The era of open voice assistants

Perceived decline of big-tech assistants

  • Many commenters report Alexa/Google Home getting slower and less reliable (timers, music, basic commands), especially compared to current LLMs.
  • Others say their devices still work fine, highlighting inconsistent experiences.
  • Several note these products generate little profit; some argue this explains lack of investment and layoffs, others say “rich companies” could afford to fix them but choose not to.

Cloud vs local economics and architecture

  • One view: using LLMs for all Alexa requests would be financially impractical at Amazon scale; GPU-heavy cloud workloads don’t pay for a “free” product.
  • Counterview: putting an expensive GPU in every home is wasteful (idle most of the time); centralized GPUs plus subscription make more sense.
  • A middle position: scaling AI infra to millions is hard either way; dedicated efficient SoCs (Apple Silicon–style) may eventually make local AI practical.

Home Assistant Voice device: role and hardware

  • Device is positioned as open, privacy-preserving “satellite”: mic + speaker + wake word on ESP32-S3 plus XMOS audio processing, connecting to a separate Home Assistant server.
  • XMOS is valued for beamforming/noise reduction so wake-word and STT work even with music or distance.
  • Users like that it has audio out and Grove connector; some expect to pair it with better speakers.
  • Sold out quickly in many regions; some already ordered many units to replace Echos.

Software stack, LLMs, and extensibility

  • Voice pipeline is modular: wake word, STT, intent/LLM, TTS can each be local or cloud (Whisper/faster-whisper, Piper, Coqui, Ollama, OpenAI, etc.).
  • Assist can first try structured “home control” intents, then fall back to a general LLM for arbitrary questions.
  • ESPHome is the main SDK; firmware and case design are open, with expectation of forks and custom hardware variants.

Privacy, openness, and ecosystem

  • Strong enthusiasm for a fully local, open, auditable alternative to Amazon/Google; many explicitly cite distrust of corporate data practices.
  • Some want to avoid any cloud use; others are fine with Nabu Casa cloud to support development and offload heavy workloads.
  • Comparisons: Home Assistant is seen as more capable and community-rich than openHAB; Mycroft is cited as an earlier, ill-fated attempt whose ideas and people partially live on here.

Concerns, trade-offs, and open questions

  • Some report past HA voice pipelines as unreliable; others find them powerful but complex to set up.
  • Worries include: HA’s weak fine-grained security model, lack of standard auth (OIDC), UI-over-YAML trend being “anti-engineer,” and unclear docs around running advanced models on user GPUs.
  • Multilingual quality and music streaming remain pain points; HA is actively crowdsourcing language support, while music depends on external provider integrations.
  • Debate over “no wake word” assistants raises technical and UX challenges (false triggers, constant compute).