I'm Owen, a student developer working on machine learning and useful software. I'm interested in what smaller models can do, how to evaluate them, and what happens when you give a model tools it can actually use.
My work spans speech models, training data, inference, and browser agents. Most of my model releases live on Hugging Face; the code and apps are here.
- Model efficiency: how architecture, model size and quantization affect quality and what can run locally.
- Speech & training data: small TTS models, synthetic datasets, and coverage beyond the easy examples.
- Evaluation: looking at individual failures as well as aggregate scores, with recordings and reproducible comparisons.
- Agents & tools: connecting models to browser interfaces and development workflows, then making their actions inspectable.
These interests connect the projects below—from training and exporting a speech model to testing the tools around it.
Inflect is my compact text-to-speech project. Nano fits the complete learned text-to-waveform path into 3.96M parameters; Micro uses 9.36M. Both generate 24 kHz English speech locally, including the waveform decoder.
Try your own sentence → · Nano weights · Micro weights · Code & evaluation
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?
These are challenge prompts, not promises of a perfect reading. If something trips it up, send the exact sentence.
- Synthetic speech data — 59,996 clips · 56 voices · ~112 hours, generated with VoxCPM2. Transcripts and generation metadata are included; access terms are on the dataset card.
- Chatterbox Nano → ONNX — an independent four-graph conversion with mixed precision, INT4 weights and KV-cache handling. A development release.
- Inflect ONNX — local inference packages for Nano and Micro, 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. Parameter counts include all learned modules loaded for synthesis.
The release reports community listening, multi-ASR intelligibility, predicted naturalness, model footprint and CPU throughput separately.
Harbor is my open-source browser agent extension. It reads the page through visual and DOM context, uses browser tools to act, and streams its progress into the side panel. You choose the model and provider.
See the page. Screenshots and structured DOM context help the agent understand what's in front of it.
Work with the browser. Tools for navigation, tabs, clicking, typing, forms and extracting information.
Follow the process. Streaming responses and tool activity show the steps as they happen.
Explore Harbor → · Browse the implementation
The part I'm interested in: connecting a model to a real interface.
A browser agent has to work across page context, model responses, tool execution and a UI that keeps the user informed. Harbor brings these together in a TypeScript / React extension: content scripts read and interact with pages, the background coordinates requests and browser tools, and the side panel displays the conversation and tool progress.
Sleeve — explore listening eras, lasting favorites, changes in taste and discoveries from your music history. Bring a Last.fm profile or imported listening files. Open the app →
Counterplot — run parameter studies, inspect classifier errors and compare scenarios with a baseline. Export an interactive experiment as one HTML file someone else can explore offline. Try the gallery →
RareSpeech — a research toolkit for rare-term TTS evaluation beside WER. Fixed specialist-word tracks, target-level alignment and original recordings let you inspect what an aggregate score misses. Automatic retention is an ASR proxy; pronunciation validation is ongoing. Inspect the evidence →
discord-bot-kit — build and test Discord bots locally. Run commands and components against the actual handlers, inspect state changes, and turn sessions into test scenarios. Includes recipes, CLI + MCP tools for coding agents. Try the ticket workflow →
Mostly Python · PyTorch · ONNX Runtime · CUDA · TypeScript · React.
Tried a model, found a strange failure, or built something with one of these? I'd like to hear about it.
owen.aw.song@gmail.com · Discord blue.os