Zijin Gu

Zijin Gu, Keith Jamison, Amy Kuceyeski et al.

Due to the low signal-to-noise ratio and limited resolution of functional MRI data, and the high complexity of natural images, reconstructing a visual stimulus from human brain fMRI measurements is a challenging task. In this work, we propose a novel approach for this task, which we call Cortex2Image, to decode visual stimuli with high semantic fidelity and rich fine-grained detail. In particular, we train a surface-based convolutional network model that maps from brain response to semantic image features first (Cortex2Semantic). We then combine this model with a high-quality image generator (Instance-Conditioned GAN) to train another mapping from brain response to fine-grained image features using a variational approach (Cortex2Detail). Image reconstructions obtained by our proposed method achieve state-of-the-art semantic fidelity, while yielding good fine-grained similarity with the ground-truth stimulus. Our code is available at: https://github.com/zijin-gu/meshconv-decoding.git.

Jason Ramapuram, Federico Danieli, Eeshan Dhekane et al. · apple-ml, berkeley

Attention is a key part of the transformer architecture. It is a sequence-to-sequence mapping that transforms each sequence element into a weighted sum of values. The weights are typically obtained as the softmax of dot products between keys and queries. Recent work has explored alternatives to softmax attention in transformers, such as ReLU and sigmoid activations. In this work, we revisit sigmoid attention and conduct an in-depth theoretical and empirical analysis. Theoretically, we prove that transformers with sigmoid attention are universal function approximators and benefit from improved regularity compared to softmax attention. Through detailed empirical analysis, we identify stabilization of large initial attention norms during the early stages of training as a crucial factor for the successful training of models with sigmoid attention, outperforming prior attempts. We also introduce FLASHSIGMOID, a hardware-aware and memory-efficient implementation of sigmoid attention yielding a 17% inference kernel speed-up over FLASHATTENTION2 on H100 GPUs. Experiments across language, vision, and speech show that properly normalized sigmoid attention matches the strong performance of softmax attention on a wide range of domains and scales, which previous attempts at sigmoid attention were unable to fully achieve. Our work unifies prior art and establishes best practices for sigmoid attention as a drop-in softmax replacement in transformers.

Richard He Bai, Tatiana Likhomanenko, Ruixiang Zhang et al. · apple-ml

Large language models have revolutionized natural language processing by leveraging self-supervised pretraining on vast textual data. Inspired by this success, researchers have investigated various compression-based speech tokenization methods to discretize continuous speech signals, enabling the application of language modeling techniques to discrete tokens. However, audio compressor introduces additional complexity and computational cost, and often fail on out-of-domain audio signals. In this work, we introduce a novel speech representation (dmel) that discretizes mel-filterbank channels into intensity bins, creating a simpler yet more effective representation compared to existing speech tokenization methods. Our approach demonstrates superior performance in preserving audio content, robustness to out-of-domain data, and offers a training-free, natural, and streamable representation. To address the high-dimensional nature of log-mel spectrograms, we propose an efficient parallel encoding and decoding method for high-dimensional tokens using an LM-style transformer architecture. This innovation enables us to develop RichTTS and RichASR, two models sharing the same architecture while achieving comparable or better results than specialized existing methods. Our results demonstrate the effectiveness of dmel in achieving high performance on both speech synthesis and recognition tasks within a unified framework, paving the way for efficient and effective joint modeling of speech and text.

Zijin Gu, Tatiana Likhomanenko, He Bai et al. · apple-ml

Language models play a central role in automatic speech recognition (ASR), yet most methods rely on text-only models unaware of ASR error patterns. Recently, large language models (LLMs) have been applied to ASR correction, but introduce latency and hallucination concerns. We revisit ASR error correction with compact seq2seq models, trained on ASR errors from real and synthetic audio. To scale training, we construct synthetic corpora via cascaded TTS and ASR, finding that matching the diversity of realistic error distributions is key. We propose correction-first decoding, where the correction model generates candidates rescored using ASR acoustic scores. With 15x fewer parameters than LLMs, our model achieves 1.5/3.3% WER on LibriSpeech test-clean/other, outperforms LLMs, generalizes across ASR architectures (CTC, Seq2seq, Transducer) and diverse domains, and provides precise corrections in the low-error regime where LLMs struggle.

Zijin Gu, Tatiana Likhomanenko, Vimal Thilak et al. · apple-ml

Sparse Mixture-of-Experts (MoE) architectures enable efficient scaling by activating only a subset of parameters for each input. However, conventional MoE routing selects each layer's experts independently, creating N^L possible expert paths -- for N experts across L layers. This far exceeds typical training set sizes, leading to statistical inefficiency as the model may not learn meaningful structure over such a vast path space. To constrain it, we propose \pathmoe, which shares router parameters across consecutive layers. Experiments on 0.9B and 16B parameter models demonstrate consistent improvements on perplexity and downstream tasks over independent routing, while eliminating the need for auxiliary load balancing losses. Analysis reveals that tokens following the same path naturally cluster by linguistic function, with \pathmoe{} producing more concentrated groups, better cross-layer consistency, and greater robustness to routing perturbations. These results offer a new perspective for understanding MoE architectures through the lens of expert paths.

