Track voices and voice-model usage across the install base of HA

ronnie_j · May 16, 2025, 3:40pm

Integration tracking is done on each installed H.A instance, and the sum of it is shown on each integration page. For example: Piper - Home Assistant show 8.5% of users have this enabled.

This kind of tracking is Wonderfull for loads of reasons: decide where to invest in, improve areas and find weaknesses etc.

Can we also count the used voices used in Piper? (eg. nl_BE_XXX or nl_NL_YYY)?

Thanks!

Edwin_D · May 16, 2025, 6:34pm

Wouldn’t that end up to be pretty much the same as when calculated by the country HA is used in? That information is available already.

ronnie_j · May 17, 2025, 7:54am

Most of the time the selection of a voice, and the country one lives in would be similar, but within each language there could be more variations or dialects (or the lack of it on the other hand).

francisp · May 17, 2025, 8:07am

As he mentioned nl_be, Belgium has 3 official languages, Dutch, French and German. A lot of countries have more official languages, as Switzerland, Canada, Italy etc…

Edwin_D · May 17, 2025, 8:27am

I know countries can be multilingual. but you can infer a lot without being too far off. You want to know French? look up the percentages of population for the countries speaking it, and estimate based on HA usage in those countries.

How actionable will the data be to HA devs? The more detailed data HA analytics gets, and the more data they collect, the more I want to opt out. One NL_nl less. And I think I might not be alone in this, as HA users tend to be privacy minded. The percentage of people opting in to analytics is not high, so expecting precise results is not an option anyway. There’s enough data to get some idea.

And if you have the numbers, then what? You will see higher usage statistics on Piper languages working well, because languages not working well don’t get used. So where would the effort go? To languages not being used? Or will it be like Google News, gravitating into a funnel? And how would one get more/better training data based on the results?

So here’s a tip: contribute to the training data. Did you know you can have piper use your own voice?

github.com/domesticatedviking/TextyMcSpeechy

docs/using_custom_voices_in_home_assistant_os.md

6673bccd8

## How to install and use a custom Piper voice model in Home Assistant OS
I put together this guide because I couldn't find an adequate explanation of how to do this anywhere else.  Please read carefully as there are currently some issues arising from the way Piper is implemented in Home Assistant that make this process more confusing than it should be.

## Install Piper add-on and Wyoming protocol integration
 1. Install the Piper Add-on (Settings > Addons > click ADD-ON STORE button > search for Piper.)
 2. Install the Wyoming protocol integration (Settings > Devices & Services > Integrations > Click ADD INTEGRATION BUTTON > Wyoming Protocol).
 3. When prompted by Wyoming Protocol for host and port, you can use `core-piper` for `host` and `10200` for `port`
    - `core-piper` is the hostname provided by the piper add-on's docker container. You can also use the ip address of the machine running the docker container. 
    - `10200` is the default port for Piper. Change it as needed.  

## Prepare your custom voice models for upload
1. Important: Per the Piper docs, your voices must be renamed according to the following scheme: `<language>_<REGION>-<name>-<quality>`, eg: `en_US-bob-medium`
  - `<language>_<REGION>` is the IETF BCP 47 format language code (eg `en_US`)
  - `name` is the name of the person or character to whom the voice belongs
  -  `quality` is the model quality. Valid values are `x_low`, `low`, `medium`, and `high`
  - rename both your `.onnx` and `.onnx.json` files, eg: `en_US-bob-medium.onnx`,`en_US-bob-medium.onnx.json` Make sure the names match EXACTLY.
  - If you don't name your model in this format, Home Assistant may not recognize it.
     
2. Update the `dataset` field in your `.onnx.json` file to reflect the new name of your model.  This is important if you want your model to show up properly in voice assistant entities created in Settings > Voice Assistants.
Custom piper models found online don't always follow the standards piper expects. The `quality` field is one I have often had to change.  It usually works fine if you set it to "Medium". I am still exploring what is expected by home assistant a bit here.

This file has been truncated. show original

francisp · May 17, 2025, 10:10am

I opted in to statistics to show Supervised is still used. Now they are going to deprecate it anyway, so opted out again.