Search

We introduce PokerBench - a benchmark for evaluating the poker-playing abilities of large language models (LLMs). As LLMs excel in traditional NLP tasks, their application to complex, strategic games like poker poses a new challenge. Poker, an incomplete information game, demands a multitude of skills such as mathematics, reasoning, planning, strategy, and a deep understanding of game theory and human psychology. This makes Poker the ideal next frontier for large language models. PokerBench consists of a comprehensive compilation of 11,000 most important scenarios, split between pre-flop and post-flop play, developed in collaboration with trained poker players. We evaluate prominent models including GPT-4, ChatGPT 3.5, and various Llama and Gemma series models, finding that all state-of-the-art LLMs underperform in playing optimal poker. However, after fine-tuning, these models show marked improvements. We validate PokerBench by having models with different scores compete with each other, demonstrating that higher scores on PokerBench lead to higher win rates in actual poker games. Through gameplay between our fine-tuned model and GPT-4, we also identify limitations of simple supervised fine-tuning for learning optimal playing strategy, suggesting the need for more advanced methodologies for effectively training language models to excel in games. PokerBench thus presents a unique benchmark for a quick and reliable evaluation of the poker-playing ability of LLMs as well as a comprehensive benchmark to study the progress of LLMs in complex game-playing scenarios. The dataset and code will be made available at: \url{https://github.com/pokerllm/pokerbench}., Comment: AAAI 2025

The amount of articulatory data available for training deep learning models is much less compared to acoustic speech data. In order to improve articulatory-to-acoustic synthesis performance in these low-resource settings, we propose a multimodal pre-training framework. On single-speaker speech synthesis tasks from real-time magnetic resonance imaging and surface electromyography inputs, the intelligibility of synthesized outputs improves noticeably. For example, compared to prior work, utilizing our proposed transfer learning methods improves the MRI-to-speech performance by 36% word error rate. In addition to these intelligibility results, our multimodal pre-trained models consistently outperform unimodal baselines on three objective and subjective synthesis quality metrics.

Speech is a hierarchical collection of text, prosody, emotions, dysfluencies, etc. Automatic transcription of speech that goes beyond text (words) is an underexplored problem. We focus on transcribing speech along with non-fluencies (dysfluencies). The current state-of-the-art pipeline SSDM suffers from complex architecture design, training complexity, and significant shortcomings in the local sequence aligner, and it does not explore in-context learning capacity. In this work, we propose SSDM 2.0, which tackles those shortcomings via four main contributions: (1) We propose a novel \textit{neural articulatory flow} to derive highly scalable speech representations. (2) We developed a \textit{full-stack connectionist subsequence aligner} that captures all types of dysfluencies. (3) We introduced a mispronunciation prompt pipeline and consistency learning module into LLM to leverage dysfluency \textit{in-context pronunciation learning} abilities. (4) We curated Libri-Dys and open-sourced the current largest-scale co-dysfluency corpus, \textit{Libri-Co-Dys}, for future research endeavors. In clinical experiments on pathological speech transcription, we tested SSDM 2.0 using nfvPPA corpus primarily characterized by \textit{articulatory dysfluencies}. Overall, SSDM 2.0 outperforms SSDM and all other dysfluency transcription models by a large margin. See our project demo page at \url{https://berkeley-speech-group.github.io/SSDM2.0/}.

Syllables are compositional units of spoken language that play a crucial role in human speech perception and production. However, current neural speech representations lack structure, resulting in dense token sequences that are costly to process. To bridge this gap, we propose a new model, Sylber, that produces speech representations with clean and robust syllabic structure. Specifically, we propose a self-supervised model that regresses features on syllabic segments distilled from a teacher model which is an exponential moving average of the model in training. This results in a highly structured representation of speech features, offering three key benefits: 1) a fast, linear-time syllable segmentation algorithm, 2) efficient syllabic tokenization with an average of 4.27 tokens per second, and 3) syllabic units better suited for lexical and syntactic understanding. We also train token-to-speech generative models with our syllabic units and show that fully intelligible speech can be reconstructed from these tokens. Lastly, we observe that categorical perception, a linguistic phenomenon of speech perception, emerges naturally in our model, making the embedding space more categorical and sparse than previous self-supervised learning approaches. Together, we present a novel self-supervised approach for representing speech as syllables, with significant potential for efficient speech tokenization and spoken language modeling.

While the language modeling objective has been shown to be deeply connected with compression, it is surprising that modern LLMs are not employed in practical text compression systems. In this paper, we provide an in-depth analysis of neural network and transformer-based compression techniques to answer this question. We compare traditional text compression systems with neural network and LLM-based text compression methods. Although LLM-based systems significantly outperform conventional compression methods, they are highly impractical. Specifically, LLMZip, a recent text compression system using Llama3-8B requires 9.5 days to compress just 10 MB of text, although with huge improvements in compression ratios. To overcome this, we present FineZip - a novel LLM-based text compression system that combines ideas of online memorization and dynamic context to reduce the compression time immensely. FineZip can compress the above corpus in approximately 4 hours compared to 9.5 days, a 54 times improvement over LLMZip and comparable performance. FineZip outperforms traditional algorithmic compression methods with a large margin, improving compression ratios by approximately 50\%. With this work, we take the first step towards making lossless text compression with LLMs a reality. While FineZip presents a significant step in that direction, LLMs are still not a viable solution for large-scale text compression. We hope our work paves the way for future research and innovation to solve this problem.

