Home Assistant Add-on: ONNX ASR

Configuration Voice Assistant
,
1

Home Assistant Add-on: ONNX ASR

Addon repository

Home Assistant add-on that uses onnx-asr for speech-to-text.

Notably, provides access to the NVIDIA NeMo Parakeet-TDT model which should be significantly faster and more accurate than Whisper for English in most cases.

Supports aarch64 Architecture Supports amd64 Architecture

Faster and better speech to text

This addon provides an English language voice recognition service which is (in theory) both better than the biggest whisper models and nearly twice as fast as even the smallest whisper model! The only drawback is it needs around 2.5GB of RAM.

This addon also supports whisper models, which can be used for other languages. It seems to be slightly faster than wyoming-faster-whisper for some models, particularly whisper-base.

This means it should be a drop-in replacement for most users!

The following benchmarks were performed on my Ryzen 5 5600X, with the English phrase “Turn on the living room lamp.”

Model Size Runtime Model Time
Parakeet wyoming-onnx-asr nemo-parakeet-tdt-0.6b-v2 0.26s
Tiny wyoming-onnx-asr onnx-community/whisper-tiny.en 0.35s
wyoming-faster-whisper tiny 0.4s
wyoming-faster-whisper tiny-int8 0.49s
wyoming-onnx-asr onnx-community/whisper-tiny 0.5s
wyoming-faster-whisper Systran/faster-distil-whisper-tiny.en 0.59s
Base wyoming-onnx-asr whisper-base 0.6s
wyoming-onnx-asr onnx-community/whisper-base.en 0.76s
wyoming-faster-whisper Systran/faster-whisper-base 0.82s
wyoming-faster-whisper Systran/faster-whisper-base.en 0.94s
Small wyoming-faster-whisper Systran/faster-distil-whisper-small.en 1.4s
wyoming-onnx-asr onnx-community/whisper-small.en 3.4s
Large wyoming-onnx-asr onnx-community/whisper-large-v3-turbo 8.1s
wyoming-faster-whisper Systran/faster-distil-whisper-large-v3 10s

About

The addon source can be found in onnx-asr-addon, which is an addon version of the wyoming-onnx-asr python module, itself heavily based on wyoming-faster-whisper. This is all made possible through the work of the developer of onnx-asr who ported the parakeet model in the first place.

Installation

This addon can be installed from my repository:
Add repository on my Home Assistant

And read the docs :slight_smile: By default the addon only sets up an english model, but it can be configured with both english and multilingual. Once running, the Wyoming integration should be auto-detected in integrations.

If you’re using Home Assistant Container, wyoming-onnx-asr provides a drop-in replacement for the wyoming-faster-whisper container as well.

4 Likes
2

The library is of particular interest to Russian-speaking users. Since it gives access to a quality local GigaAM model. If anyone is interested, I have already done a similar project (but the translation is only in the addon). At the same time allowing to get sufficient speech recognition speed on n100 cpu. This is also true for parakeet.
It would be nice to help the author add support for canary, it would add a few more languages, but it seems to have problems with it when converting to onnx
You can also manually create onnx for fastconformer (available for several European languages), but no one has done a comparison on test datasets, so it’s not known if this is better than whisper

upd.
Added the multilingual model nemo-parakeet-tdt-0.6b-v3 to my server

upd. 2

Added nemo-canary-1b-v2

1 Like
3

In my test system I installed the addon, configured both the model_en and model_multi to “auto”. The debug logs say the models are downloaded correctly.

Installed the addon for Wyoming, and the STT is viewable. However when I configure a new Voice Assistant and want to select the Onnx-asr as engine, it is not possible. It is greyed out. What am I missing?

4

I can reproduce this, and it looks like an error in how I’m presenting the language codes for the multilingual model.

I’ll try and get a fix out soon.

While trying to troubleshoot I couldn’t find a way to change the voice assistant language of a pipeline: does home assistant even support this? Language is a drop-down on each item, but I can’t see where to add extra languages at the start!

5

I’ve released 0.1.3 which should resolve this, although I’m not sure what will happen if you have both en and multilingual enabled at the moment. I’ll fix that up later today.

6

Cool i’ll try it out. Well mostly I want the multilanguage, so I will disable the EN one :slight_smile: It was more as a test why the setting was greyed out.

Thanks!

7

@dayshine there’s typo in Dockerfile, that prevents Docker container from serving. And there’s PR to fix it. Could you look please?

1 Like
8

I’m manually overriding it in my docker compose right now and I’m not able to get Home Assistant to connect to Wyoming.

9

There was one other issue I’d missed with the docker image’s default args: It was only listening on localhost :slight_smile:

v0.3.5 is publishing now which works for me with no custom command in docker-compose.

Thanks for pointing this out!

Just in case anyone struggles with getting it running, all I have is:

  onnx-asr:
    image: ghcr.io/tboby/wyoming-onnx-asr
    volumes:
      - '/mnt/user/docker-configs/hass/whisper/data:/data'
    restart: on-failure:5
    networks:
      - caddy
    mem_limit: 8G
    memswap_limit: -1
    container_name: onnx-asr

Where the only lines necessary are

  onnx-asr:
    image: ghcr.io/tboby/wyoming-onnx-asr
    volumes:
      - '/mnt/user/docker-configs/hass/whisper/data:/data'
    networks:
      - caddy

i.e.:

  • The image name
  • The model caching volume
  • The network it shares with hass in my setup
2 Likes
10

Github link in first post is incorrect, should be: GitHub - istupakov/onnx-asr: Automatic Speech Recognition in Python using ONNX models

1 Like
11

@dayshine can you add support for nemo-parakeet-tdt-0.6b-v3 in HA addon?

12

If you want to test the model, use my add-on version

13

Yup, I’ll update the library this weekend

14

Hi,
I’ve installed addon, but cannot see in speech to text services.
I clicked your my link to add Wyoming Protocol, but it wants me to enter host and port number.
Where can I get these information?
Currently my Wyoming Protocol shows piper and whisper as services.
Thank you.

15

Show a screenshot of the main page of the running addon.

16

I’ve restarted HASSOS and it works. Thank you.

17

Hey @dayshine got notification for new version, but can’t update plugin.
Error: Failed to perform the action update/install. Error updating ONNX ASR: Can’t install ghcr.io/tboby/onnx-asr/amd64:0.2.0: 404 Client Error for http+docker://localhost/v1.51/images/create?tag=0.2.0&fromImage=ghcr.io%2Ftboby%2Fonnx-asr%2Famd64&platform=linux%2Famd64: Not Found (“manifest unknown”)

18

Same for me. Running a NUC (generic x86-64).

19

Same issue here, won’t update on HAOS

20

Sorry about that, I’ll try and fix it today.

And I’ll also fix my forum notifications to not have a… 3 day delay? :smiley: