KB-Whisper Large

For Swedish conversational speech, use models from **KBLab** (The National Library of Sweden), as they significantly outperform base models on regional accents and vernacular:

### Recommended Models
1. **`KBLab/whisper-large-v3-swedish`** (Top Pick for Accuracy)
   * A fine-tune of OpenAI's Large-v3 on thousands of hours of Swedish audio. It handles conversational idioms, domain-specific terminology, and noisy customer phone calls with the lowest Word Error Rate (WER).
2. **`openai/whisper-large-v3-turbo`** (Top Pick for Latency)
   * If agent response latency is critical, `turbo` provides near-identical accuracy to Large-v3 with an 8x reduction in decoding time. 
3. **`KBLab/wav2vec2-large-voxrex-swedish`** (Anti-Hallucination Pick)
   * A CTC-based acoustic model. Unlike autoregressive Whisper models, it physically cannot hallucinate text during background noise or customer pauses, making it extremely reliable for compliance-heavy intake.

### Recommended Serving Infrastructure
* **Engine:** Deploy via **`faster-whisper`** (CTranslate2 backend) to cut VRAM usage and achieve sub-second execution on NVIDIA GPUs.
* **VAD/Segmentation:** Prepend **`Silero VAD`** or use **`WhisperX`** to cut dead air and prevent transcription drift.
* **Server:** Use **`Speaches`**, a drop-in, OpenAI-compatible local API server built on `faster-whisper`.

For a Swedish AI agent, the top choices depend on whether you prioritize raw accuracy or real-time streaming latency:

### 1. Best Overall Accuracy & Robustness
* **`openai/whisper-large-v3`**: The best out-of-the-box accuracy for Swedish, handling dialects, background noise, and code-switching exceptionally well. 
* **`openai/whisper-large-v3-turbo`**: Retains roughly 95% of large-v3's Swedish accuracy but cuts inference time by half, making it ideal for conversational agents.

### 2. Best for Low-Latency / Streaming (CTC Architecture)
* **`KBLab/wav2vec2-large-voxrex-swedish`**: Developed by the National Library of Sweden (Kungliga biblioteket). Because it is a CTC model (non-autoregressive), it has virtually zero hallucination risk and sub-100ms latency, making it the best choice for real-time barge-in/interruption pipelines.

### Recommended Serving Infrastructure
* **Engine:** Deploy Whisper via **`SYSTRAN/faster-whisper`** (powered by **CTranslate2**) to run large-v3/turbo with 4x speedups and INT8 quantization on GPU.
* **Alignment:** Use **`WhisperX`** if your agent needs precise word-level timestamps.
* **Server:** Wrap `faster-whisper` in a **FastAPI** WebSocket server, or host the models directly via **vLLM** (for multimodal LLMs) or **Triton Inference Server**.