Speech sounds convey a great deal of information about the scenes, resulting in a variety of effects ranging from reverberation to additional ambient sounds. In this paper, we manipulate input speech to sound as though it was recorded within a different scene, given an audio-visual conditional example recorded from that scene. Our model learns through self-supervision, taking advantage of the fact that natural video contains recurring sound events and textures. We extract an audio clip from a video and apply speech enhancement. We then train a latent diffusion model to recover the original speech, using another audio-visual clip taken from elsewhere in the video as a conditional hint. Through this process, the model learns to transfer the conditional example's sound properties to the input speech. We show that our model can be successfully trained using unlabeled, in-the-wild videos, and that an additional visual signal can improve its sound prediction abilities. Please see our project webpage for video results: https://tinglok.netlify.app/files/avsoundscape/, Comment: ECCV 2024

Speech dysfluency modeling is a task to detect dysfluencies in speech, such as repetition, block, insertion, replacement, and deletion. Most recent advancements treat this problem as a time-based object detection problem. In this work, we revisit this problem from a new perspective: tokenizing dysfluencies and modeling the detection problem as a token-based automatic speech recognition (ASR) problem. We propose rule-based speech and text dysfluency simulators and develop VCTK-token, and then develop a Whisper-like seq2seq architecture to build a new benchmark with decent performance. We also systematically compare our proposed token-based methods with time-based methods, and propose a unified benchmark to facilitate future research endeavors. We open-source these resources for the broader scientific community. The project page is available at https://rorizzz.github.io/

Layer normalization is a pivotal step in the transformer architecture. This paper delves into the less explored geometric implications of this process, examining how LayerNorm influences the norm and orientation of hidden vectors in the representation space. We show that the definition of LayerNorm is innately linked to the uniform vector, defined as $\boldsymbol{1} = [1, 1, 1, 1, \cdots, 1]^T \in \mathbb{R}^d$. We then show that the standardization step in LayerNorm can be understood in three simple steps: (i) remove the component of a vector along the uniform vector, (ii) normalize the remaining vector, and (iii) scale the resultant vector by $\sqrt{d}$, where $d$ is the dimensionality of the representation space. We also introduce the property of "irreversibility" for LayerNorm, where we show that the information lost during the normalization process cannot be recovered. In other words, unlike batch normalization, LayerNorm cannot learn an identity transform. While we present possible arguments for removing the component along the uniform vector, the choice of removing this component seems arbitrary and not well motivated by the original authors. To evaluate the usefulness of this step, we compare the hidden representations of LayerNorm-based LLMs with models trained using RMSNorm and show that all LLMs naturally align representations orthogonal to the uniform vector, presenting the first mechanistic evidence that removing the component along the uniform vector in LayerNorm is a redundant step. Our findings support the use of RMSNorm over LayerNorm as it is not only more computationally efficient with comparable downstream performance, but also learns a similar distribution of hidden representations that operate orthogonal to the uniform vector.

Current de-facto dysfluency modeling methods utilize template matching algorithms which are not generalizable to out-of-domain real-world dysfluencies across languages, and are not scalable with increasing amounts of training data. To handle these problems, we propose Stutter-Solver: an end-to-end framework that detects dysfluency with accurate type and time transcription, inspired by the YOLO object detection algorithm. Stutter-Solver can handle co-dysfluencies and is a natural multi-lingual dysfluency detector. To leverage scalability and boost performance, we also introduce three novel dysfluency corpora: VCTK-Pro, VCTK-Art, and AISHELL3-Pro, simulating natural spoken dysfluencies including repetition, block, missing, replacement, and prolongation through articulatory-encodec and TTS-based methods. Our approach achieves state-of-the-art performance on all available dysfluency corpora. Code and datasets are open-sourced at https://github.com/eureka235/Stutter-Solver, Comment: IEEE Spoken Language Technology Workshop 2024

Articulatory trajectories like electromagnetic articulography (EMA) provide a low-dimensional representation of the vocal tract filter and have been used as natural, grounded features for speech synthesis. Differentiable digital signal processing (DDSP) is a parameter-efficient framework for audio synthesis. Therefore, integrating low-dimensional EMA features with DDSP can significantly enhance the computational efficiency of speech synthesis. In this paper, we propose a fast, high-quality, and parameter-efficient DDSP articulatory vocoder that can synthesize speech from EMA, F0, and loudness. We incorporate several techniques to solve the harmonics / noise imbalance problem, and add a multi-resolution adversarial loss for better synthesis quality. Our model achieves a transcription word error rate (WER) of 6.67% and a mean opinion score (MOS) of 3.74, with an improvement of 1.63% and 0.16 compared to the state-of-the-art (SOTA) baseline. Our DDSP vocoder is 4.9x faster than the baseline on CPU during inference, and can generate speech of comparable quality with only 0.4M parameters, in contrast to the 9M parameters required by the SOTA., Comment: accepted for Spoken Language Technology Workshop 2024

