Whisper-Tiny Encoder (Ambarella N1-655 CVflow v3 NVP)

Whisper-Tiny Encoder is the audio encoder stage of OpenAI Whisper-Tiny, mapping 30-second 80-bin log-Mel spectrograms into dense acoustic representations ([1, 1500, 384]) on the Ambarella N1-655 CVflow v3 NVP.

Original paper: Robust Speech Recognition via Large-Scale Weak Supervision

Model Configuration:

• Reference implementation: https://github.com/openai/whisper
• Original Weight: https://huggingface.co/onnx-community/whisper-tiny
• Dataset: OpenAI Multilingual Weakly Supervised Audio Dataset (680k hours)
• License: MIT (LICENSE)
• Support Cooper version:
  • CNNGen: [3.9.0 / n1-655x] (cavalry_gen 3.0.4)

Upstream Attribution & Modification Notice

• Upstream Copyright: Copyright (c) 2022 OpenAI (openai/whisper-tiny, onnx-community/whisper-tiny), licensed under the MIT License.
• Modification Notice: Modified by Ambarella International LLC (2026). Compiled the Whisper-Tiny audio encoder ([1, 80, 3000] -> [1, 1500, 384]) via Ambarella Cooper SDK CNNGen 3.9.0 (cavalry_gen 3.0.4) into CVflow v3 N1-655 (vdg_arch = 0xB) Cavalry DAG binary format (n1-655_whisper_tiny_encoder_amba_optimized.bin).

NVP Port & Runtime Specification

• Target SoC: Ambarella N1-655 (vdg_arch = 0xB, CVflow v3 cavalry_gen 3.0.4)
• Input port: input_features [1, 80, 3000] IEEE FP16 (16 kHz PCM, 30 s window, N_FFT=400, hop=160, 80-bin Slaney log10-Mel)
• Output port: last_hidden_state [1, 1500, 384] IEEE FP16

Model	Device	Compression	Model Link
Whisper	N1-655	Amba_optimized	Model_Link