Deploy and Run ASR Models as NIM Microservices

This section provides detailed instructions for deploying and running ASR models using the NVIDIA Riva Python client and sample scripts from the Repository of the client.

Choose a model based on your language, inference mode, and capability requirements. You determine what CONTAINER_ID and NIM_TAGS_SELECTOR environment variables to use from the following table. For a complete list of models, languages, inference modes, and GPU memory requirements, refer to the ASR support matrix.

Model	Languages	Modes	Capabilities	CONTAINER_ID	NIM_TAGS_SELECTOR
Parakeet CTC English	English	Streaming + Offline	Transcription	parakeet-1-1b-ctc-en-us	Refer to the support matrix
Parakeet CTC Vietnamese	Vietnamese, English	Streaming + Offline	Transcription	parakeet-ctc-0.6b-vi	Refer to the support matrix
Parakeet CTC Spanish	Spanish, English	Streaming + Offline	Transcription	parakeet-ctc-0.6b-es	Refer to the support matrix
Parakeet CTC Mandarin	Mandarin, English	Streaming + Offline	Transcription	parakeet-ctc-0.6b-zh-cn	Refer to the support matrix
Parakeet CTC Taiwanese	Taiwanese, Mandarin, English	Streaming + Offline	Transcription	parakeet-ctc-0.6b-zh-tw	Refer to the support matrix
Parakeet TDT English	English	Offline only	Transcription, word-level timestamps	parakeet-0.6b-tdt	Refer to the support matrix
Parakeet RNNT Multilingual	25+ languages (auto-detect)	Streaming + Offline	Transcription	parakeet-1-1b-rnnt-multilingual	Refer to the support matrix
Conformer CTC Spanish	Spanish	Streaming + Offline	Transcription	riva-asr	Refer to the support matrix
Whisper Large v3	100+ languages (auto-detect)	Offline only	Transcription, translation to English	whisper-large-v3	Refer to the support matrix
Canary 1b Multilingual	26 languages	Offline only	Transcription, bidirectional translation	canary-1b	Refer to the support matrix