Recent large language models (LLMs) have enabled the development of advanced agentic systems that can integrate various tools and APIs to fulfill user queries through function calling. However, the deployment of these LLMs on the edge has not been explored since they typically require cloud-based infrastructure due to their substantial model size and computational demands. To this end, we present TinyAgent, an end-to-end framework for training and deploying task-specific small language model agents capable of function calling for driving agentic systems at the edge. We first show how to enable accurate function calling for open-source models via the LLMCompiler framework. We then systematically curate a high-quality dataset for function calling, which we use to fine-tune two small language models, TinyAgent-1.1B and 7B. For efficient inference, we introduce a novel tool retrieval method to reduce the input prompt length and utilize quantization to further accelerate the inference speed. As a driving application, we demonstrate a local Siri-like system for Apple's MacBook that can execute user commands through text or voice input. Our results show that our models can achieve, and even surpass, the function-calling capabilities of larger models like GPT-4-Turbo, while being fully deployed at the edge. We open-source our dataset, models, and installable package and provide a demo video for our MacBook assistant agent., Comment: EMNLP 2024 Demo

Speech dysfluency modeling is the core module for spoken language learning, and speech therapy. However, there are three challenges. First, current state-of-the-art solutions\cite{lian2023unconstrained-udm, lian-anumanchipalli-2024-towards-hudm} suffer from poor scalability. Second, there is a lack of a large-scale dysfluency corpus. Third, there is not an effective learning framework. In this paper, we propose \textit{SSDM: Scalable Speech Dysfluency Modeling}, which (1) adopts articulatory gestures as scalable forced alignment; (2) introduces connectionist subsequence aligner (CSA) to achieve dysfluency alignment; (3) introduces a large-scale simulated dysfluency corpus called Libri-Dys; and (4) develops an end-to-end system by leveraging the power of large language models (LLMs). We expect SSDM to serve as a standard in the area of dysfluency modeling. Demo is available at \url{https://berkeley-speech-group.github.io/SSDM/}., Comment: 2024 NeurIPS

Dysfluent speech detection is the bottleneck for disordered speech analysis and spoken language learning. Current state-of-the-art models are governed by rule-based systems which lack efficiency and robustness, and are sensitive to template design. In this paper, we propose YOLO-Stutter: a first end-to-end method that detects dysfluencies in a time-accurate manner. YOLO-Stutter takes imperfect speech-text alignment as input, followed by a spatial feature aggregator, and a temporal dependency extractor to perform region-wise boundary and class predictions. We also introduce two dysfluency corpus, VCTK-Stutter and VCTK-TTS, that simulate natural spoken dysfluencies including repetition, block, missing, replacement, and prolongation. Our end-to-end method achieves state-of-the-art performance with a minimum number of trainable parameters for on both simulated data and real aphasia speech. Code and datasets are open-sourced at https://github.com/rorizzz/YOLO-Stutter, Comment: Interspeech 2024

Electrodes for decoding speech from electromyography (EMG) are typically placed on the face, requiring adhesives that are inconvenient and skin-irritating if used regularly. We explore a different device form factor, where dry electrodes are placed around the neck instead. 11-word, multi-speaker voiced EMG classifiers trained on data recorded with this device achieve 92.7% accuracy. Ablation studies reveal the importance of having more than two electrodes on the neck, and phonological analyses reveal similar classification confusions between neck-only and neck-and-face form factors. Finally, speech-EMG correlation experiments demonstrate a linear relationship between many EMG spectrogram frequency bins and self-supervised speech representation dimensions.

As experts in voice modification, trans-feminine gender-affirming voice teachers have unique perspectives on voice that confound current understandings of speaker identity. To demonstrate this, we present the Versatile Voice Dataset (VVD), a collection of three speakers modifying their voices along gendered axes. The VVD illustrates that current approaches in speaker modeling, based on categorical notions of gender and a static understanding of vocal texture, fail to account for the flexibility of the vocal tract. Utilizing publicly-available speaker embeddings, we demonstrate that gender classification systems are highly sensitive to voice modification, and speaker verification systems fail to identify voices as coming from the same speaker as voice modification becomes more drastic. As one path towards moving beyond categorical and static notions of speaker identity, we propose modeling individual qualities of vocal texture such as pitch, resonance, and weight.

