ai-pc-benchmarks-transcription-intel-openvino

OpenVINO IR bundle for nvidia/parakeet-tdt-0.6b-v3, used by the AI PC Benchmarks Intel transcription workload.

Contents

• model/config.json — model type, feature size, subsampling.
• model/manifest.json — build metadata, reference performance numbers.
• model/vocab.txt — SentencePiece vocab, 8193 tokens, <blk> at id 8192.
• model/ir/preprocessor_fused_n240000.{xml,bin} — fixed-chunk preprocessor (15 s @ 16 kHz).
• model/ir/encoder_t1501.{xml,bin} — W8A8-quantized Conformer encoder.
• model/ir/decoder_joint_e1_t1.{xml,bin} — W8A8-quantized TDT decoder + joint.

Runtime

OpenVINO >= 2026.1. The IRs compile to Intel NPU, GPU, or CPU on first use and are cached via CACHE_DIR. No prebuilt NPU blobs ship here — the benchmark compiles them on-device the first time a given device is used.

License

CC-BY-4.0, inherited from the base model.