Skip to content

semaje

Open-source speech-to-text and text-to-speech for 1,600+ languages, on your own infrastructure.

semaje turns any recording (meetings, interviews, phone voice notes, lectures, broadcast video) into searchable, shareable transcripts, and brings dictation, live captions, and read-aloud to every web page through its browser extension. It runs only open-source models, so it works without any hosted AI service, including fully air-gapped.

Highlights

  • Every language that matters to you. Whisper for the 30 or so languages it handles best; Meta's Omnilingual ASR for 1,672 languages, including Swahili, Kikuyu, Luo, Yoruba, Amharic, Hausa, Luganda, Somali, Zulu, and hundreds more. Routing is automatic and measured: Languages.
  • Any audio. Every common audio and video format, from 8 kHz phone audio to 96 kHz studio masters and 5.1 surround (dialogue-weighted downmix), with optional noise reduction.
  • Speech out, too. Piper voices by default; MMS-TTS for about 1,100 more languages when an operator accepts its non-commercial license; custom voices for anything else.
  • Everywhere you type. The browser extension for Chrome, Edge, Brave, Opera, Vivaldi, Arc, Firefox, and Safari: semaje-extension.
  • Built for organizations. SSO, scoped API keys, tenancy, audit trails, retention, Prometheus metrics, request tracing, readiness probes, and back-pressure under load.

Quick start

git clone https://github.com/semajehq/semaje
cd semaje
cp .env.example .env
deploy/semaje cpu up
open http://localhost:8080
Profile Hardware Speech models
cpu 4 cores, 8 GB RAM Whisper base + Omnilingual CTC-300M
workstation-gpu NVIDIA 8+ GB Whisper large-v3-turbo + Omnilingual LLM-1B
server-gpu NVIDIA 16+ GB Whisper large-v3 + Omnilingual LLM-3B

Guides