wav2vec2-keyword-spotting for Intel NPUs, GPUs and CPUs: w2v2_frontend_fp16 and w2v2_head_fp16 from litert-community/wav2vec2-keyword-spotting as ONNX (tf2onnx 1.17.0, opset 17), OpenVINO IR with fp32 weights, and precompiled Intel NPU blobs. tf2onnx 1.17.0 drops the bias of 24 FULLY_CONNECTED layers of w2v2_head_fp16; these files were converted with a one-line fix of that handler (convert/patch_tf2onnx.py in the GitHub repo), and their outputs match the .tflite. On the NPU of a Core Ultra 9 386H the blob runs in 7.53 ms (w2v2_frontend_fp16) and 6.87 ms (w2v2_head_fp16) per run, numpy input to numpy output.
Scripts, raw records, and the same table for all 52 graphs of LiteRT-Models: https://github.com/john-rocky/windows-zoo
| file | bytes | sha256 |
|---|---|---|
w2v2_frontend_fp16.onnx |
18,478,369 | 51645678a786b40dad1834d998943b95488858f5181b4cdf5f8822db6095ce63 |
openvino/w2v2_frontend_fp16.xml |
183,708 | ccdf923947def70198df847a6ed07ded583f5e40e22304c2d34f3fad8e9c3ed5 |
openvino/w2v2_frontend_fp16.bin |
18,380,412 | 3a3d7d02bf8b3f25865e4b8133da47b9867b57a6ed7b667e35a39241876ae6e3 |
npu/w2v2_frontend_fp16.blob |
11,481,217 | 394cbec179b17382b55ffc121763824a04c765e5a1235791acdd79d60e7ec5af |
w2v2_head_fp16.onnx |
360,612,481 | daadaee04b8b8f5ce8cc0038e05561cec8e5b0df27f81d55254f43823d6633d7 |
openvino/w2v2_head_fp16.xml |
1,158,918 | f54cbcf27bb4c68f79217d2fded2a1fdc2ba80994b22772e7133faae2924831a |
openvino/w2v2_head_fp16.bin |
359,904,956 | 19dc678653b957a5e85736287231ec7160c7329ec54f0224ddfc5526c534103b |
npu/w2v2_head_fp16.blob |
188,002,433 | a3f9872809bbb7e22c4630aac2946e0bd721e51af3ab6d03757044eba1cd069b |
The NPU blobs were compiled with OpenVINO 2026.4.1, Intel NPU driver 32.0.100.4841 (OpenVINO reports NPU_DRIVER_VERSION 1004841, NPU_COMPILER_VERSION 524292). If import_model refuses a blob on another driver or OpenVINO version, compile the IR instead (the commented line in the snippet).
MSI Stealth 16 AI+ (Core Ultra 9 386H, RTX 5080 Laptop GPU at a 95 W power limit, Windows 11 build 26200.9711, on AC power), 2026-11-06 to 2026-11-07. On the laptop, each row is one process: load, one run, 5 warmup runs, then the median of 30 timed runs (2 warmup + 10 timed runs when the first run takes more than 3 s), numpy input to numpy output, every runtime at its default settings, one process at a time, each started after the CPU, the RTX and the Intel GPU had gone idle (the thresholds are in METHOD.md of the GitHub repo). Output check: each output against the CPU output in fp32 on the same input (ONNX Runtime's CPU EP); PASS = cosine ≥ 0.99, no NaN or Inf, and top-1 agreement ≥ 0.99 for class scores. The Snapdragon rows are Qualcomm AI Hub profile jobs of the same ONNX files (their own timer). No Snapdragon rows for w2v2_head_fp16: Qualcomm AI Hub measured the file before the bias fix (its numbers are in the GitHub table).
| processor | runtime | file | ms | load s | output check |
|---|---|---|---|---|---|
| Intel AI Boost NPU | OpenVINO 2026.4.1 | npu/ blob (import_model) |
7.53 | 0.01 | FAIL: output_0_output: cos 0.8997 < 0.99 |
| Intel AI Boost NPU | ONNX Runtime 1.31.1 + OpenVINO EP 1.8.1 | .onnx, compiled when the session loads |
9.61 | 1.82 | FAIL: output_0_output: cos 0.8999 < 0.99 |
| Intel Graphics (iGPU) | ONNX Runtime 1.31.1 + OpenVINO EP 1.8.1 | .onnx |
4.26 | 0.35 | PASS (cos ≥ 0.9999) |
| RTX 5080 Laptop (95 W limit) | ONNX Runtime 1.31.1, CUDA EP | fp16 copy of the .onnx (not in this repo) |
1.51 | 1.13 | PASS (cos ≥ 0.9999) |
| RTX 5080 Laptop (95 W limit) | ONNX Runtime 1.31.1, CUDA EP | .onnx |
2.24 | 1.05 | PASS (cos ≥ 0.9999) |
| Core Ultra 9 386H CPU | ONNX Runtime 1.31.1, CPU EP | .onnx |
12.05 | 0.05 | reference |
| Snapdragon X2 Elite NPU | Qualcomm AI Hub, precompiled QNN ONNX | .onnx compiled by AI Hub |
11.32 | 0.37 | not compared (AI Hub profile job, 2026-10-30) |
| Snapdragon X2 Elite NPU | Qualcomm AI Hub, ONNX Runtime QNN EP | .onnx, compiled when it loads |
13.61 | 7.60 | not compared (AI Hub profile job, 2026-10-30) |
| Galaxy S26 NPU | Qualcomm AI Hub, precompiled QNN ONNX | .onnx compiled by AI Hub |
11.94 | 0.17 | not compared (AI Hub profile job, 2026-10-30) |
The ONNX file against the source .tflite (LiteRT interpreter on the CPU, the same input): PASS, cosine ≥ 0.9999 for every output.
