older amazon echo dots and shows really make great hardware for this, combined with homeassistant, and can be had for very little cost. e.g. shameless plug: https://github.com/ygelfand/echolocal
In short it (optionally) keeps the excellent Amazon wake word detection, beam forming, and transcription pipeline, but firewalls off all cloud access and redirects to your HomeAssistant voice pipeline.
Oh man, that's really interesting. I have an Echo speaker hidden in a drawer somewhere because one of the grandparents got it for my son and it was incredibly annoying. Using the hardware for something good (and not having Amazon listening to the room at the drop of a hat) sounds like a good weekend hack.
I'm using this as well for an echo dot I bought on eBay for $15 + Hermes integration in Home Assistant to give instructions to my Hermes agent. Works very well, the first time in my life where i'm getting to enjoy a voice assistant without the nagging worry of all my conversations being sent off to Google/Amazon/Apple
I got an echo show rooted and ROMd that I am loving - put it in the kids room as a music player in kiosk mode, but have been now wanting one for myself with wakewords. This looks like the ticket.
the gen 1 or 2 show 8s make great devices for this kind of stuff, 5's are a little small for me. the lenovo android things (google homes) are also really nice if you want a slightly bigger screen, but not as cheap second hand.
Interesting write up, I've been vibe coding an android voice assistant using a similar pipeline without a wakeword - VAD > ASR > LLM > TTS
I've found parakeet to be much, much better than whisper and I made a 4 bit quant of parakeet 110M for my crappier android test devices, it would probably fit a raspberry pi:
Large is feasible on one core of a pi 4, but not in low latency streaming. All will work well on a 5.
I have built exactly the pipeline you're talking about with this model and gemma e4b + mtp. Works entirely on CPU at very reasonable speeds. With an LLM you can even use lighter models and it has negligible impact.
The model is actually a distillation of a different checkpoint called parakeet-ultra, same size and architecture as parakeet-tdt-0.6b-v3 but better, that's why the 1.58bit quant works so well.
I am using the old whisper.cpp to turn voice into text, and Kokoro to read text out. Works pretty well for me, but Kokoro still sounds AI and whisper is not that accurate. Will try Parakeet and Silero
Interesting read! I have my own local voice assistant and I implemented exactly the same question mark trick just the other day.
The overall pipeline is quite similar to my previous setup, but I found the STT -> LLM -> TTS pipeline just infuriatingly slow. TTS and STT with whisper/piper worked fine and reasonably fast for me locally, but the real bottleneck was always the fact that they're not running concurrently with the LLM call. I ended up forking an open source openai realtime API ESP32 plugin and patching it to my liking. It's a bit more expensive in terms of API cost but I only use it sparingly and it's so much nicer to use now.
The lack of concurrency is exactly what makes the traditional pipeline feel slow; Realtime avoids waiting for separate STT, LLM, and TTS stages. I built voicepe-realtime (https://github.com/TristanBrotherton/voicepe-realtime) for the Home Assistant Voice PE, using OpenAI Realtime with barge-in and custom wake words.
Oh you’re very welcome. I’m glad you like it - I welcome pull requests! I’ve made a lot of improvements in speed recently, I’d love to know if I’ve missed anything
Home assistant doesn't really provide much out of the box, its mostly a modular/pluggable ecosystem and one of the things you can install are various voice assistants
I used microwakeword too, and training my own wakeword was pretty simple (basically just recording a few tens of samples). Highly recommended.
I also wanted to be able to issue some voice commands for lights and other things in my home with low latency, so I added a short-circuit plugin that runs right after recognition. If the recognized command matches one of the regexes, it's immediately executed, which has a latency of under a second. Otherwise, it gets sent to my assistant, which usually will tweak the regex (or add a new one) to improve the whole system as it fails.
Beam forming microphone array, custom UI, custom case. Maybe costs more than a second hand echo dot though (hardware is probably around $30).
In short it (optionally) keeps the excellent Amazon wake word detection, beam forming, and transcription pipeline, but firewalls off all cloud access and redirects to your HomeAssistant voice pipeline.
I've been pretty impressed!
I've found parakeet to be much, much better than whisper and I made a 4 bit quant of parakeet 110M for my crappier android test devices, it would probably fit a raspberry pi:
https://huggingface.co/mewse/parakeet-tdt_ctc-110m-onnx-int4
https://huggingface.co/spaces/futo-org/asr4all-client-side
Large is feasible on one core of a pi 4, but not in low latency streaming. All will work well on a 5.
I have built exactly the pipeline you're talking about with this model and gemma e4b + mtp. Works entirely on CPU at very reasonable speeds. With an LLM you can even use lighter models and it has negligible impact.
The model is actually a distillation of a different checkpoint called parakeet-ultra, same size and architecture as parakeet-tdt-0.6b-v3 but better, that's why the 1.58bit quant works so well.
The overall pipeline is quite similar to my previous setup, but I found the STT -> LLM -> TTS pipeline just infuriatingly slow. TTS and STT with whisper/piper worked fine and reasonably fast for me locally, but the real bottleneck was always the fact that they're not running concurrently with the LLM call. I ended up forking an open source openai realtime API ESP32 plugin and patching it to my liking. It's a bit more expensive in terms of API cost but I only use it sparingly and it's so much nicer to use now.
https://m.youtube.com/shorts/3FseVyAG5uM
I used microwakeword too, and training my own wakeword was pretty simple (basically just recording a few tens of samples). Highly recommended.
I also wanted to be able to issue some voice commands for lights and other things in my home with low latency, so I added a short-circuit plugin that runs right after recognition. If the recognized command matches one of the regexes, it's immediately executed, which has a latency of under a second. Otherwise, it gets sent to my assistant, which usually will tweak the regex (or add a new one) to improve the whole system as it fails.