Uncertainty Distillation: Teaching Language Models to Express Semantic Confidence

As large language models (LLMs) are increasingly used for factual question-answering, it becomes more important for LLMs to have the capability to communicate the likelihood that their answer is correct. For these verbalized expressions of uncertainty to be meaningful, they should reflect the error rates at the expressed level of confidence. However, when prompted to express confidence, the error rates of current LLMs are inconsistent with their communicated confidences, highlighting the need for uncertainty quantification methods. Many prior methods calculate lexical uncertainty, estimating a model's confidence in the specific string it generated. In some cases, however, it may be more useful to estimate semantic uncertainty, or the model's confidence in the answer regardless of how it is verbalized. We propose a simple procedure, uncertainty distillation, to teach an LLM to verbalize calibrated semantic confidences. Using held-out data to map initial uncertainty estimates to meaningful probabilities, we create examples annotated with verbalized probabilities for supervised fine-tuning. We demonstrate our method yields verbalized confidences that correlate with observed error rates with a small fine-tuned language model as well as with larger instruction-tuned models, and find that our semantic uncertainty correlates well with lexical uncertainty on short answers.

Less is More for Synthetic Speech Detection in the Wild

Driven by advances in self-supervised learning for speech, state-of-the-art synthetic speech detectors have achieved low error rates on popular benchmarks such as ASVspoof. However, prior benchmarks do not address the wide range of real-world variability in speech. Are reported error rates realistic in real-world conditions? To assess detector failure modes and robustness under controlled distribution shifts, we introduce ShiftySpeech, a benchmark with more than 3000 hours of synthetic speech from 7 domains, 6 TTS systems, 12 vocoders, and 3 languages. We found that all distribution shifts degraded model performance, and contrary to prior findings, training on more vocoders, speakers, or with data augmentation did not guarantee better generalization. In fact, we found that training on less diverse data resulted in better generalization, and that a detector fit using samples from a single carefully selected vocoder and a single speaker achieved state-of-the-art results on the challenging In-the-Wild benchmark.

GenVC: Self-Supervised Zero-Shot Voice Conversion

Zero-shot voice conversion has recently made substantial progress, but many models still depend on external supervised systems to disentangle speaker identity and linguistic content. Furthermore, current methods often use parallel conversion, where the converted speech inherits the source utterance's temporal structure, restricting speaker similarity and privacy. To overcome these limitations, we introduce GenVC, a generative zero-shot voice conversion model. GenVC learns to disentangle linguistic content and speaker style in a self-supervised manner, eliminating the need for external models and enabling efficient training on large, unlabeled datasets. Experimental results show that GenVC achieves state-of-the-art speaker similarity while maintaining naturalness competitive with leading approaches. Its autoregressive generation also allows the converted speech to deviate from the source utterance's temporal structure. This feature makes GenVC highly effective for voice anonymization, as it minimizes the preservation of source prosody and speaker characteristics, enhancing privacy protection.

Merging Feed-Forward Sublayers for Compressed Transformers

With the rise and ubiquity of larger deep learning models, the need for high-quality compression techniques is growing in order to deploy these models widely. The sheer parameter count of these models makes it difficult to fit them into the memory constraints of different hardware. In this work, we present a novel approach to model compression by merging similar parameter groups within a model, rather than pruning away less important parameters. Specifically, we select, align, and merge separate feed-forward sublayers in Transformer models, and test our method on language modeling, image classification, and machine translation. With our method, we demonstrate performance comparable to the original models while combining more than a third of model feed-forward sublayers, and demonstrate improved performance over a strong layer-pruning baseline. For instance, we can remove over 21% of total parameters from a Vision Transformer, while maintaining 99% of its original performance. Additionally, we observe that some groups of feed-forward sublayers exhibit high activation similarity, which may help explain their surprising mergeability.

SpeechQE: Estimating the Quality of Direct Speech Translation

Recent advances in automatic quality estimation for machine translation have exclusively focused on written language, leaving the speech modality underexplored. In this work, we formulate the task of quality estimation for speech translation (SpeechQE), construct a benchmark, and evaluate a family of systems based on cascaded and end-to-end architectures. In this process, we introduce a novel end-to-end system leveraging pre-trained text LLM. Results suggest that end-to-end approaches are better suited to estimating the quality of direct speech translation than using quality estimation systems designed for text in cascaded systems. More broadly, we argue that quality estimation of speech translation needs to be studied as a separate problem from that of text, and release our data and models to guide further research in this space.

