**WhisperLiveKit** exists because of a mismatch most people discover the hard way: Whisper was trained on complete utterances, so feeding it two-second chunks in a loop cuts words mid-syllable, loses context and produces transcripts worse than the model's reputation suggests. WLK replaces that naive loop with published simultaneous-speech research and wraps the result in a self-hosted server. Apache-2.0, written in **Python**, it has reached **10,866 stars and 1,131 forks**, with **v0.2.25** tagged **1 August 2026**. The intelligence is in the commit policy rather than the model. Transcription runs on **LocalAgreement** (whisper_streaming) or the **AlignAtt** policy from Simul-Whisper/SimulStreaming, both of which decide when a hypothesis is stable enough to emit instead of resending a churning buffer. Diarization uses **Streaming Sortformer**; simultaneous translation runs through **NLLW**, built on distilled **NLLB**, covering translation from and to 200 languages. Voice activity detection cuts backend work when nobody is speaking, and the server handles multiple concurrent users. Backend choice is the project's real differentiator, because no single ASR model wins everywhere. Whisper is the default, including an **MLX** path on Apple Silicon. **Voxtral Mini**, Mistral's 4B speech model, handles 100+ languages with per-chunk language detection the maintainer notes is less English-biased than Whisper's. **SenseVoiceSmall** via FunASR covers Mandarin, Cantonese, English, Japanese and Korean. **NVIDIA Canary-1b-v2** brings 25 European languages with native word-level timestamps and AmberNet language ID. And **Qwen3-ASR** runs through either plain HF Transformers or vLLM. The Qwen3 path carries the newest engineering. The windowed default re-encodes up to 12 seconds of audio on every update; the **causal mode** instead encodes each ~2-second block exactly once through an append-only causal-KV encoder, so memory is bounded at a 15-second window and per-chunk compute stays constant no matter how long the session runs — the intended answer for many concurrent streams, energy-constrained serving, or sessions with no natural end. It uses a fine-tuned tower published as `qfuxa/qwen3-asr-0.6b-streaming`, and is **English-only for now**, with the windowed default still recommended when accuracy matters most. Integration is designed to require no rewrite. WLK exposes an **OpenAI-compatible REST endpoint** that works with the OpenAI Python SDK against a local base URL, a **Deepgram-compatible WebSocket**, and a native WebSocket with per-session `language`, `target_language`, `mode` and `token` parameters. The `wlk` CLI covers server, one-shot file transcription, SRT subtitle generation, model pull/remove and a built-in benchmark. A Chrome extension captures audio from web pages and a native SwiftUI macOS client ships in-tree. The honest friction is dependency management. Several heavy extras — `voxtral-hf`, the vLLM stacks, `qwen3-vllm-metal` — deliberately conflict and must live in separate environments, with `[tool.uv].conflicts` in `pyproject.toml` as the authoritative list; Diart diarization is pinned below Python 3.13 and not recommended against Sortformer. Word timestamps are interpolated estimates at roughly one-second precision unless you run the vLLM ForcedAligner path, and the maintainer advises planning one realtime session per GPU. Published benchmarks are 6 minutes of LibriVox audio per language measured on an **H100**, reproducible via the included script but not yet broadly replicated on other hardware — the project is openly soliciting results.