recipe(wav2vec2-deepfake-audio): add CPU audio classification configs

[fuliucansheng]

Summary

Adds CPU recipe coverage for Gustking/wav2vec2-large-xlsr-deepfake-audio-classification, a Wav2Vec2 audio classifier for real/fake speech detection. This is an L0 recipe-only contribution: current main already resolves the model through Optimum's Wav2Vec2 audio-classification path, and this PR records verified CPU fp32/fp16 coverage with checked-in recipes.

Model metadata

What the model does: Gustking/wav2vec2-large-xlsr-deepfake-audio-classification classifies audio waveforms into real and fake labels. WinML exports it as audio-classification with Wav2Vec2ForSequenceClassification.

Primary user stories:

Story	Confidence
A user supplies a 16 kHz audio waveform to obtain real/fake logits for deepfake-audio detection.	verified from checkpoint config labels, feature extractor metadata, and architecture

Supported tasks:

Task	Surface	Confidence
audio-classification	Transformers / Optimum ONNX / WinML export surface	verified

Model architecture:

Wav2Vec2ForSequenceClassification
├── Wav2Vec2Model
│   ├── convolutional feature encoder over waveform samples
│   ├── feature projection (hidden 1024)
│   └── encoder stack x24
│       ├── self-attention (16 heads)
│       └── feed-forward + GELU + residual/layernorm
└── sequence-classification head -> logits [B, 2] (`real`, `fake`)

Source/confidence: pinned checkpoint config resolves model_type: wav2vec2, architectures: [Wav2Vec2ForSequenceClassification], hidden size 1024, 24 layers, 16 attention heads, 2 labels, and Wav2Vec2FeatureExtractor.

Validation and support evidence

Baseline

Baseline was run from current origin/main at ace130f20e36df44ebf7de6cf0356235c8ee79fd with winml, version 0.2.0.

Probe	Result
winml inspect	PASS: loader AutoModelForAudioClassification, exporter Wav2Vec2OnnxConfig, WinML class WinMLModelForGenericTask; input input_values [1,16000] float32; output logits.
Optimum coverage	VENDOR-ONLY for wav2vec2: vendor ONNX tasks include audio-classification; WinML adds no extra task override.
Recipe-free build	PASS: build complete in 81.2s; model loaded as Wav2Vec2ForSequenceClassification with 409 modules and 315.7M parameters; export tagged 1144/1144 nodes; final artifact emitted.
Recipe-free perf	PASS on CPU fp32: avg 53.87 ms, p50 53.97 ms, throughput 18.56 samples/sec, RAM total +1267.2 MB.
L3 task metric	CLI-BLOCKED: winml eval --schema --task audio-classification reports the task is not supported by the eval registry, so no task metric is available through the CLI today.

Baseline winml config -p fp32 --ep cpu --device cpu and winml config -p fp16 --ep cpu --device cpu are the sources for the checked-in recipes.

Goal

Effort: L0. Goal ceiling: L2. Outcome: L0 recipe-only.

Success means the checked-in CPU fp32/fp16 recipes build structurally valid ONNX artifacts, run winml perf on CPU, and match the PyTorch reference numerically with the same waveform input. L3 is explicitly blocked by the missing audio-classification eval task.

Outcome

Highest verified goal: L2 PASS on CPU for both fp32 and fp16. Coverage is full for the targeted EP set (cpu/cpu/fp32, cpu/cpu/fp16); no accelerator tuples are claimed.

Shipped recipes:

Recipe
examples/recipes/Gustking_wav2vec2-large-xlsr-deepfake-audio-classification/cpu/cpu/audio-classification_fp32_config.json
examples/recipes/Gustking_wav2vec2-large-xlsr-deepfake-audio-classification/cpu/cpu/audio-classification_fp16_config.json

Methodology friction observed: none. Model finding: large Wav2Vec2 sequence classifiers export cleanly through existing audio-classification; task-metric eval remains CLI-blocked until WinML adds audio-classification evaluation support.

