README.md

metadata

title: BabelCast Mistral
emoji: 🎙️
colorFrom: blue
colorTo: purple
sdk: docker
pinned: false
license: mit

BabelCast — Mistral 7B Translation Pipeline

Real-time speech translation pipeline: Whisper large-v3-turbo (STT) + Mistral 7B Instruct Q5_K_M (Translation) + Qwen3-TTS (Voice Dubbing).

Stack

Component	Model	Purpose
STT	faster-whisper large-v3-turbo	Speech-to-text
LLM	Mistral 7B Instruct v0.3 (GGUF Q5_K_M)	Translation via llama.cpp
TTS	Qwen3-TTS 0.6B with CUDA graphs	Voice dubbing

API Endpoints

Method	Path	Description
GET	/health	Health check
POST	/v1/transcribe	Audio -> text (STT)
POST	/v1/translate/text	Text translation
POST	/v1/translate	Audio -> translated text (STT + LLM)
POST	/v1/tts	Text -> speech
POST	/v1/speech	Full pipeline: audio -> STT -> translate -> TTS

Requirements

• NVIDIA GPU with 16+ GB VRAM (RTX 4090 recommended)
• CUDA 12.4+

Run locally

docker build -t babelcast-mistral .
docker run --gpus all -p 8000:8000 babelcast-mistral

Docker Hub

docker pull marcosremar/babelcast-mistral:latest
docker run --gpus all -p 8000:8000 marcosremar/babelcast-mistral:latest

Deploy on GPU cloud

Works on RunPod, Vast.ai, TensorDock, or any Docker-capable GPU host. Models are downloaded on first boot (~5 min with fast internet).