Accurate modeling of the vocal tract is necessary to construct articulatory representations for interpretable speech processing and linguistics. However, vocal tract modeling is challenging because many internal articulators are occluded from external motion capture technologies. Real-time magnetic resonance imaging (RT-MRI) allows measuring precise movements of internal articulators during speech, but annotated datasets of MRI are limited in size due to time-consuming and computationally expensive labeling methods. We first present a deep labeling strategy for the RT-MRI video using a vision-only segmentation approach. We then introduce a multimodal algorithm using audio to improve segmentation of vocal articulators. Together, we set a new benchmark for vocal tract modeling in MRI video segmentation and use this to release labels for a 75-speaker RT-MRI dataset, increasing the amount of labeled public RT-MRI data of the vocal tract by over a factor of 9. The code and dataset labels can be found at \url{rishiraij.github.io/multimodal-mri-avatar/}., Comment: Interspeech 2024

Vocal tract articulation is a natural, grounded control space of speech production. The spatiotemporal coordination of articulators combined with the vocal source shapes intelligible speech sounds to enable effective spoken communication. Based on this physiological grounding of speech, we propose a new framework of neural encoding-decoding of speech -- Speech Articulatory Coding (SPARC). SPARC comprises an articulatory analysis model that infers articulatory features from speech audio, and an articulatory synthesis model that synthesizes speech audio from articulatory features. The articulatory features are kinematic traces of vocal tract articulators and source features, which are intuitively interpretable and controllable, being the actual physical interface of speech production. An additional speaker identity encoder is jointly trained with the articulatory synthesizer to inform the voice texture of individual speakers. By training on large-scale speech data, we achieve a fully intelligible, high-quality articulatory synthesizer that generalizes to unseen speakers. Furthermore, the speaker embedding is effectively disentangled from articulations, which enables accent-perserving zero-shot voice conversion. To the best of our knowledge, this is the first demonstration of universal, high-performance articulatory inference and synthesis, suggesting the proposed framework as a powerful coding system of speech.

This study presents a targeted model editing analysis focused on the latest large language model, Llama-3. We explore the efficacy of popular model editing techniques - ROME, MEMIT, and EMMET, which are designed for precise layer interventions. We identify the most effective layers for targeted edits through an evaluation that encompasses up to 4096 edits across three distinct strategies: sequential editing, batch editing, and a hybrid approach we call as sequential-batch editing. Our findings indicate that increasing edit batch-sizes may degrade model performance more significantly than using smaller edit batches sequentially for equal number of edits. With this, we argue that sequential model editing is an important component for scaling model editing methods and future research should focus on methods that combine both batched and sequential editing. This observation suggests a potential limitation in current model editing methods which push towards bigger edit batch sizes, and we hope it paves way for future investigations into optimizing batch sizes and model editing performance.

Pretrained large language models (LLMs) are currently state-of-the-art for solving the vast majority of natural language processing tasks. While many real-world applications still require fine-tuning to reach satisfactory levels of performance, many of them are in the low-data regime, making fine-tuning challenging. To address this, we propose LLM2LLM, a targeted and iterative data augmentation strategy that uses a teacher LLM to enhance a small seed dataset by augmenting additional data that can be used for fine-tuning on a specific task. LLM2LLM (1) fine-tunes a baseline student LLM on the initial seed data, (2) evaluates and extracts data points that the model gets wrong, and (3) uses a teacher LLM to generate synthetic data based on these incorrect data points, which are then added back into the training data. This approach amplifies the signal from incorrectly predicted data points by the LLM during training and reintegrates them into the dataset to focus on more challenging examples for the LLM. Our results show that LLM2LLM significantly enhances the performance of LLMs in the low-data regime, outperforming both traditional fine-tuning and other data augmentation baselines. LLM2LLM reduces the dependence on labor-intensive data curation and paves the way for more scalable and performant LLM solutions, allowing us to tackle data-constrained domains and tasks. We achieve improvements up to 24.2% on the GSM8K dataset, 32.6% on CaseHOLD, 32.0% on SNIPS, 52.6% on TREC and 39.8% on SST-2 over regular fine-tuning in the low-data regime using a Llama-2-7B student model. Our code is available at https://github.com/SqueezeAILab/LLM2LLM ., Comment: ACL 2024

ROME and MEMIT are largely believed to be two different model editing algorithms, with the major difference between them being the ability to perform batched edits. In this paper, we unify these two algorithms under a single conceptual umbrella, optimizing for the same goal, which we call the preservation-memorization objective. ROME uses an equality constraint to optimize this objective to perform one edit at a time, whereas MEMIT employs a more flexible least-square constraint that allows for batched edits. We generalize ROME and enable batched editing with equality constraint in the form of EMMET - an Equality-constrained Mass Model Editing algorithm for Transformers, a new batched memory-editing algorithm. EMMET can perform batched-edits up to a batch-size of 10,000, with very similar performance to MEMIT across multiple dimensions. With the introduction of EMMET, we truly unify ROME and MEMIT and show that both algorithms are equivalent in terms of their optimization objective, their abilities (singular and batched editing), their model editing performance and their limitations., Comment: EMNLP 2024 Findings

Recent work using Rank-One Model Editing (ROME), a popular model editing method, has shown that there are certain facts that the algorithm is unable to edit without breaking the model. Such edits have previously been called disabling edits. These disabling edits cause immediate model collapse and limits the use of ROME for sequential editing. In this paper, we show that disabling edits are an artifact of irregularities in the implementation of ROME. With this paper, we provide a more stable implementation ROME, which we call r-ROME and show that model collapse is no longer observed when making large scale sequential edits with r-ROME, while further improving generalization and locality of model editing compared to the original implementation of ROME. We also provide a detailed mathematical explanation of the reason behind disabling edits., Comment: EMNLP 2024 (Main)

