NeMo FastConformer

FastConformer is NVIDIA's efficient evolution of the Conformer speech-recognition architecture, distributed through the open-source NeMo toolkit. It replaces standard convolutional subsampling with 8x depthwise-separable subsampling and shrinks kernel sizes, making the encoder about 2.4x faster than a regular Conformer with minimal quality loss. Available in CTC and Transducer (RNNT) variants with BPE subword modeling, it underpins NVIDIA's production ASR stack — including the Parakeet TDT models that beat Whisper large-v3 on English benchmarks and the Nemotron 3.5 ASR streaming models covering 40 language-locales.

The honest picture: FastConformer is an architecture family rather than a single checkpoint, so 'using it' means picking a specific NeMo checkpoint (Parakeet, Nemotron, or a stock model) with its own language coverage and license. The NeMo toolkit's dependency pins can conflict with modern transformers versions, complicating installation. But for teams building production speech recognition — especially streaming and multilingual — it remains one of the most battle-tested open ASR foundations.