HD Transcribe
A DeepSpeech/TensorFlow speech transcription research project and web app for improving communication for people with Huntington's Disease.
- Repo: https://github.com/Kevin-Liu-01/HD-Transcribe
- Live site: https://hd-transcribe.vercel.app
- Publication / archive: IYRC paper archive and Pre-Collegiate Global Health Review article.
- Core result: fine-tuning a pre-trained DeepSpeech model on Huntington's Disease voice data improved speech recognition accuracy to 85% / WER around 0.15 in the reported test setting. Source: raw/articles/iyrc-kevin-liu-hd-transcription-2026-06-25.md; raw/articles/pghr-hd-transcribe-2026-06-25.md
What it is
HD Transcribe is both a research project and a deployed accessibility app. The research fine-tuned Mozilla DeepSpeech 0.9.3 with Huntington's Disease patient voice data using checkpointing. The application provided microphone recording and audio upload for real-time transcription, with quality-of-life features including ChatGPT 3.5 integration for health questions and text-to-speech output. Source: raw/articles/pghr-hd-transcribe-2026-06-25.md
The GitHub repo metadata describes the implementation as a live speech-to-text transcription interface for people with Huntington's Disease, built with TypeScript, Next, NextAuth, React, Tailwind, DeepSpeech, TensorFlow, accessibility, and speech-to-text topics. Source: raw/github/kevin-repos-2026-06-25.json
Research method
The study started from a pre-trained DeepSpeech model, prepared Huntington's Disease voice clips from a CHDI Foundation dataset, converted audio to mono / 16kHz, and trained with TensorFlow on Ubuntu 20.04 using GPU acceleration. The dataset reported 650 clips and roughly 2 hours of audio, split into training, validation, and test sets. Source: raw/articles/pghr-hd-transcribe-2026-06-25.md
The training ran 7 epochs, with each epoch taking about 6 hours on an NVIDIA GeForce GTX 1660 Ti. Reported WER improved from about 0.97 early in training to around 0.15 on the test set, which the article frames as roughly 85% correct transcription. Source: raw/articles/pghr-hd-transcribe-2026-06-25.md
App implementation
The web application used TypeScript, Next.js, React, Tailwind, the DeepSpeech JavaScript API, OpenAI/ChatGPT 3.5, and React-TTS. It was hosted on Vercel and designed to be accessible from phones, tablets, and desktop browsers. Source: raw/articles/pghr-hd-transcribe-2026-06-25.md
Why it matters
This is Kevin's strongest early research/accessibility artifact. It combines a real disease-context problem, model training, data preparation, a public web interface, publication/archive proof, and a clear product goal: make speech impairment easier to communicate through. It should be listed with GSET Battery Thermal Conductivity Research and CoTCodec in research/project contexts. Source: raw/articles/iyrc-kevin-liu-hd-transcription-2026-06-25.md; raw/articles/pghr-hd-transcribe-2026-06-25.md
Limits
The study itself notes the small dataset: roughly 2 hours of HD speech, with only part available for training/validation. Do not overstate it as a production medical model. Position it as a research prototype and accessibility app with promising reported results. Source: raw/articles/pghr-hd-transcribe-2026-06-25.md
Current Evidence (2026-07-01)
This page was refreshed against Kevin's live GitHub inventory, pinned repositories, PortfolioMon cards, README scrape, and downloaded portfolio screenshots. Treat this section as the current evidence anchor; deeper interpretation stays in the surrounding compiled truth.
| PortfolioMon card | Types | Site | Screenshot |
|---|---|---|---|
| #028 HD Transcribe | AI/Health | site | ![]() |
- PortfolioMon card #028; AI/Health; HP 325, ATK 95, DEF 100, SPD 75; standard card. Speech transcription tool trained on the unique vocal patterns of Huntington's Disease patients. Custom acoustic model improves recognition accuracy for dysarthric speech — bridging a critical accessibility gap in existing STT systems.
| GitHub repo | Visibility | Language | Stars | Last push | Homepage |
|---|---|---|---|---|---|
| Kevin-Liu-01/HD-Transcribe | private | TypeScript | 1 | 2023-03-15 | site |
Timeline
-
2026-07-01 | Project graph refresh checked this page against Kevin's live GitHub inventory, pinned repos, scraped PortfolioMon cards, READMEs, and portfolio screenshots; current evidence was added above. Source: raw/github/kevin-repos-2026-07-01.json; raw/github/kevin-pinned-repos-2026-07-01.json; raw/sites/kevinliu-biz-2026-07-01/portfolio-projects.json
-
2023-01-01 | HD-Transcribe repo created. Source: raw/github/kevin-repos-2026-06-25.json
-
2023-12-11 | Pre-Collegiate Global Health Review published Kevin's article, "Deep Neural Networks for Enhanced Speech Transcription: A Case Study on Huntington's Disease." Source: raw/articles/pghr-hd-transcribe-2026-06-25.md
-
2026-06-25 | IYRC archive and PGHR article captured into raw sources; live deployment verified at https://hd-transcribe.vercel.app. Source: raw/articles/iyrc-kevin-liu-hd-transcription-2026-06-25.md; raw/github/deployment-check-2026-06-25.tsv