As Large Language Models (LLMs) are integrated with human daily applications rapidly, many societal and ethical concerns are raised regarding the behavior of LLMs. One of the ways to comprehend LLMs' behavior is to analyze their personalities. Many recent studies quantify LLMs' personalities using self-assessment tests that are created for humans. Yet many critiques question the applicability and reliability of these self-assessment tests when applied to LLMs. In this paper, we investigate LLM personalities using an alternate personality measurement method, which we refer to as the external evaluation method, where instead of prompting LLMs with multiple-choice questions in the Likert scale, we evaluate LLMs' personalities by analyzing their responses toward open-ended situational questions using an external machine learning model. We first fine-tuned a Llama2-7B model as the MBTI personality predictor that outperforms the state-of-the-art models as the tool to analyze LLMs' responses. Then, we prompt the LLMs with situational questions and ask them to generate Twitter posts and comments, respectively, in order to assess their personalities when playing two different roles. Using the external personality evaluation method, we identify that the obtained personality types for LLMs are significantly different when generating posts versus comments, whereas humans show a consistent personality profile in these two different situations. This shows that LLMs can exhibit different personalities based on different scenarios, thus highlighting a fundamental difference between personality in LLMs and humans. With our work, we call for a re-evaluation of personality definition and measurement in LLMs.

Speech disfluency modeling is the bottleneck for both speech therapy and language learning. However, there is no effective AI solution to systematically tackle this problem. We solidify the concept of disfluent speech and disfluent speech modeling. We then present Hierarchical Unconstrained Disfluency Modeling (H-UDM) approach, the hierarchical extension of UDM that addresses both disfluency transcription and detection to eliminate the need for extensive manual annotation. Our experimental findings serve as clear evidence of the effectiveness and reliability of the methods we have introduced, encompassing both transcription and detection tasks., Comment: 2024 EACL. Hierarchical extension of our previous workshop paper arXiv:2312.12810

Editing knowledge in large language models is an attractive capability to have which allows us to correct incorrectly learnt facts during pre-training, as well as update the model with an ever-growing list of new facts. While existing model editing techniques have shown promise, they are usually evaluated using metrics for reliability, specificity and generalization over one or few edits. We argue that for model editing to have practical utility, we must be able to make multiple edits to the same model. With this in mind, we evaluate the current model editing methods at scale, focusing on two state of the art methods: ROME and MEMIT. We find that as the model is edited sequentially with multiple facts, it continually forgets previously edited facts and the ability to perform downstream tasks. This forgetting happens in two phases -- an initial gradual but progressive forgetting phase followed by abrupt or catastrophic forgetting phase. Both gradual and catastrophic forgetting limit the usefulness of model editing methods at scale -- the former making model editing less effective as multiple edits are made to the model while the latter caps the scalability of such model editing methods. Our analysis also highlights other key limitations of ROME and MEMIT at scale. With our work, we push for the development and evaluation of model editing methods keeping scalability in mind., Comment: ACL 2024 Findings

Dysfluent speech modeling requires time-accurate and silence-aware transcription at both the word-level and phonetic-level. However, current research in dysfluency modeling primarily focuses on either transcription or detection, and the performance of each aspect remains limited. In this work, we present an unconstrained dysfluency modeling (UDM) approach that addresses both transcription and detection in an automatic and hierarchical manner. UDM eliminates the need for extensive manual annotation by providing a comprehensive solution. Furthermore, we introduce a simulated dysfluent dataset called VCTK++ to enhance the capabilities of UDM in phonetic transcription. Our experimental results demonstrate the effectiveness and robustness of our proposed methods in both transcription and detection tasks., Comment: 2023 ASRU

Perceptual modification of voice is an elusive goal. While non-experts can modify an image or sentence perceptually with available tools, it is not clear how to similarly modify speech along perceptual axes. Voice conversion does make it possible to convert one voice to another, but these modifications are handled by black box models, and the specifics of what perceptual qualities to modify and how to modify them are unclear. Towards allowing greater perceptual control over voice, we introduce PerMod, a conditional latent diffusion model that takes in an input voice and a perceptual qualities vector, and produces a voice with the matching perceptual qualities. Unlike prior work, PerMod generates a new voice corresponding to specific perceptual modifications. Evaluating perceptual quality vectors with RMSE from both human and predicted labels, we demonstrate that PerMod produces voices with the desired perceptual qualities for typical voices, but performs poorly on atypical voices.

Streaming speech-to-avatar synthesis creates real-time animations for a virtual character from audio data. Accurate avatar representations of speech are important for the visualization of sound in linguistics, phonetics, and phonology, visual feedback to assist second language acquisition, and virtual embodiment for paralyzed patients. Previous works have highlighted the capability of deep articulatory inversion to perform high-quality avatar animation using electromagnetic articulography (EMA) features. However, these models focus on offline avatar synthesis with recordings rather than real-time audio, which is necessary for live avatar visualization or embodiment. To address this issue, we propose a method using articulatory inversion for streaming high quality facial and inner-mouth avatar animation from real-time audio. Our approach achieves 130ms average streaming latency for every 0.1 seconds of audio with a 0.792 correlation with ground truth articulations. Finally, we show generated mouth and tongue animations to demonstrate the efficacy of our methodology., Comment: Submitted to ICASSP 2024

