potgieterdl/nemotron-speech-streaming-en-0.6b-onnx-int8

nemotron-speech-streaming-en-0.6b — ONNX, calibrated static int8

A calibration-based static-int8 quantization of the ONNX export of nvidia/nemotron-speech-streaming-en-0.6b (English streaming speech recognition, 0.6B parameters).

• Peak inference memory ≈ 0.8 GB (the fp32 ONNX export runs ≈ 2.6 GB).
• Word-error rate within noise of fp32 on standard read-speech material; whispered/very-quiet speech degrades somewhat relative to fp32.
• Files: encoder.onnx + encoder.onnx.data, decoder_joint.onnx, tokenizer.model.
• Runs with ONNX Runtime; the encoder is cache-aware for incremental/streaming inference.

Licensed under the NVIDIA Open Model License — see LICENSE and NOTICE. Base model © NVIDIA Corporation.