Large Language Model (LLM) training is increasingly concentrated in homogeneous datacenters, while private data and underutilized GPUs across universities, laboratories, and edge sites remain difficult to use. This extended abstract presents preliminary results from a geo-distributed LLM training prototype that treats networking constraints as first-order design concerns. The prototype connects three heterogeneous GPU sites via cloud-hosted parameter servers, outbound-only gRPC streams, two-stage delta compression (INT8 quantization + Huffman coding, achieving up to 4× payload reduction), and fault-tolerant rejoin. In real deployments, GPT-2 Medium pretraining achieves stable loss reduction and reaches the target loss 15.2% faster in wall-clock time than the best tested baseline; Llama3-1B pretraining remains stable under larger communication pressure; and cross-site latency traces reveal site-dependent WAN spikes of up to 200s. These results motivate adaptive networking support for synchronization, compression, placement, telemetry, and recovery in geo-distributed LLM training.

Existing automatic prompt engineering methods are typically designed for discriminative tasks, where new task prompts are iteratively refined with limited feedback from a single metric reflecting a single aspect. However, these approaches are suboptimal for generative tasks, which require more nuanced guidance beyond a single numeric metric to improve the prompt and optimize multiple aspects of the generated text. To address these challenges, we propose a novel multi-aspect Critique-Suggestion-guided automatic Prompt Optimization (CriSPO) approach. CriSPO introduces a critique-suggestion module as its core component. This module spontaneously discovers aspects, and compares generated and reference texts across these aspects, providing specific suggestions for prompt modification. These clear critiques and actionable suggestions guide a receptive optimizer module to make more substantial changes, exploring a broader and more effective search space. To further improve CriSPO with multi-metric optimization, we introduce an Automatic Suffix Tuning (AST) extension to enhance the performance of task prompts across multiple metrics. We evaluate CriSPO on 4 state-of-the-art Large Language Models (LLMs) across 4 summarization and 5 Question Answering (QA) datasets. Extensive experiments show 3-4% ROUGE score improvement on summarization and substantial improvement of various metrics on QA.

Speaker Diarization (SD) is a crucial component of modern end-to-end ASR pipelines. Traditional SD systems, which are typically audio-based and operate independently of ASR, often introduce speaker errors, particularly during speaker transitions and overlapping speech. Recently, language models including fine-tuned large language models (LLMs) have shown to be effective as a second-pass speaker error corrector by leveraging lexical context in the transcribed output. In this work, we introduce a novel acoustic conditioning approach to provide more fine-grained information from the acoustic diarizer to the LLM. We also show that a simpler constrained decoding strategy reduces LLM hallucinations, while avoiding complicated post-processing. Our approach significantly reduces the speaker error rates by 24-43% across Fisher, Callhome, and RT03-CTS datasets, compared to the first-pass Acoustic SD.

The power of foundation models (FMs) lies in their capacity to learn highly expressive representations that can be adapted to a broad spectrum of tasks. However, these pretrained models require additional training stages to become effective for downstream applications. In the multi-task setting, prior works have shown empirically that specific meta-learning approaches for preparing a model for future adaptation through parameter-efficient fine-tuning (PEFT) can outperform standard retraining methods, but the mechanism of the benefits of meta-learning has been largely unexplored. We introduce a framework for generic PEFT-based meta-learning to learn a model that can easily adapt to unseen tasks. For linear models using LoRA, we show that standard retraining is provably suboptimal for finding an adaptable set of parameters and provide strict performance guarantees for our proposed method. We verify these theoretical insights through experiments on synthetic data as well as real-data vision and language tasks. We observe significant performance benefits using a simple implementation of our proposed meta-learning scheme during retraining relative to the conventional approach.

Juan Pablo Zuluaga-Gomez, Zhaocheng Huang, Xing Niu, Rohit Paturi, Sundararajan Srinivasan, Prashant Mathur, Brian Thompson, Marcello Federico. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing. 2023.

Audio-visual data allows us to leverage different modalities for downstream tasks. The idea being individual streams can complement each other in the given task, thereby resulting in a model with improved performance. In this work, we present our experimental results on action recognition and video summarization tasks. The proposed modeling approach builds upon the recent advances in contrastive loss based audio-visual representation learning. Temporally cognizant audio-visual discrimination is achieved in a Transformer model by learning with a masked feature reconstruction loss over a fixed time window in addition to learning via contrastive loss. Overall, our results indicate that the addition of temporal information significantly improved the performance of the contrastive loss based framework. We achieve an action classification accuracy of 66.2% versus the next best baseline at 64.7% on the HMDB dataset. For video summarization, we attain an F1 score of 43.5 verses 42.2 on the SumMe dataset.

