Work in progress...
work in progress...
This blog is generated using a language model (LLM) that transcribes my recorded audio in Japanese. Once I have accumulated a sufficient amount of data (approximately one year), I plan to publish a dataset that will include:
I'm using VuePress to build this blog, along with the VuePress-Hope theme for web design.
Once everything is set up, all I need to do is write in Markdown, and VuePress-Hope automatically recognizes tags and categories, formatting the documents accordingly. It also seamlessly switches between English and Japanese, which I find to be a particularly useful feature.
work in progress...
I'd like to explain about the research I've been get in touch with.
- A Comparative Study on Transformer vs RNN in Speech Applications: The comparison of RNN based models and Transformer based models on Speech to text task.
- ESPnet-ONNX: Bridging a Gap Between Research and Production: A software to convert pre-trained model of ESPnet into ONNX.
- Segment-Level Vectorized Beam Search Based on Partially Autoregressive Inference: This proposes a decoding process that combines the autoregressive and non-autoregressive process. We improved trade-off between accuracy and inference speed.
Abstract
By combining autoregressive (AR) and non-autoregressive (NAR) decoding, we developed a partially autoregressive (PAR) approach that leverages the strengths of both methods. This resulted in a 12-13x speedup with minimal accuracy degradation. A key advantage of PAR is that it doesn’t require training a new model; we can achieve this speedup using a pre-trained AR model. I've integrated this decoding algorithm into ESPnet: #5769
work in progress...
Self-Introduction
Short Bio
I majored in Statistics at the School of Economics, Nagoya University, following my high school graduation. During my undergraduate years, I interned at Human Dataware Lab and held a part-time position at Tarvo. After earning my degree from Nagoya University, I joined IBM Japan, where I was involved in system development for an insurance company. In particular, my pioneering work on generative AI, for which I served as technical lead, was widely reported in the news.