We share what we build along the way: a free call recorder for Mac, the speech and language models we tune, and the Lithuanian speech data we prepare to train them. Everything here is free to use under its license.
Software you can install and use today, with the full source code.
A free call recorder and AI meeting note taker for Mac. It records Zoom, Google Meet, Microsoft Teams, FaceTime, WhatsApp and phone calls with no bot joining the call, then gives you transcripts that show who said what.
Large language models we tune and quantize for our own GPU cluster, shared with the setup that serves them.
Qwen3.8-27B on a single RTX 5090 with the full 256K-token context. Each part of the network gets the precision it needs, so it stays closer to the original model at the same memory.
Speech recognition (speech-to-text) models fine-tuned for Lithuanian, with honest results on held-out and real-world test sets.
Lithuanian speech recognition for telephone speech, fine-tuned from NVIDIA Parakeet TDT 0.6B v3 on 426 hours of calls. It writes readable text: punctuation, capitals, and numbers, dates and amounts as digits.
Lithuanian speech datasets built from the openly licensed LIEPA-3 corpus, with transcripts in written form so models learn to output readable text.
429 hours and 417,000 utterances of spontaneous Lithuanian telephone speech. To our knowledge, the largest openly licensed Lithuanian phone speech dataset with written-form transcripts.
100.5 hours and 111,970 utterances of spontaneous dialect speech from 117 speakers in Aukštaitija, Žemaitija, Dzūkija and Suvalkija, with dialect word forms kept as spoken.
Speech technology for smaller languages moves faster when data and models are shared. We publish what we can, including results that show where a model still falls short, so others can build on it and check our work.
