jeans/home-assistant.io
jeans/home-assistant.io
1
0
Fork 0
mirror of https://github.com/home-assistant/home-assistant.io.git synced 2025-07-28 03:37:06 +00:00
Code Issues Packages Projects Releases Wiki Activity

2024.12: Iterate on the voice section

Browse Source
This commit is contained in:
Franck Nijhof 2024-12-04 16:16:00 +01:00
parent 5179fe84b7
commit e5fc70bfe4
No known key found for this signature in database
GPG Key ID: D62583BA8AB11CA3
1 changed files with 28 additions and 29 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

57
source/_posts/2024-12-04-release-202412.markdown
View File

@ -142,19 +142,20 @@ But before we get there, let's dive into the features that ship in this release!
### Let your voice assistant fall back to a LLM-based agent
[Exactly 6 months ago] we bridged the gap between our intent world and the
wonderful (and scary) world of LLMs. We allowed you to let an LLM agent control
your home instead of relying on our built-in intents.
[Exactly 6 months ago] we bridged the gap between our default conversation agent
and the wonderful (and scary) world of [LLMs]. We allowed you to let an
LLM-based conversation agent control your home instead of relying on our
built-in sentences.
This choice was an interesting first step. LLMs are generally much smarter and
more knowledgeable about the world than our built-in intents. However, they are
more knowledgeable about the world than our default agent. However, they are
often slow and/or expensive. And let’s face it: Even if the demos are cool,
90% of the commands we say in our homes are simple:
_“Turn this on”_ or _“Turn that off”_.
_“Turn this on”_ or _“Turn that off”_.
Today, we're finally allowing users to **mix** these worlds. Starting from this
release, you can set up a voice assistant in a way that will target our fast,
community-driven, built-in intents first, and only fallback to an LLM-based
release, you can set up a voice assistant that targets our fast,
community-driven conversation agent first and only falls back to an LLM-based
agent if no matches are found.
<img class="no-shadow" alt="Dialog showing the Assist pipeline configuration, showing the new option to prefer handling commands locally." src="/images/blog/2024-12/llm-fallback.png"/>
@ -167,45 +168,43 @@ does not natively understand.
<lite-youtube videoid="vThoxRIxHyI" videotitle="Let your voice assistant fall back to a LLM-based agent" posterquality="maxresdefault"></lite-youtube>
[Exactly 6 months ago]: https://www.home-assistant.io/blog/2024/06/05/release-20246/
[Exactly 6 months ago]: /blog/2024/06/05/release-20246/
[LLMs]: https://en.wikipedia.org/wiki/Language_model
### Language leaders are accelerating the pace
It is not a secret: our voice hardware will come soon. Language leaders already
got their hands on the device, and are working extremely hard to polish support
for their language.
It is not a secret: our voice hardware will be revealed on 19 December 2024.
Language leaders have already gottten their hands on the device, and are working
extremely hard to polish support for their language.
The amount of contributions on our [intents repository] (where we store the
supported sentences) skyrocketed during the last month. It is something we all
truly appreciate. More and more languages are becoming usable, or even complete!
You can follow the progress [here](https://home-assistant.github.io/intents/).
The number of contributions to our [intents repository] (where we store the
supported sentences) skyrocketed during the last month, which we all truly
appreciate. More and more languages are becoming usable or even complete!
You can follow the progress [here](https://home-assistant.github.io/intents/).
We won't list them all here, but just be aware that it is very likely that
someone is hard at work making sure your native language gets some love so
that you can speak with your home.
you can speak with your home.
[intents repository]: https://github.com/home-assistant/intents?tab=readme-ov-file
### A faster voice experience
[HASSIL], our intent parser, was built from the ground up to match a sentence to
its underlying intention extremely fast and on cheap hardware.
Our default conversation agent was built from the ground up to match a sentence
to its underlying intention extremely fast and on cheap hardware.
Release after release, we added more use-cases and more sentences to
Home Assistant, and our sentence-matching logic became slower and slower. In
some of the languages, a non-matching sentence could take more than 15 seconds
to be processed!
Release after release, we added more use cases and sentences to Home Assistant,
and our sentence-matching logic became slower and slower. In some languages,
a non-matching sentence could take more than 15 seconds to process!
This release adds a lot of love to HASSIL; we completely reengineered the way
we match sentences. On top of that, language leaders spent some time reducing
the complexity of their sentences. Some languages saw a reduction of 99% in the
number of possible combinations!
This release adds a lot of love to this matching logic; we completely
reengineered the way we match sentences. On top of that, language leaders spent
some time reducing the complexity of their sentences. Some languages saw a 99%
reduction in the number of possible combinations!
The results speak for themselves:
<lite-youtube videoid="U_CMXL2Z5NE" videotitle="A faster voice experience with HASSIL" posterquality="maxresdefault"></lite-youtube>
[HASSIL]: https://github.com/home-assistant/hassil
<lite-youtube videoid="U_CMXL2Z5NE" videotitle="A faster voice experience with Our default conversation agent" posterquality="maxresdefault"></lite-youtube>
## Revised Integration Quality Scale