Conformer-CTC Small (Ambarella N1-655 CVflow v3 NVP)

Conformer-CTC Small is a convolution-augmented Transformer speech recognition model trained with Connectionist Temporal Classification (CTC) loss for English automatic speech recognition (ASR).

Original paper: Conformer: Convolution-augmented Transformer for Speech Recognition

Model Configuration:

• Reference implementation: https://github.com/NVIDIA/NeMo
• Original Weight: https://huggingface.co/csukuangfj/sherpa-onnx-nemo-ctc-en-conformer-small
• Dataset: NeMo ASRSet (LibriSpeech, Fisher, Switchboard, WSJ, NSC, Mozilla Common Voice)
• License: Apache-2.0 (LICENSE)
• Support Cooper version:
  • CNNGen: [3.9.0 / n1-655x] (cavalry_gen 3.0.4)

Upstream Attribution & Modification Notice

• Upstream Copyright: Copyright (c) NVIDIA CORPORATION & Affiliates (NVIDIA NeMo stt_en_conformer_ctc_small) and Copyright (c) Xiaomi Corporation / sherpa-onnx contributors (csukuangfj/sherpa-onnx-nemo-ctc-en-conformer-small), licensed under the Apache License, Version 2.0.
• Modification Notice (Apache-2.0 Section 4(b)): Modified by Ambarella International LLC (2026). Fixed dynamic time dimension to static [1, 80, 1024] (10.24 s window), quantized to mixed FP16/INT8 via Ambarella CNNGen (onnx_surgery -> vas -> cavalry_gen 3.0.4), and compiled to CVflow v3 N1-655 (vdg_arch = 0xB) Cavalry DAG binary format (n1-655_conformer_ctc_small_amba_optimized.bin).

NVP Port & Runtime Specification

• Target SoC: Ambarella N1-655 (vdg_arch = 0xB, CVflow v3 cavalry_gen 3.0.4)
• Input port: audio_signal [1, 80, 1024] IEEE FP16 (16 kHz PCM -> 80-bin Kaldi log-Mel, 25 ms Povey window, 10 ms hop, per-feature mean/std normalized)
• Output port: logprobs [1, 256, 1025] IEEE FP16 (CTC greedy decode with blank ID 1024 against tokens.txt)

Model	Device	Compression	Model Link
Conformer	N1-655	Amba_optimized	Model_Link