Edda v0.2

Here are the best permissively licensed options for Danish STT:

### Top Models
1. **`openai/whisper-large-v3-turbo` or `openai/whisper-large-v3`**
   * **License:** MIT
   * **Details:** Best overall zero-shot accuracy for Danish. The `-turbo` variant reduces latency significantly while maintaining high Danish word error rate (WER) performance.
2. **`alexandrainst/whisper-large-v3-danish`**
   * **License:** Apache-2.0
   * **Details:** Created by the Alexandra Institute (the primary Danish NLP research group). It fine-tunes Whisper on Danish corpora (CoRal, NST), significantly improving transcription of native accents and local idioms over base Whisper.
3. **`alexandrainst/da-wav2vec2-large-v2`**
   * **License:** Apache-2.0
   * **Details:** A CTC-based model trained on hundreds of hours of Danish audio. It is significantly faster and uses less memory than Whisper, though it lacks built-in punctuation and formatting.

### Deployment / Serving Stack
* **Runtime:** Use **`faster-whisper`** (MIT License, powered by CTranslate2) or **`whisper.cpp`** (MIT License) for efficient local execution.
* **Server:** Wrap with **`vLLM`** (Apache-2.0, now supports audio inputs) or run natively using the **`Triton Inference Server`** (Apache-2.0) with the `faster-whisper` backend.

For commercial Danish ASR, use models with permissive licenses (MIT or Apache 2.0):

### 1. Dedicated Danish Models (Alexandra Institute / CoRal Project)
* **`alexandrainst/coral-whisper-large-v3`** (Apache 2.0): Fine-tuned specifically on Danish speech corpora (CoRal dataset). Currently the highest-accuracy open model for Danish.
* **`alexandrainst/coral-wav2vec2-large-da`** (Apache 2.0): Acoustic-only Danish model based on Wav2Vec 2.0; lower latency than Whisper, ideal for real-time streaming.

### 2. General Whisper Models
* **`openai/whisper-large-v3`** and **`openai/whisper-large-v3-turbo`** (MIT): Out-of-the-box multilingual support with strong Danish comprehension, fully cleared for commercial use.
* **`chcaa/whisper-large-v3-danish`** (Apache 2.0): A Danish fine-tune from Aarhus University's Center for Humanities Computing.

### Recommended Serving Stack
* **Runtime**: Run Whisper models using **Faster-Whisper** (CTranslate2, MIT) for high-throughput GPU/CPU inference, or **whisper.cpp** (MIT) for minimal-footprint edge deployments.
* **Server**: Deploy via **vLLM** (ASR-supported architectures) or a custom **Triton Inference Server** wrapper.