Le premier assistant vocal IA en malagasy.
Milo Voice
Le premier assistant vocal IA en malagasy.
Malagasy Intelligent Language Operator
STT → LLM → TTS
Parle en malagasy → Comprend → Repond en malagasy
---
## Le probleme
30 millions de personnes parlent malagasy. **Aucun assistant vocal ne les comprend.** Ni Siri, ni Alexa, ni Google Assistant ne supportent cette langue. Milo comble ce vide.
## La solution
Une pipeline vocale complete, **local-first**, qui tourne sur un seul GPU :
```
Micro Haut-parleur
| ^
v |
[ Whisper STT ] --> [ Claude / Mistral ] --> [ MMS-TTS ]
fine-tune MG repond en MG synthese MG
~300ms ~2s ~200ms
```
## Performances
| Metrique | Valeur |
|----------|--------|
| Latence chat texte | **~2s** |
| Latence pipeline vocale | **~2.5s** |
| VRAM utilisee | **1.6 Go** / 16 Go |
| Mode | Cloud (Claude) + fallback local (Mistral 7B) |
---
## Stack technique
### Backend (Python / FastAPI)
| Composant | Modele | Role |
|-----------|--------|------|
| **STT** | Whisper Medium fine-tune MG | Transcription audio malagasy |
| **LLM** | Claude 3.5 Haiku (cloud) / Mistral 7B Q4 (local) | Comprehension + generation |
| **TTS** | Facebook MMS-TTS `mlg` | Synthese vocale malagasy |
| **VAD** | Silero VAD | Detection d'activite vocale |
| **Traduction** | NLLB-200 600M | Traduction MG FR (endpoint dedie) |
### Frontend (React / TypeScript)
| Ecran | Description |
|-------|-------------|
| Conversation | Chat vocal temps reel via WebSocket |
| Transcription | Upload audio + transcription STT |
| Synthese | Saisie texte + generation audio TTS |
| Admin | Monitoring GPU, latence, metriques |
### Infrastructure
| Service | Port | Role |
|---------|------|------|
| FastAPI | `8000` | A …