SpeechAssist

Auto-listen
Starts when I speak Auto-speak
Says what was heard straight away AI guesses
For slurred speech: choose the meaning

Heard (raw):

—

When I speak, the app writes my words here and says them aloud. Tap the text to say it again.

Teach the app my voice

Tap Record, then I say the phrase the way I normally say it. I record each phrase 3 times. The app then recognises my way of saying it, however slurred, whenever the normal recogniser is unsure.

—

Test, check accuracy, add phrases, export

Add one of my own phrases or a name

Settings

1 · Speech recognition (hearing me)

Records my voice and turns it into text. Groq is free and fast (Whisper). OpenAI is paid but very good. ElevenLabs Scribe is very accurate. "Browser" uses the built-in recogniser (Chrome on a computer or Android; no key needed, but it is not personalised and does not work in the iPad home-screen app).

Provider Use my trained phrases (Train tab)
When the recogniser is unsure, match my speech against my own recordings. Boost very quiet speech before recognising it
Recommended: weak breath support makes speech soft, and soft audio often comes back empty. Groq API key OpenAI API key ElevenLabs Scribe model OpenAI transcription model Language Silence that ends a sentence (seconds) Microphone sensitivity

2 · Understanding (AI clean-up of slurred words)

Only used when "AI guesses" is switched on (Talk screen). An AI reads the raw transcript plus my personal words and guesses what I meant, showing the likely meanings to choose from. Claude gives the best guesses; Groq is free. With "AI guesses" off, the app writes and says exactly what it hears.

AI guesses for slurred speech (same switch as on the Talk screen) Provider Anthropic API key Claude model OpenAI chat model Groq chat model Auto-speak the best guess after a countdown Countdown before auto-speak (seconds)

3 · Voice (speaking for me)

ElevenLabs gives a natural voice, and can clone my own voice from old recordings. Without a key the app uses the free built-in voice.

ElevenLabs API key

Voice

The list shows ElevenLabs' built-in voices; pick one and press ▶ Test voice to compare. "Load my voices" adds the voices in your ElevenLabs account (needs a key with Voices: Read), such as my cloned voice.

Or paste a voice ID A typed ID overrides the list above. ElevenLabs model (list refreshes from your account when you Load voices) Speaking speed Fallback browser voice

4 · My words (names, places, medicines)

One per line. These are fed to the recogniser and the AI so names like "Priya" or "suction" are recognised even when slurred. The app also learns from corrections.

About me (helps the AI guess well)

5 · Access & display

Text size Theme Switch scanning Auto-scan speed (seconds per item)

Tip: iPad (iPadOS 18+) has built-in eye tracking with dwell-click, and Windows has Eye Control. Both work with this app because every button is large and needs only a single tap.

6 · CLARIS clarifier (experimental)

CLARIS (IIIT Hyderabad / TCS Research, 2026) re-synthesises slurred or whispered speech as clear speech. When enabled, each utterance is also sent to the CLARIS service; the app transcribes the clarified version too and adds a "Hear clarified speech" button. Run it with the files in the claris-service folder.

Use CLARIS

Service address

CLARIS runs on the family laptop and connects to this site automatically while claris-service\run.ps1 is running.

Which CLARIS model Also transcribe the clarified audio and give the AI both transcripts Play clarified speech in my own voice (ElevenLabs Speech-to-Speech)
Needs an ElevenLabs voice selected above. Costs like normal speech generation.

7 · Data

⬆ Import backup

Everything is stored on this device only. Keys never leave the browser except to call the services above. Export a backup before changing devices.