Large language models demonstrate proficiency on phonetic tasks, such as rhyming, without explicit phonetic or auditory grounding. In this work, we investigate how \verb|Llama-3.2-1B-Instruct| represents token-level phonetic information. Our results suggest that Llama uses a rich internal model of phonemes to complete phonetic tasks. We provide evidence for high-level organization of phoneme representations in its latent space. In doing so, we also identify a ``phoneme mover head" which promotes phonetic information during rhyming tasks. We visualize the output space of this head and find that, while notable differences exist, Llama learns a model of vowels similar to the standard IPA vowel chart for humans, despite receiving no direct supervision to do so.

Automatic Speech Recognition has advanced with self-supervised learning, enabling feature extraction directly from raw audio. In Wav2Vec, a CNN first transforms audio into feature vectors before the transformer processes them. This study examines CNN-extracted information for monophthong vowels using the TIMIT corpus. We compare MFCCs, MFCCs with formants, and CNN activations by training SVM classifiers for front-back vowel identification, assessing their classification accuracy to evaluate phonetic representation.

In recent years, the impact of self-supervised speech Transformers has extended to speaker-related applications. However, little research has explored how these models encode speaker information. In this work, we address this gap by identifying neurons in the feed-forward layers that are correlated with speaker information. Specifically, we analyze neurons associated with k-means clusters of self-supervised features and i-vectors. Our analysis reveals that these clusters correspond to broad phonetic and gender classes, making them suitable for identifying neurons that represent speakers. By protecting these neurons during pruning, we can significantly preserve performance on speaker-related task, demonstrating their crucial role in encoding speaker information.

We present PAST, a novel end-to-end framework that jointly models phonetic information alongside signal reconstruction, eliminating the need for external pretrained models. Unlike previous approaches that rely on pretrained self-supervised models, PAST employs supervised phonetic data, directly integrating domain knowledge into the tokenization process via auxiliary tasks. Additionally, we introduce a streamable, causal variant of PAST, enabling real-time speech applications. Results demonstrate that PAST surpasses existing evaluated baseline tokenizers across common evaluation metrics, including phonetic representation and speech reconstruction. Notably, PAST also achieves superior performance when serving as a speech representation for speech language models, further highlighting its effectiveness as a foundation for spoken language generation. To foster further research, we release the full implementation. For code, model checkpoints, and samples see: https://pages.cs.huji.ac.il/adiyoss-lab/PAST

As malicious users increasingly employ phonetic substitution to evade hate speech detection, researchers have investigated such strategies. However, two key challenges remain. First, existing studies have overlooked the Korean language, despite its vulnerability to phonetic perturbations due to its phonographic nature. Second, prior work has primarily focused on constructing datasets rather than developing architectural defenses. To address these challenges, we propose (1) PHonetic-Informed Substitution for Hangul (PHISH) that exploits the phonological characteristics of the Korean writing system, and (2) Mixed Encoding of Semantic-pHonetic features (MESH) that enhances the detector's robustness by incorporating phonetic information at the architectural level. Our experimental results demonstrate the effectiveness of our proposed methods on both perturbed and unperturbed datasets, suggesting that they not only improve detection performance but also reflect realistic adversarial behaviors employed by malicious users.

Generative error correction (GER) with large language models (LLMs) has emerged as an effective post-processing approach to improve automatic speech recognition (ASR) performance. However, it often struggles with rare or domain-specific words due to limited training data. Furthermore, existing LLM-based GER approaches primarily rely on textual information, neglecting phonetic cues, which leads to over-correction. To address these issues, we propose a novel LLM-based GER approach that targets rare words and incorporates phonetic information. First, we generate synthetic data to contain rare words for fine-tuning the GER model. Second, we integrate ASR's N-best hypotheses along with phonetic context to mitigate over-correction. Experimental results show that our method not only improves the correction of rare words but also reduces the WER and CER across both English and Japanese datasets.

Understanding the neural mechanisms underlying speech production is essential for both advancing cognitive neuroscience theory and developing practical communication technologies. In this study, we investigated magnetoencephalography signals to decode phones from brain activity during speech production and perception (passive listening and voice playback) tasks. Using a dataset comprising 17 participants, we performed pairwise phone classification, extending our analysis to 15 phonetic pairs. Multiple machine learning approaches, including regularized linear models and neural network architectures, were compared to determine their effectiveness in decoding phonetic information. Our results demonstrate significantly higher decoding accuracy during speech production (76.6%) compared to passive listening and playback modalities (~51%), emphasizing the richer neural information available during overt speech. Among the models, the Elastic Net classifier consistently outperformed more complex neural networks, highlighting the effectiveness of traditional regularization techniques when applied to limited and high-dimensional MEG datasets. Besides, analysis of specific brain frequency bands revealed that low-frequency oscillations, particularly Delta (0.2-3 Hz) and Theta (4-7 Hz), contributed the most substantially to decoding accuracy, suggesting that these bands encode critical speech production-related neural processes. Despite using advanced denoising methods, it remains unclear whether decoding solely reflects neural activity or if residual muscular or movement artifacts also contributed, indicating the need for further methodological refinement. Overall, our findings underline the critical importance of examining overt speech production paradigms, which, despite their complexity, offer opportunities to improve brain-computer interfaces to help individuals with severe speech impairments.

ABSTRACT Speaker verification systems are crucial for authenticating identity through voice. Traditionally, these systems focus on comparing feature vectors, overlooking the speech's content. The lexical content L determines the phonetic structure a measure of the frequency or duration of phonemes, as a P, which in turn determines the acoustics A. Thus, any production crucial cue in speaker verification. A novel Phoneme-Debiasing Attention of a signal actually represents the draws of all three variables. Framework (PDAF) is introduced, integrating with existing Content-agnostic verification systems, however, only consider the attention frameworks to mitigate biases caused by phonetic dominance. This approach paves the way for more accurate and reliable identity authentication through voice.

Recently, there have been efforts to encode the linguistic information of speech using a self-supervised framework for speech synthesis. However, predicting representations from surrounding representations can inadvertently entangle speaker information in the speech representation. This paper aims to remove speaker information by exploiting the structured nature of speech, composed of discrete units like phonemes with clear boundaries. A neural network predicts these boundaries, enabling variable-length pooling for event-based representation extraction instead of fixed-rate methods. The boundary predictor outputs a probability for the boundary between 0 and 1, making pooling soft. The model is trained to minimize the difference with the pooled representation of the data augmented by time-stretch and pitch-shift. To confirm that the learned representation includes contents information but is independent of speaker information, the model was evaluated with libri-light's phonetic ABX task and SUPERB's speaker identification task.

In this work, we study the features extracted by English self-supervised learning (SSL) models in cross-lingual contexts and propose a new metric to predict the quality of feature representations. Using automatic speech recognition (ASR) as a downstream task, we analyze the effect of model size, training objectives, and model architecture on the models' performance as a feature extractor for a set of topologically diverse corpora. We develop a novel metric, the Phonetic-Syntax Ratio (PSR), to measure the phonetic and synthetic information in the extracted representations using deep generalized canonical correlation analysis. Results show the contrastive loss in the wav2vec2.0 objective facilitates more effective cross-lingual feature extraction. There is a positive correlation between PSR scores and ASR performance, suggesting that phonetic information extracted by monolingual SSL models can be used for downstream tasks in cross-lingual settings. The proposed metric is an effective indicator of the quality of the representations and can be useful for model selection.