thainph/node-trans

Real-time audio translation app

125

stars

commits

JavaScript

primary language

Jul 9, 2026

updated

README

Node Trans

Real-time audio transcription and translation app. Captures audio from microphone, system audio, or both. Supports two transcription modes: Soniox (cloud, built-in translation) and Local Whisper (fully offline).

Runs as a web app (browser) or desktop app (Electron). Supports macOS, Windows, and Linux.

Features

Real-time speech-to-text with live partial results
Automatic translation (cloud via Soniox, or local via Ollama / LibreTranslate)
Speaker diarization — identifies who is speaking (Speaker 1, Speaker 2, ...)
Session history with speaker renaming and Markdown export
Always-on-top overlay window (Electron)
Bilingual UI (English / Vietnamese)

Quick Start

npm install
npm run start      # web app — open http://localhost:3000

For Electron:

npm run electron:dev

See DEV-BUILD.md for build instructions and troubleshooting.

Hardware Requirements

Only relevant for Local Whisper mode. Soniox only needs an internet connection.

macOS

Config	RAM	Usable features
Minimum	8 GB	Whisper tiny/base + Ollama gemma3:1b, no diarization
Recommended	Apple Silicon 16 GB	Whisper medium/large-v3-turbo + Ollama gemma3:4b + diarization
Optimal	Apple Silicon 32 GB	All features with large models, very smooth

Apple Silicon recommended (M1+): unified memory allows CPU and GPU to share RAM — Whisper (Metal), Ollama (Metal), and diarization (MPS) are all hardware-accelerated, 4–6× faster than Intel at the same model size.

Windows

Config	RAM	GPU	Usable features
Minimum	16 GB	None	Whisper tiny/base + small Ollama (slow), no diarization
Recommended	16 GB	NVIDIA 8 GB VRAM (RTX 3060+)	Whisper medium + Ollama gemma3:4b + diarization
Optimal	32 GB	NVIDIA 12 GB+ VRAM (RTX 4070 Ti+)	All features comfortably

AMD GPUs are not supported. Without an NVIDIA GPU, diarization cannot run in real-time. Install CUDA Toolkit for GPU acceleration.

Linux

Config	RAM	GPU	Usable features
Minimum	16 GB	None	Whisper tiny/base + small Ollama (slow), no diarization
Recommended	16 GB	NVIDIA 8 GB VRAM (RTX 3060+)	Whisper medium + Ollama gemma3:4b + diarization
Optimal	32 GB	NVIDIA 12 GB+ VRAM (RTX 4070 Ti+)	All features comfortably

Requires PulseAudio or PipeWire (with PulseAudio compatibility) for audio device management. Most modern distros (Ubuntu 22.04+, Fedora, etc.) include one of these by default. Install ffmpeg via your package manager (sudo apt install ffmpeg).

Disk space (Local Whisper)

Component	Size
Whisper base model	~150 MB
Whisper large-v3-turbo model	~1.6 GB
Ollama + gemma3:4b model	~3.5 GB
Python venv for diarization (torch)	~4–6 GB
pyannote diarization model	~1 GB
Total (full setup)	~10–12 GB

Soniox (Cloud)

Requires a free API key from soniox.com.

Configure in Settings → Engine → Soniox API Key, or via .env:

SONIOX_API_KEY=your_api_key_here

Soniox handles transcription, translation, and speaker diarization out of the box — no additional setup needed.

Local Whisper (Offline)

No API key required. All processing happens on-device.

1. Build whisper.cpp (one-time):

npm run setup:whisper

2. Download a model:

cd node_modules/nodejs-whisper/cpp/whisper.cpp
bash models/download-ggml-model.sh base   # or: tiny, small, medium, large-v3-turbo

3. Select in app: Settings → Engine → Local Whisper → choose your model.

4. Translation (optional): configure Ollama or LibreTranslate in Settings → Engine.

Whisper Models

Model	Size	RAM	Speed	Quality
`tiny`	~75 MB	~1 GB	Very fast	Low
`base`	~150 MB	~1 GB	Fast	Medium
`small`	~500 MB	~2 GB	Medium	Good
`medium`	~1.5 GB	~5 GB	Slow	High
`large-v3-turbo`	~1.6 GB	~4 GB	Medium	Very high
`large`	~3 GB	~8 GB	Slowest	Best

Recommended: large-v3-turbo on Apple Silicon, base on low-end hardware.

Speaker Diarization (Local Whisper)

Identifies individual speakers using pyannote-audio. Requires Python 3.10–3.12 and a free HuggingFace token.

1. Install Python dependencies:

npm run setup:diarize

Or via the UI: Settings → Engine → Setup Diarization.