Data-driven unit discovery in self-supervised learning (SSL) of speech has embarked on a new era of spoken language processing. Yet, the discovered units often remain in phonetic space and the units beyond phonemes are largely underexplored. Here, we demonstrate that a syllabic organization emerges in learning sentence-level representation of speech. In particular, we adopt "self-distillation" objective to fine-tune the pretrained HuBERT with an aggregator token that summarizes the entire sentence. Without any supervision, the resulting model draws definite boundaries in speech, and the representations across frames exhibit salient syllabic structures. We demonstrate that this emergent structure largely corresponds to the ground truth syllables. Furthermore, we propose a new benchmark task, Spoken Speech ABX, for evaluating sentence-level representation of speech. When compared to previous models, our model outperforms in both unsupervised syllable discovery and learning sentence-level representation. Together, we demonstrate that the self-distillation of HuBERT gives rise to syllabic organization without relying on external labels or modalities, and potentially provides novel data-driven units for spoken language modeling.

Self-Supervised Learning (SSL) based models of speech have shown remarkable performance on a range of downstream tasks. These state-of-the-art models have remained blackboxes, but many recent studies have begun "probing" models like HuBERT, to correlate their internal representations to different aspects of speech. In this paper, we show "inference of articulatory kinematics" as fundamental property of SSL models, i.e., the ability of these models to transform acoustics into the causal articulatory dynamics underlying the speech signal. We also show that this abstraction is largely overlapping across the language of the data used to train the model, with preference to the language with similar phonological system. Furthermore, we show that with simple affine transformations, Acoustic-to-Articulatory inversion (AAI) is transferrable across speakers, even across genders, languages, and dialects, showing the generalizability of this property. Together, these results shed new light on the internals of SSL models that are critical to their superior performance, and open up new avenues into language-agnostic universal models for speech engineering, that are interpretable and grounded in speech science.

Unlike other data modalities such as text and vision, speech does not lend itself to easy interpretation. While lay people can understand how to describe an image or sentence via perception, non-expert descriptions of speech often end at high-level demographic information, such as gender or age. In this paper, we propose a possible interpretable representation of speaker identity based on perceptual voice qualities (PQs). By adding gendered PQs to the pathology-focused Consensus Auditory-Perceptual Evaluation of Voice (CAPE-V) protocol, our PQ-based approach provides a perceptual latent space of the character of adult voices that is an intermediary of abstraction between high-level demographics and low-level acoustic, physical, or learned representations. Contrary to prior belief, we demonstrate that these PQs are hearable by ensembles of non-experts, and further demonstrate that the information encoded in a PQ-based representation is predictable by various speech representations.

Vocoder models have recently achieved substantial progress in generating authentic audio comparable to human quality while significantly reducing memory requirement and inference time. However, these data-hungry generative models require large-scale audio data for learning good representations. In this paper, we apply contrastive learning methods in training the vocoder to improve the perceptual quality of the vocoder without modifying its architecture or adding more data. We design an auxiliary task with mel-spectrogram contrastive learning to enhance the utterance-level quality of the vocoder model under data-limited conditions. We also extend the task to include waveforms to improve the multi-modality comprehension of the model and address the discriminator overfitting problem. We optimize the additional task simultaneously with GAN training objectives. Our results show that the tasks improve model performance substantially in data-limited settings.

As large language models (LLM) evolve in their capabilities, various recent studies have tried to quantify their behavior using psychological tools created to study human behavior. One such example is the measurement of "personality" of LLMs using self-assessment personality tests developed to measure human personality. Yet almost none of these works verify the applicability of these tests on LLMs. In this paper, we analyze the reliability of LLM personality scores obtained from self-assessment personality tests using two simple experiments. We first introduce the property of prompt sensitivity, where three semantically equivalent prompts representing three intuitive ways of administering self-assessment tests on LLMs are used to measure the personality of the same LLM. We find that all three prompts lead to very different personality scores, a difference that is statistically significant for all traits in a large majority of scenarios. We then introduce the property of option-order symmetry for personality measurement of LLMs. Since most of the self-assessment tests exist in the form of multiple choice question (MCQ) questions, we argue that the scores should also be robust to not just the prompt template but also the order in which the options are presented. This test unsurprisingly reveals that the self-assessment test scores are not robust to the order of the options. These simple tests, done on ChatGPT and three Llama2 models of different sizes, show that self-assessment personality tests created for humans are unreliable measures of personality in LLMs.

Humans encode information into sounds by controlling articulators and decode information from sounds using the auditory apparatus. This paper introduces CiwaGAN, a model of human spoken language acquisition that combines unsupervised articulatory modeling with an unsupervised model of information exchange through the auditory modality. While prior research includes unsupervised articulatory modeling and information exchange separately, our model is the first to combine the two components. The paper also proposes an improved articulatory model with more interpretable internal representations. The proposed CiwaGAN model is the most realistic approximation of human spoken language acquisition using deep learning. As such, it is useful for cognitively plausible simulations of the human speech act.

