How to Deploy NeMo Speech for Speech Recognition with Docker

In this tutorial you will deploy NeMo Speech for automatic speech recognition using the official NVIDIA container. The whole setup takes about ten minutes and gives you a GPU-ready ASR service you can point at your own audio files.

Step 1: System requirements

NeMo is built for GPU acceleration. For a smooth experience, plan for at least 4 GB RAM and a CUDA-capable GPU; production workloads benefit from 8 GB RAM and a modern GPU like the A10 or L40.

Step 2: Run with Docker

Create a docker-compose.yml:

services:
  nemo-speech:
    image: nvcr.io/nvidia/nemo:latest
    restart: unless-stopped
    ports:
      - "8080:8080"
    volumes:
      - ./data/nemo-speech:/data
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: 1
              capabilities: [gpu]

Step 3: Transcribe your first audio

Start the stack with docker compose up -d, then use the Python API inside the container to run a Parakeet ASR model on any WAV file:

import nemo.collections.asr as asr
model = asr.models.ASRModel.from_pretrained("nvidia/parakeet-tdt-0.6b-v2")
text = model.transcribe(["audio.wav"])
print(text)

That is it — you now have a self-hosted, state-of-the-art speech recognition pipeline. Because everything runs in Docker, rolling back, upgrading, and moving between machines is trivial.