Per-EP/device/precision results

Tier	EP / Device	Precision	Verdict	Evidence
L0 build	CPUExecutionProvider / cpu	fp32	PASS	Recipe build complete in 64.1s; optimized model is loadable with input input_values [1,16000] float32, output logits [1,2], 897 nodes, FLOAT initializers, and 1204.3 MB external data colocated with model.onnx.
L0 build	CPUExecutionProvider / cpu	fp16	PASS	Recipe build complete in 77.8s with --precision fp16; fp16 stage complete in 13.1s; optimized model is loadable with 899 nodes, FLOAT16 initializers, and 602.2 MB external data.
L1 perf	CPUExecutionProvider / cpu	fp32	PASS	Avg 55.87 ms, p50 56.70 ms, throughput 17.90 samples/sec, RAM total +1267.4 MB.
L1 perf	CPUExecutionProvider / cpu	fp16	PASS	Avg 75.01 ms, p50 73.46 ms, throughput 13.33 samples/sec, RAM total +1276.0 MB; perf reports model precision fp16.
L2 compare	CPUExecutionProvider / cpu	fp32	PASS	PyTorch vs ONNX logits on the same deterministic waveform: cosine 1.0, max absolute diff 1.6391277313232422e-06.
L2 compare	CPUExecutionProvider / cpu	fp16	PASS	PyTorch vs ONNX logits on the same deterministic waveform: cosine 1.0, max absolute diff 0.0008488893508911133.
L3 eval	CPUExecutionProvider / cpu	fp32	CLI-BLOCKED	winml eval --schema --task audio-classification reports unsupported task; supported task list does not include audio-classification.
L3 eval	CPUExecutionProvider / cpu	fp16	CLI-BLOCKED	Same unsupported-task blocker as fp32; no task-metric value is reported.

Delta

Both recipes are generated from current winml config output with explicit CPU targeting. The fp32 recipe preserves quant: null; the fp16 recipe preserves the generated fp16 quantization block and model metadata. No source code or tests are changed, and examples/recipes/README.md is untouched.

Reducibility: recipe-only is appropriate because auto-config already resolves task, loader, exporter, and WinML inference class correctly; this PR is a per-checkpoint verified CPU coverage declaration rather than a class-wide code fix.

Analyze summary

Static rule analysis completed with partial-success exit semantics: QNN/OpenVINO rows were complete, while VitisAI has no rule data and is all-unknown. This is compatibility analysis, not runtime execution.

Component-level summary

Artifact	Architecture coverage	Mapping	Actionable EP findings
fp32	convolutional feature encoder; 24x Wav2Vec2 encoder attention/FFN; sequence-classification head	1144 exported nodes tagged during HTP export; 897 optimized ONNX ops analyzed	QNN NPU partial: GELU-related Div, Erf, Add, Mul; OpenVINO NPU fully supported; VitisAI: no rule data.
fp16	convolutional feature encoder; 24x Wav2Vec2 encoder attention/FFN; sequence-classification head	1144 exported nodes tagged during HTP export; 899 optimized ONNX ops analyzed	QNN NPU partial: GELU-related Div, Erf, Add, Mul; OpenVINO NPU fully supported; VitisAI: no rule data.

Op-level summary

Artifact	Graph	Dominant ops	EP roll-up
fp32	897 ops / 14 types	Add 227; MatMul 194; Mul 112; Transpose 111; Reshape 96; LayerNormalization 57; Div 32; Erf 32	QNN: 865 supported, 32 partial, 0 unsupported, 0 unknown. OpenVINO: 897 supported. VitisAI: 897 unknown.
fp16	899 ops / 15 types	Add 227; MatMul 194; Mul 112; Transpose 111; Reshape 96; LayerNormalization 57; Div 32; Erf 32	QNN: 867 supported, 32 partial, 0 unsupported, 0 unknown. OpenVINO: 899 supported. VitisAI: 899 unknown.

Reproduce commands

$MODEL = "Gustking/wav2vec2-large-xlsr-deepfake-audio-classification"
$OUT = "temp/repro_Gustking_wav2vec2_large_xlsr_deepfake_audio_classification"

uv run winml inspect -m $MODEL --format json
uv run winml config -m $MODEL -p fp32 --ep cpu --device cpu -o "$OUT/config_fp32.json"
uv run winml config -m $MODEL -p fp16 --ep cpu --device cpu -o "$OUT/config_fp16.json"
uv run winml build -c examples/recipes/Gustking_wav2vec2-large-xlsr-deepfake-audio-classification/cpu/cpu/audio-classification_fp32_config.json -m $MODEL -o "$OUT/fp32" --ep cpu --device cpu --no-analyze --no-compile --rebuild
uv run winml build -c examples/recipes/Gustking_wav2vec2-large-xlsr-deepfake-audio-classification/cpu/cpu/audio-classification_fp16_config.json -m $MODEL -o "$OUT/fp16" --ep cpu --device cpu --precision fp16 --no-analyze --no-compile --rebuild
uv run winml perf -m "$OUT/fp32/model.onnx" --ep cpu --device cpu --iterations 3 --warmup 1
uv run winml perf -m "$OUT/fp16/model.onnx" --ep cpu --device cpu --iterations 3 --warmup 1
uv run winml eval --schema --task audio-classification
uv run winml analyze --model "$OUT/fp32/model.onnx" --ep all --output "$OUT/fp32/analyze_all.json"
uv run winml analyze --model "$OUT/fp16/model.onnx" --ep all --output "$OUT/fp16/analyze_all.json"