Understanding the neural implementation of complex human behaviors is one of the major goals in neuroscience. To this end, it is crucial to find a true representation of the neural data, which is challenging due to the high complexity of behaviors and the low signal-to-ratio (SNR) of the signals. Here, we propose a novel unsupervised learning framework, Neural Latent Aligner (NLA), to find well-constrained, behaviorally relevant neural representations of complex behaviors. The key idea is to align representations across repeated trials to learn cross-trial consistent information. Furthermore, we propose a novel, fully differentiable time warping model (TWM) to resolve the temporal misalignment of trials. When applied to intracranial electrocorticography (ECoG) of natural speaking, our model learns better representations for decoding behaviors than the baseline models, especially in lower dimensional space. The TWM is empirically validated by measuring behavioral coherence between aligned trials. The proposed framework learns more cross-trial consistent representations than the baselines, and when visualized, the manifold reveals shared neural trajectories across trials., Comment: Accepted at ICML 2023

In this paper, we study articulatory synthesis, a speech synthesis method using human vocal tract information that offers a way to develop efficient, generalizable and interpretable synthesizers. While recent advances have enabled intelligible articulatory synthesis using electromagnetic articulography (EMA), these methods lack critical articulatory information like excitation and nasality, limiting generalization capabilities. To bridge this gap, we propose an alternative MRI-based feature set that covers a much more extensive articulatory space than EMA. We also introduce normalization and denoising procedures to enhance the generalizability of deep learning methods trained on MRI data. Moreover, we propose an MRI-to-speech model that improves both computational efficiency and speech fidelity. Finally, through a series of ablations, we show that the proposed MRI representation is more comprehensive than EMA and identify the most suitable MRI feature subset for articulatory synthesis.

To build speech processing methods that can handle speech as naturally as humans, researchers have explored multiple ways of building an invertible mapping from speech to an interpretable space. The articulatory space is a promising inversion target, since this space captures the mechanics of speech production. To this end, we build an acoustic-to-articulatory inversion (AAI) model that leverages self-supervision to generalize to unseen speakers. Our approach obtains 0.784 correlation on an electromagnetic articulography (EMA) dataset, improving the state-of-the-art by 12.5\%. Additionally, we show the interpretability of these representations through directly comparing the behavior of estimated representations with speech production behavior. Finally, we propose a resynthesis-based AAI evaluation metric that does not rely on articulatory labels, demonstrating its efficacy with an 18-speaker dataset.

Articulatory representation learning is the fundamental research in modeling neural speech production system. Our previous work has established a deep paradigm to decompose the articulatory kinematics data into gestures, which explicitly model the phonological and linguistic structure encoded with human speech production mechanism, and corresponding gestural scores. We continue with this line of work by raising two concerns: (1) The articulators are entangled together in the original algorithm such that some of the articulators do not leverage effective moving patterns, which limits the interpretability of both gestures and gestural scores; (2) The EMA data is sparsely sampled from articulators, which limits the intelligibility of learned representations. In this work, we propose a novel articulatory representation decomposition algorithm that takes the advantage of guided factor analysis to derive the articulatory-specific factors and factor scores. A neural convolutive matrix factorization algorithm is then employed on the factor scores to derive the new gestures and gestural scores. We experiment with the rtMRI corpus that captures the fine-grained vocal tract contours. Both subjective and objective evaluation results suggest that the newly proposed system delivers the articulatory representations that are intelligible, generalizable, efficient and interpretable., Comment: Accepted to 2023 ICASSP. Camera Ready

Estimation of fundamental frequency (F0) in voiced segments of speech signals, also known as pitch tracking, plays a crucial role in pitch synchronous speech analysis, speech synthesis, and speech manipulation. In this paper, we capitalize on the high time and frequency resolution of the pseudo Wigner-Ville distribution (PWVD) and propose a new PWVD-based pitch estimation method. We devise an efficient algorithm to compute PWVD faster and use cepstrum-based pre-filtering to avoid cross-term interference. Evaluating our approach on a database with speech and electroglottograph (EGG) recordings yields a state-of-the-art mean absolute error (MAE) of around 4Hz. Our approach is also effective at voiced/unvoiced classification and handling sudden frequency changes.

Generative deep neural networks are widely used for speech synthesis, but most existing models directly generate waveforms or spectral outputs. Humans, however, produce speech by controlling articulators, which results in the production of speech sounds through physical properties of sound propagation. We introduce the Articulatory Generator to the Generative Adversarial Network paradigm, a new unsupervised generative model of speech production/synthesis. The Articulatory Generator more closely mimics human speech production by learning to generate articulatory representations (electromagnetic articulography or EMA) in a fully unsupervised manner. A separate pre-trained physical model (ema2wav) then transforms the generated EMA representations to speech waveforms, which get sent to the Discriminator for evaluation. Articulatory analysis suggests that the network learns to control articulators in a similar manner to humans during speech production. Acoustic analysis of the outputs suggests that the network learns to generate words that are both present and absent in the training distribution. We additionally discuss implications of articulatory representations for cognitive models of human language and speech technology in general., Comment: ICASSP 2023

