arXiv Science⌕ Search

arXiv subjects

Ethan Zou

Publications and source records attributed to Ethan Zou.

3 recordsLinked to original sources

Hy ASR 3.0 Technical Report

We present Hy ASR 3.0, a large language model (LLM)-based automatic speech recognition (ASR) system that delivers strong recognition accuracy, supports flexible contextual conditioning, and maintains robust recognition performance across diverse speech conditions. Hy ASR 3.0 supports Mandarin Chinese, English and 13 Chinese dialect groups, including Cantonese. Among the evaluated systems, it achieves the lowest aggregate error rates on public Mandarin Chinese, English, and Cantonese benchmarks and the lowest error rates on the corresponding internal test sets, alongside first-tier dialect recognition performance. Beyond audio-only transcription, Hy ASR 3.0 supports multiple forms of external context, including hotword lists, question--answer dialogues, and summaries, enabling the use of task-relevant information during recognition. The model also leverages intra-utterance context to disambiguate homophonic and phonetically similar keywords using information from other parts of the same utterance, without requiring a separate contextual input. Evaluations demonstrate strong contextual recognition in both external-context and intra-utterance settings, as well as robust recognition of long-form speech, children's and older adults' speech, low-volume speech, and speech in noisy environments. This report describes the model architecture and training methodology and presents evaluations across recognition accuracy, context utilization, and robustness.

cs.CL↗

Tardis 2.0: Optimized Time Traveling Coherence for Relaxed Consistency Models

Cache coherence scalability is a big challenge in shared memory systems. Traditional protocols do not scale due to the storage and traffic overhead of cache invalidation. Tardis, a recently proposed coherence protocol, removes cache invalidation using logical timestamps and achieves excellent scalability. The original Tardis protocol, however, only supports the Sequential Consistency (SC) memory model, limiting its applicability. Tardis also incurs extra network traffic on some benchmarks due to renew messages, and has suboptimal performance when the program uses spinning to communicate between threads. In this paper, we address these downsides of Tardis protocol and make it significantly more practical. Specifically, we discuss the architectural, memory system and protocol changes required in order to implement the TSO consistency model on Tardis, and prove that the modified protocol satisfies TSO. We also describe modifications for Partial Store Order (PSO) and Release Consistency (RC). Finally, we propose optimizations for better leasing policies and to handle program spinning. On a set of benchmarks, optimized Tardis improves on a full-map directory protocol in the metrics of performance, storage and network traffic, while being simpler to implement.

cs.AR↗

Optimizing Path ORAM for Cloud Storage Applications

We live in a world where our personal data are both valuable and vulnerable to misappropriation through exploitation of security vulnerabilities in online services. For instance, Dropbox, a popular cloud storage tool, has certain security flaws that can be exploited to compromise a user's data, one of which being that a user's access pattern is unprotected. We have thus created an implementation of Path Oblivious RAM (Path ORAM) for Dropbox users to obfuscate path access information to patch this vulnerability. This implementation differs significantly from the standard usage of Path ORAM, in that we introduce several innovations, including a dynamically growing and shrinking tree architecture, multi-block fetching, block packing and the possibility for multi-client use. Our optimizations together produce about a 77% throughput increase and a 60% reduction in necessary tree size; these numbers vary with file size distribution.

cs.CR↗