I'm Owen, a student developer interested in machine learning, efficient AI, and building useful software. My projects include small speech models, inference experiments, browser agents, and music stats.
Most of my ML releases live on Hugging Face. You'll find the code and apps here.
Hugging Face ↗ · Email ↗ · Discord blue.os
- Smaller models: how architecture and model size affect quality, inference cost, and what can run locally.
- Better training data: synthetic speech, coverage, and testing what a dataset actually teaches a model.
- Models in real software: ONNX conversion, quantization, browser inference, and agents that use tools.
These interests show up in different parts of my projects: model weights and datasets on Hugging Face, inference code and apps on GitHub.
Inflect is my compact text-to-speech project. Nano fits the whole learned text-to-waveform path into 3.96M parameters. Micro uses 9.36M. Both make 24 kHz English speech locally, including the waveform decoder.
Try both voices → · Weights · Code & results
Give the little one a difficult sentence.
Try one in the demo, or bring your own:
At 12:07, Dr. Nguyen ordered 3.5 kilograms of oranges.
The bass player caught a bass. Read what you've read.
Wait—did you say fifteen, or fifty?
The demo runs Nano and Micro on your text. These are challenge prompts, not claims that every reading is correct. If a sentence trips it up, send the exact text.
- Synthetic speech data — 59,996 clips, 56 voices, ~112 hours generated with VoxCPM2, with transcripts and generation metadata. Access terms are on the dataset card.
- Chatterbox Nano → ONNX — an independent four-graph conversion with mixed precision, INT4 weights, and KV-cache handling. Development release.
- Inflect ONNX releases — local inference packages alongside the PyTorch models.
Under the hood: architecture, evaluation, and runtime.
Inflect v2 uses a compact VITS-family architecture: transformer text encoding, stochastic duration prediction, residual coupling flows, and an integrated neural decoder. The parameter counts include all learned modules loaded for synthesis.
The release reports community listening, multi-ASR intelligibility, predicted naturalness, footprint, and CPU throughput separately.
Harbor — browser agents with visual and DOM context, browser tools, streaming progress, and your choice of model.
Sleeve — a music-history explorer: listening eras, changing taste, discoveries, and a next record to try. Open it →
Also: Mr. Tutor, an AI tutoring bot for Discord.
Mostly Python, PyTorch, ONNX Runtime, CUDA, TypeScript, and React.
Tried something I've made? I'd like to hear how it went.
owen.aw.song@gmail.com · Discord blue.os