Generic pre-trained speech and text representations promise to reduce the need for large labeled datasets on specific speech and language tasks. However, it is not clear how to effectively adapt these representations for speech emotion recognition. Recent public benchmarks show the efficacy of several popular self-supervised speech representations for emotion classification. In this study, we show that the primary difference between the top-performing representations is in predicting valence while the differences in predicting activation and dominance dimensions are less pronounced. However, we show that even the best-performing HuBERT representation underperforms on valence prediction compared to a multimodal model that also incorporates text representation. We address this shortcoming by injecting lexical information into the speech representation using the multimodal model as a teacher. To improve the efficacy of our approach, we propose a novel estimate of the quality of the emotion predictions, to condition teacher-student training. We report new audio-only state-of-the-art concordance correlation coefficient (CCC) values of 0.757, 0.627, 0.671 for activation, valence and dominance predictions, respectively, on the MSP-Podcast corpus, and also state-of-the-art values of 0.667, 0.582, 0.545 on the IEMOCAP corpus.

Many of the recent advances in speech separation are primarily aimed at synthetic mixtures of short audio utterances with high degrees of overlap.Most of these approaches need an additional stitching step to stitch the separated speech chunks for long form audio. Since most of the approaches involve Permutation Invariant training (PIT), the order of separated speech chunks is nondeterministic and leads to difficulty in accurately stitching homogenous speaker chunks for downstream tasks like Automatic Speech Recognition (ASR).Also, most of these models are trained with synthetic mixtures and do not generalize to real conversational data.In this paper, we propose a speaker conditioned separator trained on speaker embeddings extracted directly from the mixed signal using an over-clustering based approach.This model naturally regulates the order of the separated chunks without the need for an additional stitching step.We also introduce a data sampling strategy with real and synthetic mixtures which generalizes well to real conversation speech.With this model and data sampling technique, we show significant improvements in speaker-attributed word error rate (SA-WER) on Hub5 data.

Training deep neural networks for automatic speech recognition (ASR) requires large amounts of transcribed speech. This becomes a bottleneck for training robust models for accented speech which typically contains high variability in pronunciation and other semantics, since obtaining large amounts of annotated accented data is both tedious and costly. Often, we only have access to large amounts of unannotated speech from different accents. In this work, we leverage this unannotated data to provide semantic regularization to an ASR model that has been trained only on one accent, to improve its performance for multiple accents. We propose Accent Pre-Training (Acc-PT), a semi-supervised training strategy that combines transfer learning and adversarial training. Our approach improves the performance of a state-of-the-art ASR model by 33% on average over the baseline across multiple accents, training only on annotated samples from one standard accent, and as little as 105 minutes of unannotated speech from a target accent.

Speaker diarization accuracy can be affected by both acoustics and conversation characteristics. Determining the cause of diarization errors is difficult because speaker voice acoustics and conversation structure co-vary, and the interactions between acoustics, conversational structure, and diarization accuracy are complex. This paper proposes a methodology that can distinguish independent marginal effects of acoustic and conversation characteristics on diarization accuracy by remixing conversations in a factorial design. As an illustration, this approach is used to investigate gender-related and language-related accuracy differences with three diarization systems: a baseline system using subsegment x-vector clustering, a variant of it with shorter subsegments, and a third system based on a Bayesian hidden Markov model. Our analysis shows large accuracy disparities for the baseline system primarily due to conversational structure, which are partially mitigated in the other two systems. The illustration thus demonstrates how the methodology can be used to identify and guide diarization model improvements.

In this paper, we present our work using ultrasonic sensing of speech for digit recognition. First, a set of spectral ultrasonic features are developed and tuned in order to achieve optimal performance for the digit recognition task. Using these features, we demonstrate an overall accuracy of 33.00% on a digit recognition task using HMMs with recordings from 6 speakers. The results indicate that ultrasonic sensing of speech is viable, but that further work is needed to achieve word accuracies that match those of audio. Finally, experimental results are presented which demonstrate that fusing information from ultrasound and audio sources show marginal improvements over audio-only performances.

Gaussian mixture models are a very successful method for modeling the output distribution of a state in a hidden Markov model (HMM). However, this approach is limited by the assumption that the dynamics of speech features are linear and can be modeled with static features and their derivatives. In this paper, a nonlinear mixture autoregressive model is used to model state output distributions (MAR-HMM). Estimation of model parameters is extended to handle vector features. MAR-HMMs are shown to provide superior performance to comparable Gaussian mixture model-based HMMs (GMM-HMM) with lower complexity on two pilot classification tasks.