretical properties and better empirical performance (in terms of phone recognition) than some of the prior techniques we generalize.
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 distilled prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. Learned from the 10,348 direct Codex labels and 10,348 direct Gemma labels. Candidate is the union of thresholded teacher heads; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels or direct frontier model labels.
Codex and Gemma teacher scores by category
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.001 | 0.000 |
| Meta-epidemiology (narrow) | 0.000 | 0.000 |
| Meta-epidemiology (broad) | 0.001 | 0.000 |
| Bibliometrics | 0.001 | 0.002 |
| Science and technology studies | 0.000 | 0.000 |
| Scholarly communication | 0.000 | 0.000 |
| Open science | 0.001 | 0.000 |
| Research integrity | 0.000 | 0.001 |
| Insufficient payload (model declined to judge) | 0.000 | 0.004 |
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 teacher head, 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".