voicebm-ha — the VoiceBM integration for Home Assistant

A Home Assistant integration that adds speaker identity to your voice pipeline — who is speaking, on the same audio your existing Whisper already transcribes. It installs via HACS and is backed by an engine add-on that runs the voice biometrics.

Built on VoiceBM by @cybericebyte. All of the voice-biometrics engine — speaker embeddings, gallery matching, enrollment, the whole identity core — is @cybericebyte's original work, released under the MIT License (© 2025 cybericebyte). This project does not claim that work. It is a standalone Home Assistant packaging of it, so VoiceBM installs cleanly via HACS and runs as a first-class part of a Home Assistant voice pipeline. Both copyright notices are retained in LICENSE, as MIT requires. Please star and support the original project.

See DIRECTION.md for why this repo is integration-first and where it's headed.

The integration, and the engine behind it

The product is the integration. It's what you install, configure, and see:

• custom_components/voicebm/ — the HA integration (HACS-installable). It registers as a Speech-to-text engine that transcribes through your existing Whisper and adds speaker identity on the same audio, and it provides a native VoiceBM Speakers sidebar panel (rename / merge / delete / enroll), served through Home Assistant so it works over HTTPS with no separate port.

Behind it runs one dependency:

• voicebm-engine-addon/ — the engine add-on. It runs the original VoiceBM biometrics; it lives in its own container because the ML needs a glibc ONNX runtime, which can't run inside HA Core's Alpine environment. You install it once, and the integration configures and starts it for you (it self-heals the add-on into integration mode over the Supervisor, the same companion pattern Z-Wave JS and Matter use).

The two communicate over MQTT + a shared /share/voicebm folder. Transcription never depends on identity: if the engine is down you still get Whisper's transcript, so adding VoiceBM as your STT engine can't break your voice pipeline.

Install

1. The integration (HACS)

1. HACS → Integrations → ⋮ → Custom repositories → add https://github.com/sam3gp8/voicebm-ha, category Integration.
2. Install VoiceBM (Whisper + Speaker Identity) → restart Home Assistant.

(Or copy custom_components/voicebm/ into config/custom_components/ manually.)

2. The engine add-on it needs

Copy the voicebm-engine-addon/ folder into /addons/ on your HAOS box so /addons/voicebm/config.yaml exists, then Settings → Add-ons → Add-on Store → ⋮ → Check for updates → install VoiceBM engine from Local add-ons. The folder is flat (config.yaml, Dockerfile, rootfs.tar.gz, docs) so any copy method works. First start builds the image on-device (needs internet, a few minutes).

Prefer an add-on repository? Point Settings → Add-ons → Add-on Store → ⋮ → Repositories at https://github.com/sam3gp8/voicebm-ha and install from there.

3. Wire it up

1. Settings → Devices & Services → Add Integration → VoiceBM (auto-picks your Whisper if it's the only other STT engine).
2. Settings → Voice assistants → your pipeline → Speech-to-text → VoiceBM (Whisper + Identity).
3. Speak a couple of commands, then open VoiceBM Speakers in the sidebar → Pending voices → name yourself. Done.

Full setup, GPU notes, and the JARVIS / Ollama integration are documented in voicebm-engine-addon/DOCS.md and custom_components/voicebm/ READMEs.

Prerequisites

• MQTT configured in Home Assistant.
• A working Whisper add-on (VoiceBM transcribes through it — it isn't replaced).
• Home Assistant OS or Supervised (the engine add-on needs the Supervisor).

Credit & license

Original VoiceBM: @cybericebyte — https://github.com/cybericebyte/VoiceBM. Licensed MIT (© 2025 cybericebyte); see LICENSE, retained from upstream. This fork's added HA integration and add-on packaging are provided under the same MIT terms.