NeMoscribe

A multi-lingual CLI transcriber (batch + live streaming) built on NVIDIA Nemotron 3.5 ASR. NeMoscribe has two subcommands transcribe and stream. The transcribe subcommand can read audio or video files and generate transcriptions in TXT, JSONL and SRT formats. The stream subcommand can capture multiple live sources like mic, system and also file in a single run then generates the same transcription formats. NeMoscribe is a wrapper around NVIDIA's Nemotron ASR model which supports 40 language-locales in three quality tiers (en-US, en-GB, ar-AR are "transcription-ready"). NeMoscribe was built and tested for English and Arabic (--language accepts any supported language by the model).

Requirements

• Python >= 3.12

Dependency	Needed for	Linux	Windows
ffmpeg + ffprobe	All audio loading, system capture	apt install ffmpeg / pacman -S ffmpeg	winget install ffmpeg
PortAudio	mic source	apt install libportaudio2 / pacman -S portaudio	bundled in the sounddevice wheel — nothing to do
PipeWire/Pulse	system source	usually preinstalled	not supported (see matrix)

Installation

Quick Start

Transcribe talk.mp4 and generate talk.txt talk.jsonl and talk.srt.

nemoscribe transcribe talk.mp4

Transcribe a live stream from your mic. This will show your speech as you talk Ctrl-C ends it.

nemoscribe stream mic

Transcribe a meeting live, save the audio and generate transcription files. This is the full set, and the naming law: every .wav pairs with the same-named .srt/.jsonl/.txt; stream-X.wav is the playable session — mpv stream-X.wav auto-loads its subtitles — and stream-X-me.wav etc. are clean per-speaker stems.

nemoscribe stream mic:me system:them --split --save-audio

Support Matrix

Performance

Caches

Fine-Tuning

Development