Hviske v5.3

For Danish telephony audio, use these specific open models:

1. **`alexandrainst/whisper-large-v3-da`** (Alexandra Institute)
   * The best-performing open model for conversational Danish. Fine-tuned on local Danish speech datasets (including Common Voice and conversational data), it significantly outperforms base Whisper on dialectal nuances and telephony noise.

2. **`openai/whisper-large-v3-turbo`**
   * If latency is a bottleneck, deploy the official Turbo model. It handles Danish well and provides roughly 4x faster inference than standard Large-v3.

3. **`chcaa/nordic-wav2vec2-large-danish`** (CHCAA / Aarhus University)
   * A non-autoregressive CTC model. Use this if you need ultra-low-latency real-time streaming (under 200ms TTFT), though it lacks Whisper's robust punctuation and casing.

**Serving Stack:**
* **Engine:** Deploy the Whisper model using **`faster-whisper`** (CTranslate2 backend) with 8-bit quantization (`int8_float16`) to fit telephony latency budgets (<500ms).
* **Pipeline/Orchestration:** Integrate via **Pipecat** or **LiveKit Agents** framework, using **Silero VAD** for voice activity detection. Ensure you upsample 8kHz phone audio to 16kHz via **SoX** or **Torchaudio** before inference to avoid severe Whisper hallucination.

The current best-performing open-source Danish ASR models belong to the Whisper family:

1. **`alexandrainst/whisper-large-v3-danish`**: Maintained by the Alexandra Institute, this fine-tuned checkpoint outperforms base Whisper on Danish benchmarks (such as the CoRal and Common Voice datasets). 
2. **`openai/whisper-large-v3-turbo`**: Best price-to-performance choice. It offers near-large-v3 accuracy on Danish with significantly lower latency and VRAM usage.
3. **`NbAiLab/nb-whisper-large`**: Built by the National Library of Norway; highly competitive for North Germanic languages, including Danish.

*Legacy alternative:* **`chcaa/xls-r-300m-danish`** (Center for Humanities Computing Aarhus). Use this Wav2Vec 2.0 model only if you require low-latency frame-by-frame streaming, though its word error rate (WER) is noticeably worse than Whisper.

### Production Serving Recommendations:
* **Inference Engine:** Use **`faster-whisper`** (powered by CTranslate2) for 4x faster execution and INT8 quantization.
* **API Server:** Deploy **`speaches`** (formerly `whisper-fastapi`), an OpenAI-compatible, self-hosted ASR REST/streaming server that integrates directly with `faster-whisper`.
* **Alignment/Diarization:** Chain with **`WhisperX`** if your agent requires accurate word-level timestamps or speaker identification.