Archives

Connect the five dimensions of ASR through language model fusion, knowledge sources, and the constraints of offline, streaming, and on-device deployment.

From DTW and HMMs to CTC, LAS, and RNN-T: how continuous speech maps to discrete text, and the tradeoffs between alignment constraints, output dependencies, and streaming deployment.

From ASR applications, WER/CER, and real-world challenges to MFCCs, CNNs, Transformers, Conformers, and self-supervised speech representations.

An introduction to speech tasks, sound, sampling rates, Fourier transforms, STFT, Mel spectrograms, and MFCCs.

A visual comparison of omni model architectures across audio encoding, Thinker, Talker, vocoder, discrete codecs, continuous latents, and ASR.

How Fun-ASR moved full-context audio encoding ahead of LM admission with batching, single-flight deduplication, and LRU caching.

Investigating a subtle batch-invariance bug in MOSS-TTS Local v1.5's reference-audio cache, where content-addressed keys hid execution-shape-dependent values.

Design agentic alpha discovery system

System design for Google Translate: ML framing, architecture, NMT training, and inference pipeline

Reading notes on the Minimax Speech 2.0 paper: technical details of a speech synthesis model

Reading notes on the Seedream 3.0 technical report

Weekly podcast notes: Japan's health insurance system, short dramas entering Hollywood, Tesla/NVIDIA, Google's Willow quantum chip, and more

A review and reflection on setting-based mystery novels, centered on A Sweet Death for the Famous Detective

Reading notes on the Seedream 2.0 technical report: a Chinese-English bilingual image generation foundation model

Weekly podcast notes: Silicon Valley 101 on MicroStrategy/Bitcoin, a Latent Space interview, and more

Thoughts on exercise, warm-up and rehab, the benefits of exercise, and different happy hormones

Some thoughts on why I started writing a blog: recording ideas, keeping my ability to think, and building a personal knowledge base