| processor | runtime | file | ms | load s | output check |
|---|---|---|---|---|---|
| Intel AI Boost NPU | OpenVINO 2026.4.1 | npu/ blob (import_model) |
6.87 | 0.11 | PASS (cos ≥ 0.9999) |
| Intel AI Boost NPU | ONNX Runtime 1.31.1 + OpenVINO EP 1.8.1 | .onnx, compiled when the session loads |
7.27 | 19.44 | PASS (cos ≥ 0.9999) |
| Intel Graphics (iGPU) | ONNX Runtime 1.31.1 + OpenVINO EP 1.8.1 | .onnx |
6.72 | 2.89 | PASS (cos ≥ 0.9999) |
| RTX 5080 Laptop (95 W limit) | ONNX Runtime 1.31.1, CUDA EP | fp16 copy of the .onnx (not in this repo) |
7.52 | 1.24 | PASS (cos ≥ 0.9992) |
| RTX 5080 Laptop (95 W limit) | ONNX Runtime 1.31.1, CUDA EP | .onnx |
8.59 | 1.35 | PASS (cos ≥ 0.9999) |
| Core Ultra 9 386H CPU | ONNX Runtime 1.31.1, CPU EP | .onnx |
17.71 | 0.47 | reference |
The ONNX file against the source .tflite (LiteRT interpreter on the CPU, the same input): PASS, cosine ≥ 0.9999 for every output.
Download this repo, then run the snippets from its folder:
pip install huggingface_hub
hf download mlboydaisuke/wav2vec2-keyword-spotting-OpenVINO --local-dir wav2vec2-keyword-spotting-OpenVINO
cd wav2vec2-keyword-spotting-OpenVINO
OpenVINO (pip install openvino==2026.4.1):
import numpy as np
import openvino as ov
core = ov.Core()
if "NPU" not in core.available_devices:
raise SystemExit(f"no Intel NPU visible to OpenVINO: {core.available_devices}")
with open("npu/w2v2_frontend_fp16.blob", "rb") as f:
compiled = core.import_model(f.read(), "NPU") # precompiled: no compile at load
# or compile the IR on this machine: compiled = core.compile_model("openvino/w2v2_frontend_fp16.xml", "NPU")
# replace the zeros with your preprocessed input (float32, the shapes printed by compiled.inputs)
feeds = {i.get_any_name(): np.zeros(list(i.shape), np.float32) for i in compiled.inputs}
outputs = compiled(feeds)
print({o.get_any_name(): outputs[o].shape for o in compiled.outputs})
ONNX Runtime with the OpenVINO EP (pip install onnxruntime==1.31.1 onnxruntime-ep-openvino==1.8.1):
import numpy as np
import onnxruntime as ort
import onnxruntime_ep_openvino as ov_ep
ort.register_execution_provider_library(ov_ep.get_ep_name(), ov_ep.get_library_path())
npu = [d for d in ort.get_ep_devices()
if d.ep_name == ov_ep.get_ep_name() and d.ep_metadata.get("ov_device") == "NPU"]
if not npu:
raise SystemExit("no Intel NPU visible to the OpenVINO EP")
so = ort.SessionOptions()
so.add_provider_for_devices(npu, {})
sess = ort.InferenceSession("w2v2_frontend_fp16.onnx", sess_options=so) # compiles for the NPU here
# replace the zeros with your preprocessed input (float32)
feeds = {i.name: np.zeros(i.shape, np.float32) for i in sess.get_inputs()}
outputs = sess.run(None, feeds)
print([o.shape for o in outputs])
The same lines run w2v2_head_fp16 with its own file names.
windows-model-zooThe branch windows-model-zoo adds the 2 files that windows-model-zoo runs. unfixed/: the conversion without the bias fix above, with its fp16 copy. To use the model, take main's files.
Apache-2.0, the licence of litert-community/wav2vec2-keyword-spotting. The upstream model: superb/wav2vec2-base-superb-ks.
w2v2_head_fp16.onnx 361 MBopenvino/w2v2_head_fp16.bin 360 MBw2v2_frontend_fp16.onnx 18 MBopenvino/w2v2_frontend_fp16.bin 18 MBlitert-community/wav2vec2-keyword-spotting