Zakaria Aldeneh, Skyler Seto, Maureen de Seyssel et al.

Modern ASR systems are typically trained on large-scale pseudo-labeled, in-the-wild data spanning multiple domains. While such heterogeneous data benefit generalist models designed for broad deployment, they pose challenges for specialist models targeting specific domains: specialist models lack the capacity to learn from all available data, and one must pay closer attention to addressing the mismatch between training and test conditions. In this work, we study targeted data selection as a strategy to address these challenges, selecting relevant subsets from 100k hours of in-the-wild training data to optimize performance on target domains. We represent speech samples using embeddings that capture complementary characteristic--speaker attributes, phonetic content, and semantic meaning--and analyze how relevance and diversity along these axes when performing data selection affect downstream ASR performance. Our experiments with CTC-based Conformer models show that training on a strategically selected 5% subset can exceed the performance of models trained on the full dataset by up to 36.8% relative WER reduction on target domains.

Yizhe Zhang, Richard Bai, Zijin Gu et al. · apple-ml

Language models usually use left-to-right (L2R) autoregressive factorization. However, L2R factorization may not always be the best inductive bias. Therefore, we investigate whether alternative factorizations of the text distribution could be beneficial in some tasks. We investigate right-to-left (R2L) training as a compelling alternative, focusing on multiple-choice questions (MCQs) as a test bed for knowledge extraction and reasoning. Through extensive experiments across various model sizes (2B-8B parameters) and training datasets, we find that R2L models can significantly outperform L2R models on several MCQ benchmarks, including logical reasoning, commonsense understanding, and truthfulness assessment tasks. Our analysis reveals that this performance difference may be fundamentally linked to multiple factors including calibration, computability, and directional conditional entropy. We analyze the impact of these factors through controlled simulation studies using arithmetic tasks, where the impacting factors can be better disentangled. Our work demonstrates that exploring alternative factorizations of the text distribution can lead to improvements in LLM capabilities and provides theoretical insights into optimal factorization towards approximating human language distribution, and when each reasoning order might be more advantageous. Our code and checkpoints are released at https://github.com/apple/ml-reversal-blessing.

Jason Ramapuram, Eeshan Gunesh Dhekane, Amitis Shidani et al.

Speech-only spoken language models (SLMs) lag behind text and text-speech models in performance, with recent discrete autoregressive (AR) SLMs indicating significant computational and data demands to match text models. Since discretizing continuous speech for AR creates bottlenecks, we explore whether continuous diffusion (CD) SLM is more viable. To quantify the SLMs linguistic quality, we introduce the phoneme Jensen-Shannon divergence (pJSD) metric. Our analysis reveals CD SLMs, mirroring AR behavior, exhibit scaling laws for validation loss and pJSD, and show optimal token-to-parameter ratios decreasing as compute scales. However, for the latter, loss becomes insensitive to choice of data and model sizes, showing potential for fast inference. Scaling CD SLMs to 16B parameters with tens of millions of hours of conversational data enables generation of emotive, prosodic, multi-speaker, multilingual speech, though achieving long-form coherence remains a significant challenge.

Richard He Bai, Zijin Gu, Tatiana Likhomanenko et al.

The latency bottleneck of traditional text-to-speech (TTS) systems fundamentally hinders the potential of streaming large language models (LLMs) in conversational AI. These TTS systems, typically trained and inferenced on complete utterances, introduce unacceptable delays, even with optimized inference speeds, when coupled with streaming LLM outputs. This is particularly problematic for creating responsive conversational agents where low first-token latency is critical. In this paper, we present SpeakStream, a streaming TTS system that generates audio incrementally from streaming text using a decoder-only architecture. SpeakStream is trained using a next-step prediction loss on interleaved text-speech data. During inference, it generates speech incrementally while absorbing streaming input text, making it particularly suitable for cascaded conversational AI agents where an LLM streams text to a TTS system. Our experiments demonstrate that SpeakStream achieves state-of-the-art latency results in terms of first-token latency while maintaining the quality of non-streaming TTS systems.

Tatiana Likhomanenko, Luke Carlson, Richard He Bai et al. · apple-ml

The emergence of large language models (LLMs) has transformed spoken dialog systems, yet the optimal architecture for real-time on-device voice agents remains an open question. While end-to-end approaches promise theoretical advantages, cascaded systems (CSs) continue to outperform them in language understanding tasks, despite being constrained by sequential processing latency. In this work, we introduce ChipChat, a novel low-latency CS that overcomes traditional bottlenecks through architectural innovations and streaming optimizations. Our system integrates streaming (a) conversational speech recognition with mixture-of-experts, (b) state-action augmented LLM, (c) text-to-speech synthesis, (d) neural vocoder, and (e) speaker modeling. Implemented using MLX, ChipChat achieves sub-second response latency on a Mac Studio without dedicated GPUs, while preserving user privacy through complete on-device processing. Our work shows that strategically redesigned CSs can overcome their historical latency limitations, offering a promising path forward for practical voice-based AI agents.

