diarizeapp/redimnet2-b3-onnx

PalabraAI ReDimNet2-B3 Speaker Embedding ONNX

Canonical repository for PalabraAI ReDimNet2-B3 Speaker Embedding ONNX maintained by the diarizeapp ecosystem.

Overview

State-of-the-art ReDimNet2-B3 speaker verification model extracting 192-dimensional speaker embeddings for short-turn duration-invariant speaker diarization. Provides both waveform-input (redimnet2_b3_lm.onnx) and STFT-free 80-channel log-mel filterbank input (redimnet2_b3_lm_fbank.onnx) for direct execution across CPU, DirectML, CUDA, and Qualcomm Hexagon NPU.

Hardware Acceleration & Execution Providers

• Hexagon HTP NPU (Snapdragon X Elite): Supports DirectML, CUDA, Hexagon HTP (QNN) via STFT-free fbank input graph.
• DirectML / CUDA: Fully supported via standard ONNX Runtime.
• CPU (x86_64 / ARM NEON): Universal SIMD fallback.

Provenance & Integrity

Maintained by diarizeapp to guarantee persistent asset availability and prevent 404 dependency breakages.