Skip to main content
DysVoxa
Status

Development stages and current situation

What the DysVoxa beta does today, which development stages are complete or underway, and what has to be demonstrated before a release.

Current situation

Stated plainly:

  • A functioning Windows beta exists. The phrase workflow — record, review, choose, speak — runs end to end.
  • Recognition quality varies by speaker and can be wrong. DysVoxa is not validated for any particular speaker or severity level.
  • The central open question is speaker adaptation: whether fine-tuning the recognizer on a user's own enrollment recordings makes recognition reliable enough to be useful. That question drives the current development work.
  • DysVoxa is not a medical device and must not be relied on for safety-critical communication. An established fallback method should remain available.

Development stages

Four stages from the existing beta to a release. Each stage has to be demonstrated — not assumed — before the next one is treated as real.

  1. 1

    Phrase workflow

    Complete

    The core loop exists in a working Windows beta: record one short phrase, run local recognition, review the text and any alternatives, edit or choose, then press Speak for synthesized output. Nothing is spoken automatically.

  2. 2

    Speaker-adaptation evidence

    In progress

    Experiments measure whether adapting a speech-recognition model to one speaker's enrollment recordings improves accuracy for that speaker. Development uses the public TORGO dysarthric-speech corpus and the developer's own recordings — no user audio is collected or used.

  3. 3

    In-app enrollment and on-device training

    In progress

    Guided enrollment phrases inside the app, local fine-tuning of the recognizer on the user's own computer, and a held-out check: a personal model becomes active only if it beats the current model on enrollment data it did not train on. Retraining never runs silently.

  4. 4

    Validation and packaged release

    Planned

    An offline installer verified on ordinary hardware with the network disabled, per-speaker results reported honestly (including where recognition does not improve), and a free release.

Technical facts

Platform
Windows 10 and 11, 64-bit
Recognition
Local on-device models — Parakeet (standard mode) or Whisper (advanced mode)
Voice output
Piper text-to-speech; optional VB-CABLE routing
Network
Not required for the local workflow once installed
Telemetry
None
Reference hardware
Intel i5-8250U-class CPU, 16 GB RAM, no dedicated GPU
Release intent
Free; no paid tier is planned
Operated by
SIA DysVoxa, registered in Latvia

What is not claimed

  • No claim that recognition works reliably for any given speaker — evidence is still being gathered.
  • No claim that suggested interpretations recover the intended meaning; the user reviews and decides.
  • Demo videos document single runs of the current beta, not validated performance.
  • No timeline is promised for the release stages above.

Support the work

DysVoxa is developed by SIA DysVoxa, a registered company in Latvia. It is not a charity, so a gift may not be tax-deductible. A donation is a gift, not a purchase, and it does not create a paid tier. The intended release is free.

Donate with PayPalBack on IndiegogoPropose a sponsorship

Follow the work

The blog documents development notes and measurement results as they land. Testing interest can be registered by email.

Development notes on the blogRegister testing interestWatch the current demos