Run Your First NVIDIA Speech NIM Microservice for Automatic Speech Recognition

Why it matters

Voice input reaches an agent as text only after a speech model has transcribed it, and the choice between streaming and offline transcription shapes latency for everything downstream.

What to practise

Start the ASR NIM with NIM_TAGS_SELECTOR="name=parakeet-1-1b-ctc-en-us,mode=all", exposing HTTP on port 9000 and gRPC on 50051, and wait for /v1/health/ready. List the loaded models, then transcribe the sample WAV in streaming mode with transcribe_file.py over gRPC and realtime_asr_client.py over WebSocket, and in offline mode with transcribe_file_offline.py over gRPC and /v1/audio/transcriptions over HTTP. Watch partial results arrive while streaming, and note that mode=all loads both modes at a higher GPU memory cost.

How it fits

This is the speech branch of 7.5.1: a transcript from a dedicated ASR NIM, separate from the visual track. Weigh its GPU allocation against routing audio through an Omni VLM instead.