Featured Models

NeMo’s ASR collection supports several model architectures. This page covers the key model families and their capabilities. For pretrained checkpoints, see All Checkpoints. For config file details, see Configuration Files.

Parakeet

Parakeet is a family of ASR models with a FastConformer Encoder and CTC, RNN-T, or TDT decoders.

• Parakeet-TDT-0.6B V3 — 25 languages, PnC, blazing fast

• Parakeet-TDT-0.6B V2 — English-only, PnC, blazing fast

• Parakeet-TDT/CTC-110M — Edge deployment

• Nemotron-Speech-Streaming — Real-time streaming

• Multitalker-Parakeet — Multi-speaker streaming

Canary

Canary models are encoder-decoder models with a FastConformer Encoder and Transformer Decoder [ASR-MODELS2]. They support ASR in 25 EU languages, speech translation (AST), and punctuation/capitalization (PnC).

• Canary-1B V2 — Flagship: 25 languages, PnC, timestamps

• Canary-Qwen-2.5B — English only, PnC, highest accuracy

• Canary-1B Flash / 180M Flash — Optimized for speed

Canary supports chunked and streaming inference.

Conformer

The Conformer [ASR-MODELS1] combines self-attention and convolution modules. NeMo supports CTC, Transducer, and HAT variants.

• Conformer-CTC: Non-autoregressive, uses EncDecCTCModelBPE

• Conformer-Transducer: Autoregressive, uses EncDecRNNTBPEModel

• Conformer-HAT: Separates labels and blank predictions for better external LM integration (paper)

Configs: examples/asr/conf/conformer/

Fast-Conformer

Fast Conformer has 8x depthwise convolutional subsampling and reduced kernel sizes, making it ~2.4x faster than standard Conformer with minimal quality loss. Supports Longformer-style local attention for audio >1 hour.

Configs: examples/asr/conf/fastconformer/

Cache-aware Streaming Conformer

Streaming models trained with limited right context for real-time inference with caching to avoid duplicate computation. Supports three modes: fully causal, regular look-ahead, and chunk-aware look-ahead (recommended).

• Tutorial notebook

• Simulation script: examples/asr/asr_cache_aware_streaming/speech_to_text_cache_aware_streaming_infer.py

• Supports multiple look-aheads with att_context_size lists

Configs: examples/asr/conf/fastconformer/cache_aware_streaming/

With Prompt Conditioning (RNN-T only): Cache-aware streaming RNN-T model with language-ID prompt conditioning for multilingual ASR via EncDecRNNTBPEModelWithPrompt. The streaming inference script accepts a target_lang flag to select the prompt at runtime (see RNN-T with Prompt Conditioning Configuration). Config: fastconformer_transducer_bpe_streaming_prompt.yaml

Multitalker Streaming

Streaming multi-speaker ASR based on cache-aware FastConformer with speaker kernel injection [ASR-MODELS3]. Deploys one model instance per speaker for robust transcription of overlapped speech.

• Model card

• Tutorial

Hybrid-Transducer-CTC

Models with both RNN-T and CTC decoders trained jointly. Switch at inference time via asr_model.change_decoding_strategy(decoder_type='ctc' or 'rnnt').

• EncDecHybridRNNTCTCBPEModel (BPE) / EncDecHybridRNNTCTCModel (char)

• Configs: examples/asr/conf/fastconformer/hybrid_transducer_ctc/

With Prompt Conditioning: Extends Hybrid models with learnable prompt embeddings for multilingual/multi-domain ASR via EncDecHybridRNNTCTCBPEModelWithPrompt. Config: fastconformer_hybrid_transducer_ctc_bpe_prompt.yaml

References

[ASR-MODELS1]

Anmol Gulati, James Qin, Chung-Cheng Chiu, Niki Parmar, Yu Zhang, Jiahui Yu, Wei Han, Shibo Wang, Zhengdong Zhang, Yonghui Wu, and others. Conformer: convolution-augmented transformer for speech recognition. arXiv preprint arXiv:2005.08100, 2020.

[ASR-MODELS2]

Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems, 6000–6010. 2017.

[ASR-MODELS3]

Weiqing Wang, Taejin Park, Ivan Medennikov, Jinhan Wang, Kunal Dhawan, He Huang, Nithin Rao Koluguri, Jagadeesh Balam, and Boris Ginsburg. Speaker Targeting via Self-Speaker Adaptation for Multi-talker ASR. In Interspeech 2025, 5498–5502. 2025. doi:10.21437/Interspeech.2025-2142.