Bibliographic record
Abstract
Voice based personal assistants are part of our daily lives. Their performance suffers in the presence of signal distortions, such as noise, reverberation, and competing speakers. This thesis addresses the problem of extracting the signal of interest in such challenging conditions by first localizing the target speaker and using the location to extract the target speech. In a first stage, a common situation is considered when the target speaker utters a known word or sentence such as the wake-up word of a distant-microphone voice command system. A method that exploits this text information in order to improve the speaker localization performance in the presence of competing speakers is proposed. The proposed solution uses a speech recognition system to align the wake-up word to the corrupted speech signal. A model spectrum representing the aligned phones is used to compute an identifier which is then used by a deep neural network to localize the target speaker. Results on simulated data show that the proposed method reduces the localization error rate compared to the classical GCC-PHAT method. Similar improvements are observed on real data. Given the estimated location of the target speaker, speech separation is performed in three stages. In the first stage, a simple delay-and-sum (DS) beamformer is used to enhance the signal impinging from that location which is then used in the second stage to estimate a time-frequency mask corresponding to the localized speaker using a neural network. This mask is used to compute the second-order statistics and to derive an adaptive beamformer in the third stage. A multichannel, multispeaker, reverberated, noisy dataset --- inspired from the famous WSJ0-2mix dataset --- was generated and the performance of the proposed pipeline was investigated in terms of the word error rate (WER). To make the system robust to localization errors, a Speaker LOcalization Guided Deflation (SLOGD) based approach which estimates the sources iteratively is proposed. At each iteration the location of one speaker is estimated and used to estimate a mask corresponding to that speaker. The estimated source is removed from the mixture before estimating the location and mask of the next source. The proposed method is shown to outperform Conv-TasNet. Finally, we consider the problem of explaining the robustness of neural networks used to compute time-frequency masks to mismatched noise conditions. We employ the so-called SHAP method to quantify the contribution of every time-frequency bin in the input signal to the estimated time-frequency mask. We define a metric that summarizes the SHAP values and show that it correlates with the WER achieved on separated speech. To the best of our knowledge, this is the first known study on neural network explainability in the context of speech separation.
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 | 0.001 |
| Meta-epidemiology (narrow) | 0.001 | 0.001 |
| Meta-epidemiology (broad) | 0.001 | 0.000 |
| Bibliometrics | 0.000 | 0.001 |
| Science and technology studies | 0.000 | 0.000 |
| Scholarly communication | 0.001 | 0.001 |
| Open science | 0.001 | 0.000 |
| Research integrity | 0.001 | 0.001 |
| Insufficient payload (model declined to judge) | 0.000 | 0.001 |
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".