Zijin Gu, Tatiana Likhomanenko, Navdeep Jaitly

Mixture-of-experts (MoE) architectures have expanded from language modeling to automatic speech recognition (ASR). Traditional MoE methods, such as the Switch Transformer, route experts independently within each layer. Our analysis reveals that routers in most layers make expert choices that are not strongly correlated with the choices of the routers in other layers. To increase the cooperation between experts in different layers and encourage greater specialization, we use a shared router across different MoE layers. We call this model Omni-router Transformer. Extensive experiments on a large-scale pseudo-labeled dataset and evaluations across 10 diverse, out-of-domain ASR benchmarks demonstrate that the Omni-router Transformer is able to achieve lower training loss and consistently outperform dense and Switch Transformer models, reducing average word error rates by 11.2% and 8.2%, respectively, while providing structured expert usage and improved robustness to diverse data.

Santiago Cuervo, Skyler Seto, Maureen de Seyssel et al.

Large Language Models (LLMs) can be adapted to extend their text capabilities to speech inputs. However, these speech-adapted LLMs consistently underperform their text-based counterparts--and even cascaded pipelines--on language understanding tasks. We term this shortfall the text-speech understanding gap: the performance drop observed when a speech-adapted LLM processes spoken inputs relative to when the original text-based LLM processes the equivalent text. Recent approaches to narrowing this gap either rely on large-scale speech synthesis of text corpora, which is costly and heavily dependent on synthetic data, or on large-scale proprietary speech datasets, which are not reproducible. As a result, there remains a need for more data-efficient alternatives for closing the text-speech understanding gap. In this work, we analyze the gap as driven by two factors: (i) forgetting of text capabilities during adaptation, and (ii) cross-modal misalignment between speech and text. Based on this analysis, we introduce SALAD--Sample-efficient Alignment with Learning through Active selection and cross-modal Distillation--which combines cross-modal distillation with targeted synthetic data to improve alignment while mitigating forgetting. Applied to 3B and 7B LLMs, SALAD achieves competitive performance with a strong open-weight model across broad-domain benchmarks in knowledge, language understanding, and reasoning, while training on over an order of magnitude less speech data from public corpora.

Zijin Gu, Keith Jamison, Mert Sabuncu et al.

Encoding models that predict brain response patterns to stimuli are one way to capture this relationship between variability in bottom-up neural systems and individual's behavior or pathological state. However, they generally need a large amount of training data to achieve optimal accuracy. Here, we propose and test an alternative personalized ensemble encoding model approach to utilize existing encoding models, to create encoding models for novel individuals with relatively little stimuli-response data. We show that these personalized ensemble encoding models trained with small amounts of data for a specific individual, i.e. ~300 image-response pairs, achieve accuracy not different from models trained on ~20,000 image-response pairs for the same individual. Importantly, the personalized ensemble encoding models preserve patterns of inter-individual variability in the image-response relationship. Additionally, we show the proposed approach is robust against domain shift by validating on a prospectively collected set of image-response data in novel individuals with a different scanner and experimental setup. Finally, we use our personalized ensemble encoding model within the recently developed NeuroGen framework to generate optimal stimuli designed to maximize specific regions' activations for a specific individual. We show that the inter-individual differences in face areas responses to images of animal vs human faces observed previously is replicated using NeuroGen with the ensemble encoding model. Our approach shows the potential to use previously collected, deeply sampled data to efficiently create accurate, personalized encoding models and, subsequently, personalized optimal synthetic images for new individuals scanned under different experimental conditions.

Zijin Gu, Keith W. Jamison, Meenakshi Khosla et al.

Functional MRI (fMRI) is a powerful technique that has allowed us to characterize visual cortex responses to stimuli, yet such experiments are by nature constructed based on a priori hypotheses, limited to the set of images presented to the individual while they are in the scanner, are subject to noise in the observed brain responses, and may vary widely across individuals. In this work, we propose a novel computational strategy, which we call NeuroGen, to overcome these limitations and develop a powerful tool for human vision neuroscience discovery. NeuroGen combines an fMRI-trained neural encoding model of human vision with a deep generative network to synthesize images predicted to achieve a target pattern of macro-scale brain activation. We demonstrate that the reduction of noise that the encoding model provides, coupled with the generative network's ability to produce images of high fidelity, results in a robust discovery architecture for visual neuroscience. By using only a small number of synthetic images created by NeuroGen, we demonstrate that we can detect and amplify differences in regional and individual human brain response patterns to visual stimuli. We then verify that these discoveries are reflected in the several thousand observed image responses measured with fMRI. We further demonstrate that NeuroGen can create synthetic images predicted to achieve regional response patterns not achievable by the best-matching natural images. The NeuroGen framework extends the utility of brain encoding models and opens up a new avenue for exploring, and possibly precisely controlling, the human visual system.