Improving AI Voice Analysis with Local Accent Datasets

When it comes to building voice-based technology for seniors, one voice does not fit all.

Age, region, culture, and health all leave distinct marks on how we speak. A softly spoken grandmother in rural France, a retired farmer in inland Portugal, and an elderly teacher from northern Spain — each carries a unique accent, rhythm, and cadence.

To truly serve them, SENSOLIFE must understand them.

This month, we took a major leap forward by training our AI models on elderly-accented datasets from across Europe, starting with Portugal, Spain, and France.

The Problem: Accents vs. Algorithms

Modern speech-to-text systems like Whisper perform well with standard speech patterns. But the voices of elderly users — often slower, quieter, and marked by subtle articulation shifts — are underrepresented in AI training corpora.

Add in regional accents, and suddenly, a polite “Bom dia” becomes a puzzle to transcribe.

Without adaptation, this leads to:

• Missed commands
• Incorrect responses
• Frustrating repetition for users who already may feel insecure about memory or clarity

That’s not acceptable. So we built a plan to teach our AI how seniors actually sound — not how textbooks say they should.

How We Collected the Data

We began by assembling a growing dataset of voice samples, sourced from:

• Local interviews (with consent) in care homes
• Archived oral history recordings
• Publicly available audio from older speakers in European linguistic corpora
• Simulated dementia-affected speech (to train sensitivity without needing diagnoses)

We then worked with native linguists and phonetic experts to tag common vocal traits — including hesitations, filler sounds, elongated vowels, and softened consonants.

Tuning the Model: What Changed

Using these new data streams, we fine-tuned our AI models to:

• Boost transcription accuracy for speech marked by aging-related vocal changes
• Recognize common regional phonetic patterns, like northern Castilian “s” softening or Alentejo drawl
• Distinguish between healthy pauses and disorientation patterns, improving early-warning detection for cognitive decline
• Handle multilingual households where a user might shift between French and Portuguese spontaneously

All this, without increasing response time or power consumption on the device.

Early Results

Even before full deployment, our updated voice model is showing:

• +31% improvement in transcription accuracy for elderly users with strong regional accents
• Fewer failed interactions due to misunderstood phrases
• Increased user confidence — especially among first-time users

It’s not just about clarity. It’s about respecting the way people naturally speak — not asking them to adjust to the machine.

More Languages, More Inclusion

This is just the beginning. SENSOLIFE’s architecture now supports seamless switching across over 20 languages. But we’re going deeper — not just adding languages, but honoring accents.

Our next phase includes regional voice adaptation for:

• Belgian French (Walloon)
• Galician Spanish
• Swiss German
• Southern Italian dialects

We believe inclusion isn’t checking a box — it’s tuning your ears to every voice that matters.