🔊 Voice / TTS Architecture
Four-tier text-to-speech fallback chain. Best-quality cloud voices when online; bundled neural models when not. Used by Prism AAC, in-call notifications, accessibility narration, and any app feature that speaks.---
🎚️ Four-Tier Fallback Chain
| Tier | Engine | Quality | Offline | Cost | When |
|---|---|---|---|---|---|
| 1 | Inworld TTS-2 (cloud) | Best-in-class — natural prosody, 60+ voices, voice cloning | — | Per-character billing; free for ro/uk/ru/de/ko/ar (Synalux absorbs cost) | Default for paid tiers; default for free tier in subsidized languages |
| 1.5 | Kokoro-82M neural (WASM) | Very good — locally-run neural voice | ✅ en/es/fr/pt/ja/zh | $0 | Free tier non-subsidized langs; offline; Tier 1 unavailable |
| 2 | OS Web Speech API premium voices | Good — varies by OS (Apple voices best) | ✅ | $0 | Tier 1.5 unavailable in user's language; bandwidth-saver mode |
| 3 | WASM espeak-ng | Acceptable — robotic but always works | ✅ | $0 | Last resort; covers 100+ languages |
The chain is automatic — no user configuration required. The voice picker (paid) lets users override their default with any Tier 1 voice.
---
🎤 Voice Cloning (Paid)
"Speak in MY voice — I trained it last week."
* 3-minute training sample uploaded once via the Voice Picker UI.
* Inworld voice-clone training takes ~10 minutes; the cloned voice is then available alongside the standard 60+ voices.
* Workspace-scoped — clones are bound to the workspace; staff voices auto-shared, patient voices private.
* Use cases: AAC users speak in a parent's voice; clinicians who prefer to hear notes in their own voice; demo videos for parents/caregivers.
---
🌍 Subsidized Languages (Free Tier)
Synalux absorbs Inworld TTS-2 cost for these languages on the free tier so users in regions where the platform serves underserved populations get premium voices at no cost:
ro (Romanian) · uk (Ukrainian) · ru (Russian) · de (German) · ko (Korean) · ar (Arabic)
---
🩺 Why This Matters Clinically
* Children with speech impairments depend on AAC's TTS for their actual voice — robotic Tier-3 fallbacks aren't acceptable for daily use; Tier 1 / 1.5 must succeed.
* Bilingual households need accurate prosody in both languages — Tier 1 is critical for non-English.
* Offline reliability — a child at school without Wi-Fi must still have their device speak. Tier 1.5 + 2 + 3 are bundled; the device always talks.
---
🏗️ Architecture
``
POST /api/v1/tts Generate TTS audio (auto tier-route based on lang + tier + connectivity)
POST /api/v1/tts/public Anonymous TTS for AAC widgets (rate-limited per IP)
GET /api/v1/tts/voices List available voices for the user's tier + language
POST /api/v1/voices/clone Submit a voice-clone training sample (paid)
`
Routing logic:
`
synthesize(text, lang, voice?) →
if voice && Tier1.available(voice): return Tier1.speak(text, voice)
if Tier1.available(lang) && (paid || Tier1.subsidizes(lang)): return Tier1.speak(text)
if Tier1_5.supports(lang): return Tier1_5.speak(text) // Kokoro
if Tier2.has(lang): return Tier2.speak(text) // Web Speech API
return Tier3.speak(text) // espeak-ng
``
---
💳 Plans
| | Free | Standard | Advanced | Enterprise |
|---|---|---|---|---|
| Default Inworld voice | ✅ | ✅ | ✅ | ✅ |
| All 60+ Inworld voices | — | ✅ | ✅ | ✅ |
| Voice cloning | — | — | ✅ | ✅ |
| In-call clinical dictation TTS feedback | — | ✅ | ✅ | ✅ |
| Custom voice library (workspace-curated) | — | — | — | ✅ |
See full pricing →---
🔄 Where TTS Is Used
* Prism AAC — picture tiles → speech, keyboard speak button, math panel AI tutor.
* In-call notifications — meeting started, patient joined, recording started.
* Accessibility narration — for users with low vision; reads notifications + form errors.
* AAC chat — assistant responses can be auto-spoken with the user's selected voice.