A free, open speech recognition model for Lithuanian — a Whisper-class alternative tuned on Lithuanian speech to cut errors roughly threefold. Runs fully on-device on Apple Silicon, or as a full research checkpoint.
Word error rate on held-out Lithuanian speech — down from 33.57% for the stock multilingual base. About 3× fewer errors.
| Model | WER (all) | Read | Spontaneous |
|---|---|---|---|
| Multilingual base (stock) | 33.57% | 33.08% | 34.07% |
| Lithuanian fine-tune | 10.87% | 12.74% | 8.99% |
Real, unedited output on held-out Lithuanian audio. On Lithuanian the stock multilingual model often drops out of the language entirely — into Portuguese, Russian, English, or gibberish — while the fine-tune stays in Lithuanian and even writes numbers out as Lithuanian words. It isn't flawless (see the last line), but across the full test set it averages 10.87% WER. Output is lowercase and unpunctuated by design.
Built for Lithuanian, not adapted as an afterthought.
Most multilingual speech models treat Lithuanian as a long-tail language. This is a full fine-tune of NVIDIA's Parakeet-TDT-0.6b-v3 on the LIEPA-3 Lithuanian corpus — read and spontaneous speech from Vilnius University and CLARIN-LT — measured on a held-out test set with Lithuanian diacritics preserved.
If you've been looking for a Whisper-class model for Lithuanian, this is built for exactly that job: accurate Lithuanian ASR, free to use, and small enough to run on-device. Full method, benchmarks, and usage are on the NeMo model card. The on-device CoreML build for Apple Silicon is available separately: download the Lithuanian Parakeet model on Hugging Face.
Released under CC-BY-4.0. Built on nvidia/parakeet-tdt-0.6b-v3 (© NVIDIA, CC-BY-4.0) and the LIEPA-3 corpus (Vilnius University / CLARIN-LT, CC-BY-4.0).
Everything you need to know about Lithuanian speech recognition on-device.
It is a fine-tune of NVIDIA Parakeet TDT 0.6B v3 trained specifically on Lithuanian speech using the LIEPA corpus — a large Lithuanian audio dataset from Vilnius University and CLARIN-LT covering both read and spontaneous speech. The result is a dedicated Lithuanian ASR model, not a general multilingual one.
Word Error Rate (WER) measures the percentage of words a model gets wrong. Lower is better. The stock multilingual Parakeet base achieves 33.57% WER on Lithuanian — meaning roughly one word in three is wrong. This fine-tune achieves 10.87%, a ~68% relative improvement (about 3× fewer errors). On spontaneous Lithuanian speech the improvement is even larger: 8.99% vs. 34.07% for the base.
LIEPA (Lietuvių kalbos parengiamojo periodo garsynas) is a Lithuanian speech corpus maintained by Vilnius University and CLARIN-LT. It includes both read and spontaneous speech recordings with full Lithuanian diacritics preserved, making it the standard benchmark for Lithuanian ASR research.
Yes. The CoreML version runs fully offline on Apple Silicon (M1 and later) using Apple's Neural Engine. It requires no internet connection, sends no audio to a server, and weighs approximately 483 MB. There is no cloud subscription. The full NeMo checkpoint also runs on Linux GPU servers for research use.
Whisnap is a real-time macOS dictation app powered by this model. Press a hotkey, speak Lithuanian, and the transcription is pasted directly into your active application. It uses the on-device CoreML build, so everything stays on your Mac. You can also download the Lithuanian Parakeet CoreML model on Hugging Face and integrate it into your own project.
The model is released under CC-BY-4.0 — free for research and commercial use with attribution. It builds on NVIDIA's Parakeet-TDT-0.6b-v3 (CC-BY-4.0) and the LIEPA-3 corpus (Vilnius University / CLARIN-LT, CC-BY-4.0).
Whisper (multilingual) and stock Parakeet both struggle with Lithuanian as a low-resource language, often switching to Portuguese, Russian, or English mid-transcription. This fine-tune stays in Lithuanian and correctly handles Lithuanian diacritics (ą, č, ę, ė, į, š, ų, ū, ž). It is not based on Whisper but achieves Whisper-comparable accuracy goals while running faster on Apple Silicon via CoreML and the Neural Engine.