Fixing pronunciation in a specific voice for my language

mdimjasevic · February 18, 2025, 10:38am

Hi everyone!

I’ve started using Home Assistant Voice Preview Edition about 10 days ago. It’s fun to control my home with voice!

I use Croatian as my language of choice to control my home as that’s my native language. I want to contribute to improving support for Croatian in a number of ways, including dusting off the language in Mozilla’s Common Voice for the speech-to-text part.

What I want to ask here is how to fix how a voice (it’s called “GabrijelaNeural” in voice debug logs) for its pronunciation of the Croatian equivalent “Timer set for 3 minutes”? The form of “3” when pronounced in Croatian should have been “tri” instead of “triju” in this specific context, which also affects how the rest of the sentence (and pronunciation) should go. So, instead of “Tajmer postavljen na triju minuta”, it should have been “Tajmer postavljen na tri minute”.

Initially I looked at GitHub - home-assistant/intents: Intents to be used with Home Assistant , but my understanding so far is that this improvement is not to be done there, but in the voice itself, wherever it comes from.

Where is the “GabrijelaNeural” voice to be found and how to fix this issue?

EDIT: I looked some more into this and I have the same “3” issue with another voice (“SreckoNeural”) so perhaps this is not about the voice itself.

How to identify the source of this issue?

Mayhem_SWE · February 18, 2025, 11:31am

Considering you seem to be using different words, this seems like it doesn’t have anything to do with pronunciation but with grammar/sentence generation.

mdimjasevic · February 18, 2025, 1:02pm

Thanks!

Would you know where to find a root cause, i.e., where should I look to find how sentences are generated?

mchk · February 18, 2025, 1:41pm

Is the bug reproduced on this site?
If so, it is on the Azure side.
If the problem is critical, try using a different tts provider

arturpragacz · February 18, 2025, 2:57pm

github.com/home-assistant/intents

responses/hr/HassStartTimer.yaml

33f2a622d

---
language: hr
responses:
  intents:
    HassStartTimer:
      default: >
        {% set h = slots.hours if slots.hours is defined else none %}
        {% set m = slots.minutes if slots.minutes is defined else none %}
        {% set s = slots.seconds if slots.seconds is defined else none %}
        {% set h_text = h ~ (' sat'  if h in [ "1", 'jedan'] else ' sata') if h else '' %}
        {% set m_text = (30 if m in ['pola', 'pol', '1/2'] else m) ~ (' minuta' if m in [ "1", 'jedna'] else ' minuta') if m else '' %}
        {% set s_text = (30 if s in ['pola', 'pol', '1/2'] else s) ~ (' sekunda' if s in [ "1", 'jedna'] else ' sekundi') if s else '' %}
        {% set text_list = [ h_text, m_text, s_text] | select() | list %}
        {% set text = text_list[:-1] | join(', ') ~ ' i ' ~ text_list[-1] if text_list | count > 2 else text_list | join(' i ') %}
        {% set name = (' nazvan ' ~ slots.name | trim) if slots.name is defined else '' %}
        Timer postavljen na {{ text }}{{ name }}
      command: >
        {% set h = slots.hours if slots.hours is defined else none %}
        {% set m = slots.minutes if slots.minutes is defined else none %}
        {% set s = slots.seconds if slots.seconds is defined else none %}

This file has been truncated. show original

mdimjasevic · February 23, 2025, 1:32pm

Thank you! Indeed, the problem is in how a response is generated.

I’ve submitted a fix for this so hopefully it gets incorporated: