Best local speech-to-text models for Slovenian
For Slovenian today, start with models that either explicitly list Slovenian or belong to a mature multilingual family with established Slovenian support. Do not assume a new multilingual ASR model covers sl unless the source says so.
1. Parakeet TDT 0.6B v3
NVIDIA’s current model card explicitly includes Slovenian. That removes one of the biggest uncertainties in local ASR selection. It is a 600M model with a CC BY 4.0 license, so attribution obligations matter if you use or redistribute it.
2. Whisper
Whisper’s multilingual models remain a practical baseline because there are many local runtimes and model sizes. If your machine is modest, start with small; on a capable GPU, turbo is a strong speed/accuracy option. OpenAI’s current repository gives approximate VRAM guidance of ~2 GB for small, ~5 GB for medium and ~6 GB for turbo.
3. Nemotron 3.5 ASR Streaming
Nemotron is interesting when native streaming is central, but the current NVIDIA card puts Slovenian in the adaptation-ready tier. NVIDIA states that adaptation-ready locales require fine-tuning on in-domain data to enable full transcription. That is materially different from out-of-the-box support.
What about Qwen3-ASR?
Qwen3-ASR is a strong modern family with offline and streaming support, but Slovenian is not in the official supported-language list reviewed for the 0.6B and 1.7B releases. That makes it a poor default recommendation for Slovenian until the official coverage changes or you validate it independently.
Recommended test order
- Parakeet TDT 0.6B v3.
- Whisper turbo on a suitable GPU, or Whisper small on lighter hardware.
- Nemotron 3.5 only if you are prepared to adapt/fine-tune for Slovenian or NVIDIA changes its support tier.
Use the matcher to adapt this order to your hardware.
Reviewed against primary sources on September 13, 2026. Model behavior and licenses can change; re-check the linked model source before production use.