Obiente Translate

Obiente Translate is a lightweight app for live, multi-source transcription and instant translation. It turns spoken conversations into readable text in real time and presents them in a clean, distraction-free "Focus" view — ideal for events, kiosks, and live captioning.

Why people use it

• Make spoken content readable and shareable instantly.
• Help multilingual audiences follow live conversations in real time.
• Perfect for online calls, language exchange, and multilingual friendships.
• Provide a clear, centered display for public installations and events.

Core features

• Live transcription from microphones and system audio.
• Optional real-time translation into target languages.
• Focus Mode: line-by-line, centered Lyrics style display.
• Participant chips for quick toggling and speaker clarity.
• Voice Activity Detection (VAD): Automatically filters out silence and background noise, reducing network traffic and preventing hallucinations from silent audio chunks.
• Noise Gate: Suppresses low-level audio to improve transcription quality.

Quick start

pnpm install
pnpm dev    # start the dev server (usually http://localhost:3000)

Local Go speech server

# Default dev path. Docker builds whisper.cpp, includes ffmpeg, and serves on
# http://127.0.0.1:8080. Put models in gowhisper/models.
pnpm models:base
pnpm dev:go

Useful variants

# Use another mounted model from gowhisper/models.
pnpm models:turbo
MODEL_NAME=large-v3-turbo pnpm dev:go

# Bare-metal fallback if you want to debug CGO locally.
pnpm dev:go:local

Use multilingual model names such as tiny, base, or large-v3-turbo. Avoid .en models for this app; Whisper .en models only transcribe English.

The Vite dev server now proxies /ws/transcribe, /transcribe, /translate, /detect, /languages, and /healthz to http://127.0.0.1:8080, so local development uses the Go path by default. Set VITE_WHISPER_ENDPOINT or VITE_WHISPER_STREAMING_ENDPOINT to point the frontend elsewhere.

Translation notes

• The Go server exposes LibreTranslate-compatible /translate, /detect, and /languages endpoints and fans out multi-target translations concurrently.
• Dev and production use https://libretranslate.obiente.cloud by default. Keep translation traffic routed through the Go server so caching and multi-target fan-out stay consistent.
• Translation results are cached in-process and source=target requests return immediately.
• The browser often records WebM/Ogg/MP4 for chunked mode. The Docker dev server includes ffmpeg; streaming mode sends PCM and does not depend on it.

Build for production

pnpm build
pnpm preview

Audio Filtering

To optimize network usage and transcription quality, the app includes intelligent audio filtering:

Voice Activity Detection (VAD)

• Automatically detects whether audio contains speech or just silence/noise
• Prevents sending empty chunks that waste bandwidth and cause hallucinations
• Uses RMS energy analysis and Zero Crossing Rate (ZCR) detection
• Only sends audio chunks when actual speech is detected

Noise Gate

• Suppresses low-level background noise and hiss
• Applies envelope-following gain reduction with configurable attack/release
• Cleans up audio before sending to the transcription server
• Default threshold: -40dB (ratio: 10:1)

Configuration

The filters automatically scale to the audio sample rate. Default settings:

• Energy threshold: 0.01 (RMS)
• Max ZCR: 0.3 (higher = more noise-like)
• Min speech duration: ~0.1 seconds
• Noise gate threshold: 0.01 with 10:1 ratio

TL;DR Obiente Translate makes international speech readable and accessible in real time with a minimal, usable UI.