2. Create a HuggingFace READ token at huggingface.co/settings/tokens, then accept terms for:

3. Enter the token in Settings → Engine → HuggingFace Token.

On first use, the app downloads ~1 GB of models. Subsequent starts are ready in ~10–15 seconds.

System Audio Capture

System audio capture (from Zoom, Meet, YouTube, etc.) works natively — no virtual audio drivers needed. Pre-built binaries are included in the repo.

macOS: ScreenCaptureKit (requires Screen Recording permission, macOS 13+)
Windows: WASAPI loopback (no special permissions needed)
Linux: PulseAudio / PipeWire monitor sources (via ffmpeg)

To rebuild from source: npm run setup:audiocap (requires Swift on macOS, .NET 8 SDK for Windows cross-compile)

Documentation

File	Contents
DEV-BUILD.md	Development setup, build commands, troubleshooting
ARCHITECTURE.md	System architecture, data flow, component overview

Tech Stack

Layer	Technology
Frontend	React 19, Vite, Tailwind CSS v4, Socket.IO Client
Backend	Node.js, Express 5, Socket.IO
Desktop	Electron
Audio capture (mic)	ffmpeg (avfoundation / dshow / pulse)
Audio capture (sys)	audiocap (macOS / Windows), ffmpeg+pulse (Linux)
STT (cloud)	Soniox API
STT (local)	nodejs-whisper (whisper.cpp)
Speaker diarization	pyannote-audio 3.1 (Python)
Translation (local)	Ollama / LibreTranslate
Database	SQLite (better-sqlite3)

Contributors

thainph

51 commits

collyn

3 commits

Similar repos

homelab-00/TranscriptionSuite

A fully local and private Speech-To-Text app, offering multiple model backends, diarization &…

720

lxp135/reco-tran

一个录音并转写文字的小工具

kizuna-ai-lab/sokuji

Real-time two-way speech translation for bilingual meetings — auto-detects the spoken language and…

1,280

nanameru/stt-test

Sharrnah/whispering

Whispering Tiger - OpenAI's whisper (and other models) with OSC and Websocket support. Allowing…

544

OpenWhispr/openwhispr

Voice-to-text dictation app with local (Nvidia Parakeet/Whisper) and cloud models (BYOK).…

8,009

chidiwilliams/buzz

Buzz transcribes and translates audio offline on your personal computer. Powered by OpenAI's…

21,357

collabora/WhisperLive

A nearly-live implementation of OpenAI's Whisper.

4,255

Languages

JavaScript

82.5%

Python

6.8%

CSS

4.4%

Swift

3.5%

1.5%

HTML

1.3%

thainph/node-trans

Real-time audio translation app

125

stars

commits

JavaScript

primary language

Jul 9, 2026

updated

README

Node Trans

Runs as a web app (browser) or desktop app (Electron). Supports macOS, Windows, and Linux.

Features

Real-time speech-to-text with live partial results
Automatic translation (cloud via Soniox, or local via Ollama / LibreTranslate)
Speaker diarization — identifies who is speaking (Speaker 1, Speaker 2, ...)
Session history with speaker renaming and Markdown export
Always-on-top overlay window (Electron)
Bilingual UI (English / Vietnamese)

Quick Start

npm install
npm run start      # web app — open http://localhost:3000

For Electron:

npm run electron:dev

See DEV-BUILD.md for build instructions and troubleshooting.

Hardware Requirements

Only relevant for Local Whisper mode. Soniox only needs an internet connection.

macOS

Config	RAM	Usable features
Minimum	8 GB	Whisper tiny/base + Ollama gemma3:1b, no diarization
Recommended	Apple Silicon 16 GB	Whisper medium/large-v3-turbo + Ollama gemma3:4b + diarization
Optimal	Apple Silicon 32 GB	All features with large models, very smooth

Windows

Config	RAM	GPU	Usable features
Minimum	16 GB	None	Whisper tiny/base + small Ollama (slow), no diarization
Recommended	16 GB	NVIDIA 8 GB VRAM (RTX 3060+)	Whisper medium + Ollama gemma3:4b + diarization
Optimal	32 GB	NVIDIA 12 GB+ VRAM (RTX 4070 Ti+)	All features comfortably

AMD GPUs are not supported. Without an NVIDIA GPU, diarization cannot run in real-time. Install CUDA Toolkit for GPU acceleration.

