Home Assistant app providing multi-model speech-to-text — Whisper, Parakeet, SenseVoice, Qwen3-ASR, and more on a single transcribe.cpp (GGUF) runtime. See the transcribe.cpp supported-models table and its per-model cards under docs/models/ for the full list of built-in models.
- Runs on your hardware — no cloud, no API key, no per-minute bill.
- Every model family on one runtime — Whisper, Parakeet, SenseVoice, Qwen3-ASR and more as GGUF on a single transcribe.cpp engine.
- One STT entity per downloaded model, so a pipeline can pick the right one per language or per use case.
- WebSocket streaming — audio reaches the server while you are still speaking, so decoding overlaps capture.
- Measure before you choose — run one recording through up to three models side by side, and score candidates against transcripts you typed yourself.
- Discovered by Home Assistant through the Supervisor, so the companion integration needs no URL or key typed in.
Dashboard — throughput, latency and real-time factor measured on this machine over the last 24 hours.
Transcribe — run one recording through up to three models and compare what each returned, and how long it took.
Models — the catalog plus what is on disk, with the p50 and run count measured from this deployment's own history.
History — every transcription, with its audio, waveform and timings; select rows to delete a batch of them.
Evaluation — score candidate models against hand-typed references: accuracy you judged, inference time, real-time factor, memory and load time.
The same run two other ways — every output beside the reference it was scored against…
…and as a grid, hardest sample first and strongest model first, where the number is how many models got that sample right.
See cortex-stt/DOCS.md for full install, configuration, discovery, and troubleshooting instructions.
Heads-up for Proxmox VE / KVM users. The pre-built binary's bundled ggml inference kernels require AVX + AVX2 + FMA + F16C + BMI2 + SSE 4.2 (Intel Haswell 2013+ /
x86-64-v3). PVE's defaultqemu64/kvm64/x86-64-v2-AESCPU types mask AVX/AVX2/FMA from the guest — change the HAOS VM's CPU Type tohost(orx86-64-v3) and cold-boot the VM (reboot is not enough). The addon's init oneshot detects the missing flags and prints a readable diagnostic instead of crash-looping. Full steps in DOCS.md.
Issues and pull requests are welcome.
cortex-stt/CONTRIBUTING.md— dev setup, the checks, and the PR flow.AGENTS.md— the module tree, cross-module guarantees and the API endpoint reference.cortex-stt/CONTEXT.md— the domain vocabulary both of those use.cortex-stt/docs/adr/— the decisions behind the runtime, the vendored catalog and the evaluation store.
- transcribe.cpp — the single GGUF/ggml runtime powering every model family.
- handy — the desktop dictation app whose runtime + model catalog this project is built on.
MIT — see LICENSE.md.