Local AI Guide

See the speed before choosing hardware

Local speech recognition & synthesis

Transcribe Korean speech live and save text with Nemotron 3.5 ASR

Follow Korean speech on screen and save the finalized transcript as text.

Nemotron 3.5 ASR1 is a 600M-parameter streaming speech recognition model. With NVIDIA NeMo-Speech.cpp, you can connect the system microphone and watch Korean transcription update as you speak. Korean ko-KR is available for transcription without additional adaptation.

Contents
Requirements and key details
  • The 600M model processes audio in chunks and extends the transcript with each new segment.
  • Korean ko-KR is one of the 32 locales ready to use without adaptation.
  • Review the live preview and the final transcript saved to a file separately.

Live text changes before a sentence is finalized

Nemotron 3.5 ASR processes audio in small chunks and extends the text with what it has heard so far. Unlike transcribing once after a recording ends, the displayed words can change while the microphone is active. Each new chunk adds context, so treat short phrases on screen as provisional and review the finalized utterance after a pause. NVIDIA model card

The model card lists 40 language locales with different readiness levels. Nineteen are transcription-ready and 13 are broad-coverage, so both groups can transcribe without adaptation. The remaining eight need adaptation with domain data. Korean ko-KR is transcription-ready. NVIDIA model card

An audio waveform displayed on a laptop beside a microphone
The transcript updates while audio arrives from the microphone.

Install the runtime and start Korean microphone input

NeMo-Speech.cpp provides an installer for macOS and Linux and a PowerShell installer for Windows. Follow the official instructions for your OS, open a new terminal, then run nemo-speech doctor and nemo-speech model list. The first command reports installed audio and inference2 backends; the second lists model aliases available in that build.

Install NeMo-Speech.cpp on macOS or Linux
curl -fsSL https://github.com/NVIDIA/NeMo-Speech.cpp/raw/main/scripts/install.sh | sh
Install NeMo-Speech.cpp from Windows PowerShell
irm https://github.com/NVIDIA/NeMo-Speech.cpp/raw/main/scripts/install.ps1 | iex
Check the CLI in a new terminal
nemo-speech doctor
nemo-speech model list

Allow microphone access for the terminal, then say a short phrase somewhere quiet. The command explicitly selects the Nemotron 3.5 alias and saves the final Korean transcript to live-ko.txt.

Start Korean microphone transcription
nemo-speech pull nemotron-3.5
nemo-speech transcribe --live --model nemotron-3.5 --output live-ko.txt

Compare the live preview with the saved final transcript

After you pause, the default endpointing setting detects silence and finalizes the utterance while microphone capture continues. To finish the full transcript, press Ctrl-C once so the stream can flush. The final text appears in the terminal, and the live-ko.txt option also saves it in the working directory. Open that file to compare words that changed during the live preview. NeMo-Speech.cpp CLI guide

Microphone input does not create an audio file. When transcription ends, correct names and numbers in the saved text, then replay the matching passage if you have an original recording.

A desk with headphones, a microphone, and a transcription workspace
Compare interim text with finalized sentences from the same recording.

How long until words appear?

For live transcription, the wait between speaking and seeing words matters. The model collects 80–1120 ms audio chunks, then adds inference time before updating the transcript. Smaller chunks update more often; larger chunks gather more audio first. Read a sentence with your intended device and microphone, and check both the first visible words and the finalized line.

A recorder, notes, and headphones for reviewing speech recognition
After stopping the microphone, review the complete transcript and saved captions.

Terminology notes

  1. Automatic Speech Recognition — Technology that recognizes spoken language in audio and converts it to text.

    Back to the text

  2. Inference — The process of using a trained model to compute an output for an input. Here, local inference means running the model on the user’s device.

    Back to the text

Frequently asked questions

What should you check first in this article?
Nemotron 3.5 ASR is a 600M-parameter streaming speech recognition model. With NVIDIA NeMo-Speech.cpp, you can connect the system microphone and watch Korean transcription update as you speak. Korean `ko-KR` is available for transcription without additional adaptation.
Where can you check the conditions and sources?
Nemotron 3.5 ASR processes audio in small chunks and extends the text with what it has heard so far. Unlike transcribing once after a recording ends, the displayed words can change while the microphone is active. Each new chunk adds context, so treat short phrases on screen as provisional and review the finalized utterance after a pause. [NVIDIA model card](https://huggingface.co/nvidia/nemotron-3.5-asr-streaming-0.6b)