saga-2-m · Danish Speech Recognition

capacit-ai/saga-2-m — Danish ASR that pairs a RoPE-patched Whisper-large-v3-turbo encoder with a Qwen3-0.6B decoder through a frame-stacking projector (1.24B params, bf16). Output is cased and punctuated Danish.

Upload a file or record from your microphone — any format librosa reads. Clips longer than 30 s are split at silences into ≤ 28 s pieces and re-joined automatically. saga-2-m is ranked #1 on the Danish ASR Leaderboard (mean 10.68 WER / 5.10 CER across five Danish benchmarks).

Example clips come from the FLEURS da_dk test split (CC-BY-4.0), used with attribution.