Recent self-supervised learning (SSL) models have proven to learn rich representations of speech, which can readily be utilized by diverse downstream tasks. To understand such utilities, various analyses have been done for speech SSL models to reveal which and how information is encoded in the learned representations. Although the scope of previous analyses is extensive in acoustic, phonetic, and semantic perspectives, the physical grounding by speech production has not yet received full attention. To bridge this gap, we conduct a comprehensive analysis to link speech representations to articulatory trajectories measured by electromagnetic articulography (EMA). Our analysis is based on a linear probing approach where we measure articulatory score as an average correlation of linear mapping to EMA. We analyze a set of SSL models selected from the leaderboard of the SUPERB benchmark and perform further layer-wise analyses on two most successful models, Wav2Vec 2.0 and HuBERT. Surprisingly, representations from the recent speech SSL models are highly correlated with EMA traces (best: r = 0.81), and only 5 minutes are sufficient to train a linear model with high performance (r = 0.77). Our findings suggest that SSL models learn to align closely with continuous articulations, and provide a novel insight into speech SSL.

In the articulatory synthesis task, speech is synthesized from input features containing information about the physical behavior of the human vocal tract. This task provides a promising direction for speech synthesis research, as the articulatory space is compact, smooth, and interpretable. Current works have highlighted the potential for deep learning models to perform articulatory synthesis. However, it remains unclear whether these models can achieve the efficiency and fidelity of the human speech production system. To help bridge this gap, we propose a time-domain articulatory synthesis methodology and demonstrate its efficacy with both electromagnetic articulography (EMA) and synthetic articulatory feature inputs. Our model is computationally efficient and achieves a transcription word error rate (WER) of 18.5% for the EMA-to-speech task, yielding an improvement of 11.6% compared to prior work. Through interpolation experiments, we also highlight the generalizability and interpretability of our approach.

In this paper, we propose a novel unsupervised text-to-speech acoustic model training scheme, named UTTS, which does not require text-audio pairs. UTTS is a multi-speaker speech synthesizer that supports zero-shot voice cloning, it is developed from a perspective of disentangled speech representation learning. The framework offers a flexible choice of a speaker's duration model, timbre feature (identity) and content for TTS inference. We leverage recent advancements in self-supervised speech representation learning as well as speech synthesis front-end techniques for system development. Specifically, we employ our recently formulated Conditional Disentangled Sequential Variational Auto-encoder (C-DSVAE) as the backbone UTTS AM, which offers well-structured content representations given unsupervised alignment (UA) as condition during training. For UTTS inference, we utilize a lexicon to map input text to the phoneme sequence, which is expanded to the frame-level forced alignment (FA) with a speaker-dependent duration model. Then, we develop an alignment mapping module that converts FA to UA. Finally, the C-DSVAE, serving as the self-supervised TTS AM, takes the predicted UA and a target speaker embedding to generate the mel spectrogram, which is ultimately converted to waveform with a neural vocoder. We show how our method enables speech synthesis without using a paired TTS corpus in AM development stage. Experiments demonstrate that UTTS can synthesize speech of high naturalness and intelligibility measured by human and objective evaluations. Audio samples are available at our demo page https://neurtts.github.io/utts\_demo/., Comment: IEEE/ACM Transactions on Audio, Speech, and Language Processing ( Volume: 31)

Disentangling content and speaking style information is essential for zero-shot non-parallel voice conversion (VC). Our previous study investigated a novel framework with disentangled sequential variational autoencoder (DSVAE) as the backbone for information decomposition. We have demonstrated that simultaneous disentangling content embedding and speaker embedding from one utterance is feasible for zero-shot VC. In this study, we continue the direction by raising one concern about the prior distribution of content branch in the DSVAE baseline. We find the random initialized prior distribution will force the content embedding to reduce the phonetic-structure information during the learning process, which is not a desired property. Here, we seek to achieve a better content embedding with more phonetic information preserved. We propose conditional DSVAE, a new model that enables content bias as a condition to the prior modeling and reshapes the content embedding sampled from the posterior distribution. In our experiment on the VCTK dataset, we demonstrate that content embeddings derived from the conditional DSVAE overcome the randomness and achieve a much better phoneme classification accuracy, a stabilized vocalization and a better zero-shot VC performance compared with the competitive DSVAE baseline., Comment: Accepted to 2022 Interspeech. Demo link is here https://jlian2.github.io/Improved-Voice-Conversion-with-Conditional-DSVAE/

Most of the research on data-driven speech representation learning has focused on raw audios in an end-to-end manner, paying little attention to their internal phonological or gestural structure. This work, investigating the speech representations derived from articulatory kinematics signals, uses a neural implementation of convolutive sparse matrix factorization to decompose the articulatory data into interpretable gestures and gestural scores. By applying sparse constraints, the gestural scores leverage the discrete combinatorial properties of phonological gestures. Phoneme recognition experiments were additionally performed to show that gestural scores indeed code phonological information successfully. The proposed work thus makes a bridge between articulatory phonology and deep neural networks to leverage informative, intelligible, interpretable,and efficient speech representations., Comment: Accepted to 2022 Interspeech. Code is publicly available at https://github.com/Berkeley-Speech-Group/ema_gesture