HLTCOE JHU Submission to the Voice Privacy Challenge 2024

We present a number of systems for the Voice Privacy Challenge, including voice conversion based systems such as the kNN-VC method and the WavLM voice Conversion method, and text-to-speech (TTS) based systems including Whisper-VITS. We found that while voice conversion systems better preserve emotional content, they struggle to conceal speaker identity in semi-white-box attack scenarios; conversely, TTS methods perform better at anonymization and worse at emotion preservation. Finally, we propose a random admixture system which seeks to balance out the strengths and weaknesses of the two category of systems, achieving a strong EER of over 40% while maintaining UAR at a respectable 47%.

Privacy versus Emotion Preservation Trade-offs in Emotion-Preserving Speaker Anonymization

Advances in speech technology now allow unprecedented access to personally identifiable information through speech. To protect such information, the differential privacy field has explored ways to anonymize speech while preserving its utility, including linguistic and paralinguistic aspects. However, anonymizing speech while maintaining emotional state remains challenging. We explore this problem in the context of the VoicePrivacy 2024 challenge. Specifically, we developed various speaker anonymization pipelines and find that approaches either excel at anonymization or preserving emotion state, but not both simultaneously. Achieving both would require an in-domain emotion recognizer. Additionally, we found that it is feasible to train a semi-effective speaker verification system using only emotion representations, demonstrating the challenge of separating these two modalities.

Efficiency-Effectiveness Tradeoff of Probabilistic Structured Queries for Cross-Language Information Retrieval

Probabilistic Structured Queries (PSQ) is a cross-language information retrieval (CLIR) method that uses translation probabilities statistically derived from aligned corpora. PSQ is a strong baseline for efficient CLIR using sparse indexing. It is, therefore, useful as the first stage in a cascaded neural CLIR system whose second stage is more effective but too inefficient to be used on its own to search a large text collection. In this reproducibility study, we revisit PSQ by introducing an efficient Python implementation. Unconstrained use of all translation probabilities that can be estimated from aligned parallel text would in the limit assign a weight to every vocabulary term, precluding use of an inverted index to serve queries efficiently. Thus, PSQ's effectiveness and efficiency both depend on how translation probabilities are pruned. This paper presents experiments over a range of modern CLIR test collections to demonstrate that achieving Pareto optimal PSQ effectiveness-efficiency tradeoffs benefits from multi-criteria pruning, which has not been fully explored in prior work. Our Python PSQ implementation is available on GitHub(https://github.com/hltcoe/PSQ) and unpruned translation tables are available on Huggingface Models(https://huggingface.co/hltcoe/psq_translation_tables).

Low-Resource Named Entity Recognition with Cross-Lingual, Character-Level Neural Conditional Random Fields

Low-resource named entity recognition is still an open problem in NLP. Most state-of-the-art systems require tens of thousands of annotated sentences in order to obtain high performance. However, for most of the world's languages, it is unfeasible to obtain such annotation. In this paper, we present a transfer learning scheme, whereby we train character-level neural CRFs to predict named entities for both high-resource languages and low resource languages jointly. Learning character representations for multiple related languages allows transfer among the languages, improving F1 by up to 9.8 points over a loglinear CRF baseline.

Where does In-context Translation Happen in Large Language Models

Self-supervised large language models have demonstrated the ability to perform Machine Translation (MT) via in-context learning, but little is known about where the model performs the task with respect to prompt instructions and demonstration examples. In this work, we attempt to characterize the region where large language models transition from in-context learners to translation models. Through a series of layer-wise context-masking experiments on \textsc{GPTNeo2.7B}, \textsc{Bloom3B}, \textsc{Llama7b} and \textsc{Llama7b-chat}, we demonstrate evidence of a "task recognition" point where the translation task is encoded into the input representations and attention to context is no longer necessary. We further observe correspondence between the low performance when masking out entire layers, and the task recognition layers. Taking advantage of this redundancy results in 45\% computational savings when prompting with 5 examples, and task recognition achieved at layer 14 / 32. Our layer-wise fine-tuning experiments indicate that the most effective layers for MT fine-tuning are the layers critical to task recognition.