Vocare brings speech AI training in-house on GIGABYTE system

Adelaide startup Vocare is training its pronunciation model locally on a GIGABYTE AI TOP ATOM it won at SouthStart 2026, reducing cloud costs and keeping learner audio on hardware it controls. The setup supports more than 30,000 annotated voice samples as Vocare expands its speech AI for multilingual aged care and disability workers.

Why it matters: - Vocare's in-house setup lets the Adelaide startup train speech AI without hourly cloud fees. - The system keeps annotated workplace voice recordings on hardware Vocare controls. - The company says that matters for privacy, speed and scaling a product used by multilingual care workers.

What happened: - Vocare now trains its pronunciation model on a GIGABYTE AI TOP ATOM it won at SouthStart 2026 and set up with XENON in April 2026. - The startup helps multilingual aged care and disability support workers speak English more clearly. - The model listens as a worker speaks, detects mispronounced phonemes and assigns exercises to improve clarity. - Vocare says the system now serves live pronunciation assessments and is where the team trains and tests new models.

The details: - Vocare was founded in 2025 by linguist Seshnie Taylor and now has more than 225 users. - An initial grant from the Australian Institute for Machine Learning gave Vocare 300 hours with machine learning engineers to build its core code. - When that project ended in March 2026, the code needed a GPU to run, but Vocare had no budget to buy one. - Renting cloud GPU time would have added an hourly charge to every training run. - Vocare trains on a proprietary dataset of more than 30,000 annotated workplace voice samples. - The ATOM has 128GB of unified memory and 4TB of NVMe storage. - XENON Solutions Architect Malcolm Haak worked with Vocare's machine learning and software engineers in April 2026 to bring the ATOM online and set up secure remote access. - Vocare's published 42-day pilot with one provider showed one participant's score rising from 33.8% to 72.2% after repeating a single exercise 72 times. - Taylor said the hardware helped Vocare say it is stored locally and hosted locally, and that it sped the work up. - Linh Ung, Vocare's machine learning engineer, said all training experiments sit on one machine and learner audio never goes to a third-party cloud API.

Between the lines: - Vocare's move in-house reduces reliance on cloud vendors and gives the startup more control over sensitive workplace audio. - The local hardware setup also fits a product built for providers that may care about data handling and turnaround time. - The startup appears to be using the ATOM as both a privacy play and a cost-control tool. - Vocare's focus for providers has shifted toward staff retention as it looks beyond Australia.

What's next: - Taylor says the FoundHer Launch Grant will help Vocare reach more multilingual care workers across aged care and disability support. - Vocare sees potential in New Zealand, Canada and the United Kingdom. - The full deployment story is available on the XENON website.

The bottom line: - Vocare has turned a prize system into core infrastructure for training, testing and running its speech AI locally.