Robustesse des modèles neuronaux pour le traitement automatique de la parole
Bibliographic record
Abstract
Automatic Speech Recognition has become a popular tool with numerous applications; it also serves as the enabling technology for other speech tasks such as SLU or AST. In ASR, the speech signal is first uttered by the speaker, transmitted through the environment, before being captured by a recording device and processed by a machine learning model. However, each of those steps can be a source of variability and lead to transcription errors, impacting the system's robustness.In this thesis, we investigate various factors influencing the processing of speech by machines. Specifically, we focus on French pre-trained models fine-tuned for ASR. We begin by presenting our work on accent robustness. Through numerous experiments, we evaluate the resilience of the model to accent variation and explore different ways to close the gaps between accents. Particularly, we examine the effect of accent ratios in the training set. Besides, we introduce CEREALES, a new dataset of Quebec French.Going beyond accents, we are also interested in the impact of demographic variables on ASR performance. Using the Common Voice corpus, we highlight the model's biases and try to reduce them by using voluntarily biased training sets. Finally, the last chapter explores the issue of acoustic robustness using keyword recognition models: we show how ID and OOD performances are correlated and investigate how the training data or input features influence the robustness.
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 | 0.004 | 0.002 |
| Meta-epidemiology (narrow) | 0.000 | 0.000 |
| Meta-epidemiology (broad) | 0.000 | 0.000 |
| Bibliometrics | 0.000 | 0.001 |
| Science and technology studies | 0.001 | 0.000 |
| Scholarly communication | 0.001 | 0.000 |
| Open science | 0.002 | 0.000 |
| Research integrity | 0.000 | 0.000 |
| Insufficient payload (model declined to judge) | 0.000 | 0.000 |
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".