A review-first communication tool for users and caregivers
DysVoxa is an experimental Windows app for people whose speech can be difficult for conventional recognition systems to understand. It turns one short phrase into reviewable text options and speaks only after the user chooses or edits the text and presses Speak.
What DysVoxa does
DysVoxa is being developed to support phrase-by-phrase communication without assuming the first transcript is correct. It can collect several possible interpretations for the user to review.
You control each phrase
Start and stop recording yourself. The normal target is a short phrase of around 2–15 words, not continuous background listening.
You control the wording
Keep the original, choose another option, edit or type different text, use a quick phrase, or cancel and try again.
You control the output
Nothing speaks automatically. An explicit Speak action starts a separate synthesized Piper voice.
A phrase-by-phrase workflow
Each phrase follows a review-first path; recognition or a suggestion can still be wrong.
1. Record
Start recording, speak one short phrase, and stop when the phrase is complete.
2. Recognize locally
Standard uses Parakeet and Advanced uses Whisper on the Windows computer. Neither is guaranteed to be better.
3. Review possibilities
Inspect the original recognition and any available alternatives. Choose, edit, type, or cancel.
4. Press Speak
Only then does Piper read the selected text through speakers or the selected output path.
What to know before testing
- DysVoxa is an experimental Windows 10 and 11 desktop beta, not a phone app.
- Recognition and correction suggestions can be incomplete, unrelated, fluent but wrong, or unavailable.
- The project cannot promise reliable performance for any cause or severity of dysarthria.
- Do not rely on DysVoxa for emergency, medical, or other safety-critical communication.
- Keep the user's established communication fallback available.
Voice profiles and caregiver support
A caregiver, family member, clinician, or tester can help set up information that matters to one speaker without treating the profile as proof of general accuracy.
- Add important names, places, vocabulary, and frequently used phrases
- Mark critical phrases and recurring recognition mistakes
- Save a correction only when the user deliberately chooses to remember it
- Enroll phrase recordings for local matching
- Inspect and delete saved profile information when needed
A profile does not retrain Parakeet or Whisper. It provides local evidence to later stages and may be most useful for familiar phrases.
Local and cloud privacy choices
The intended Local workflow is self-contained, while optional Cloud mode is a deliberate BYOK choice.
- Parakeet and Whisper speech recognition process microphone audio on the Windows PC
- Piper speech generation, profiles, settings, and saved repairs stay local in Local or Off mode
- Normal Local use does not require internet after the offline installation is complete
- Cloud mode may send recognized text plus relevant saved vocabulary or repair examples
- The cloud correction path does not send raw microphone audio, and provider billing and retention terms apply
Reference Windows hardware
The current runtime direction is CPU-only. Performance and release gates are evaluated against a modest reference computer rather than a dedicated GPU.
Windows 10 or 11, 64-bit
The reference processor is in the Intel i5-8250U class with four cores. A dedicated graphics card is not intended to be required.
16 GB RAM and audio devices
Testing assumes 16 GB RAM, a microphone, and speakers or headphones. VB-CABLE is optional for experimental call routing.
This is a reference configuration, not a promise of a particular response time or recognition quality on every computer.
Who DysVoxa may help
The project is being designed for several groups, but its limited development data cannot establish effectiveness across causes, severities, accents, ages, or speech patterns.
- People with dysarthria
- People with other speech differences that conventional recognition handles poorly
- People who can speak a short phrase and want to review text before synthesized output
- People who also want direct typing and reusable quick phrases
- Caregivers, clinicians, family members, and testers supporting setup
- People who find saved phrases useful even when automatic recognition performs poorly
See the current beta
Watch short recordings of the current Windows workflow. A demo shows one run, not validated performance. VB-CABLE call routing remains experimental until remote-listener tests pass.
Join testing when available
Register your interest in testing the Windows beta. Feedback can help identify where recognition, suggestions, timing, and setup succeed or fail.
Fund dysarthria speech recognition that stays free to use.
One-time donations, Indiegogo backing, and organizational sponsorship pay for safety work, Windows packaging, and broader evaluation. SIA DysVoxa is a registered company in Latvia, so gifts may not be tax-deductible.