README.md

metadata

title: PARLE Light — Speech-to-Speech
emoji: 🗣️
colorFrom: green
colorTo: blue
sdk: docker
pinned: false

PARLE Light — Speech-to-Speech Pipeline

Lightweight version of the PARLE speech-to-speech pipeline for fast testing and low-VRAM deployments.

Stack

Component	Model	VRAM
STT	Faster Whisper Small	~460 MB
LLM	Qwen2.5-0.5B-Instruct	~1 GB
TTS	gTTS (Google TTS API)	0

Total: ~1.5 GB VRAM — runs on any GPU.

Deploy with SkyPilot

sky launch -c parle-light skypilot.yaml --cloud vast --gpus RTX3090:1 -i 5 --down -y

Endpoints

• POST /api/stream-audio — SSE stream from audio
• POST /api/text — SSE stream from text
• GET /health — Health check
• GET /capabilities — Model info
• WS /ws/stream — WebSocket stream