Finnish Voice Issues
# support
d
I am having issues with the way the bot speaks back numerical data in Finnish, specifically phone numbers and dates. Is there a specific package I could add or any model that performs best specifically with Finnish. Thank you in advance!
r
Hi this is usually not a model issue but how the text is formatted before it hits TTS. Finnish struggles with raw numbers, so the fix is to normalize the output by converting numbers, dates, and phone numbers into fully written Finnish before playback. You can also improve it by switching TTS providers, since Finnish quality varies a lot between them. If you want, I can help you set up a clean formatting layer so everything sounds natural. @DenisNMetodiev
d
I suppose I could go with the normalization of the output, so I should just configure it to get the information, before saying it back, save it in its immediate memory, then play back to the customer, correct? @User
s
Hi! Improving Finnish numerical speech mostly comes down to how the text is prepared before it reaches the voice engine. Here is the shortest way to fix it:
1. Prompting (Most Effective) Since automatic formatting is language-agnostic, explicitly instruct your LLM in the System Prompt to format numbers for Finnish speech. Example: "Always write out phone numbers digit-by-digit (e.g., 'nolla nolla...') and dates in full Finnish spoken format (e.g., 'kymmenes toukokuuta')."
2. Adjust the Voice Formatting Plan You can disable the default formatting so it doesn't interfere with your prompt instructions:
Copy code
json
"voice": {
  "chunkPlan": {
    "formatPlan": { "enabled": false }
  }
}
3. Recommended Transcriber For the best accuracy in Finnish, we recommend using Speechmatics. It generally handles Finnish accents and vocabulary better than other providers.
4. Endpointing If the bot interrupts or pauses awkwardly, ensure you are using Transcription-based endpointing, as our "Smart Endpointing" is currently optimized for English.