Linux

Config	RAM	GPU	Usable features
Minimum	16 GB	None	Whisper tiny/base + small Ollama (slow), no diarization
Recommended	16 GB	NVIDIA 8 GB VRAM (RTX 3060+)	Whisper medium + Ollama gemma3:4b + diarization
Optimal	32 GB	NVIDIA 12 GB+ VRAM (RTX 4070 Ti+)	All features comfortably

Disk space (Local Whisper)

Component	Size
Whisper base model	~150 MB
Whisper large-v3-turbo model	~1.6 GB
Ollama + gemma3:4b model	~3.5 GB
Python venv for diarization (torch)	~4–6 GB
pyannote diarization model	~1 GB
Total (full setup)	~10–12 GB

Soniox (Cloud)

Requires a free API key from soniox.com.

Configure in Settings → Engine → Soniox API Key, or via .env:

SONIOX_API_KEY=your_api_key_here

Soniox handles transcription, translation, and speaker diarization out of the box — no additional setup needed.

Local Whisper (Offline)

No API key required. All processing happens on-device.

1. Build whisper.cpp (one-time):

npm run setup:whisper

2. Download a model:

cd node_modules/nodejs-whisper/cpp/whisper.cpp
bash models/download-ggml-model.sh base   # or: tiny, small, medium, large-v3-turbo

3. Select in app: Settings → Engine → Local Whisper → choose your model.

4. Translation (optional): configure Ollama or LibreTranslate in Settings → Engine.

Whisper Models

Model	Size	RAM	Speed	Quality
`tiny`	~75 MB	~1 GB	Very fast	Low
`base`	~150 MB	~1 GB	Fast	Medium
`small`	~500 MB	~2 GB	Medium	Good
`medium`	~1.5 GB	~5 GB	Slow	High
`large-v3-turbo`	~1.6 GB	~4 GB	Medium	Very high
`large`	~3 GB	~8 GB	Slowest	Best

Recommended: large-v3-turbo on Apple Silicon, base on low-end hardware.

Speaker Diarization (Local Whisper)

Identifies individual speakers using pyannote-audio. Requires Python 3.10–3.12 and a free HuggingFace token.

1. Install Python dependencies:

npm run setup:diarize

Or via the UI: Settings → Engine → Setup Diarization.

2. Create a HuggingFace READ token at huggingface.co/settings/tokens, then accept terms for:

3. Enter the token in Settings → Engine → HuggingFace Token.

On first use, the app downloads ~1 GB of models. Subsequent starts are ready in ~10–15 seconds.

System Audio Capture

System audio capture (from Zoom, Meet, YouTube, etc.) works natively — no virtual audio drivers needed. Pre-built binaries are included in the repo.

macOS: ScreenCaptureKit (requires Screen Recording permission, macOS 13+)
Windows: WASAPI loopback (no special permissions needed)
Linux: PulseAudio / PipeWire monitor sources (via ffmpeg)

To rebuild from source: npm run setup:audiocap (requires Swift on macOS, .NET 8 SDK for Windows cross-compile)

Documentation

File	Contents
DEV-BUILD.md	Development setup, build commands, troubleshooting
ARCHITECTURE.md	System architecture, data flow, component overview

Tech Stack

Layer	Technology
Frontend	React 19, Vite, Tailwind CSS v4, Socket.IO Client
Backend	Node.js, Express 5, Socket.IO
Desktop	Electron
Audio capture (mic)	ffmpeg (avfoundation / dshow / pulse)
Audio capture (sys)	audiocap (macOS / Windows), ffmpeg+pulse (Linux)
STT (cloud)	Soniox API
STT (local)	nodejs-whisper (whisper.cpp)
Speaker diarization	pyannote-audio 3.1 (Python)
Translation (local)	Ollama / LibreTranslate
Database	SQLite (better-sqlite3)

Similar repos

homelab-00/TranscriptionSuite

A fully local and private Speech-To-Text app, offering multiple model backends, diarization &…

720

lxp135/reco-tran

一个录音并转写文字的小工具

kizuna-ai-lab/sokuji

Real-time two-way speech translation for bilingual meetings — auto-detects the spoken language and…

1,280

nanameru/stt-test

Sharrnah/whispering

Whispering Tiger - OpenAI's whisper (and other models) with OSC and Websocket support. Allowing…

544

OpenWhispr/openwhispr

Voice-to-text dictation app with local (Nvidia Parakeet/Whisper) and cloud models (BYOK).…

8,009

chidiwilliams/buzz

Buzz transcribes and translates audio offline on your personal computer. Powered by OpenAI's…

21,357

collabora/WhisperLive

A nearly-live implementation of OpenAI's Whisper.

4,255

Contributors

thainph

51 commits

collyn

3 commits

Languages

JavaScript

82.5%

Python

6.8%

CSS

4.4%

Swift

3.5%

1.5%

HTML

1.3%