recipe(wav2vec2): add jakeBland/wav2vec-vm-finetune (audio-classification)

[codykk]

Add audio-classification recipes for jakeBland/wav2vec-vm-finetune

Adds fp16, w8a16, and w8a8 quantized recipe variants for jakeBland/wav2vec-vm-finetune (Wav2Vec2ForSequenceClassification, audio-classification task). The w8a8 variant delivers a 2.1x speedup and 71% memory reduction vs the fp32 baseline on CPU.

---

1. Recipe path(s)

• examples/recipes/jakeBland_wav2vec-vm-finetune/cpu/cpu/audio-classification_fp16_config.json
• examples/recipes/jakeBland_wav2vec-vm-finetune/cpu/cpu/audio-classification_w8a16_config.json
• examples/recipes/jakeBland_wav2vec-vm-finetune/cpu/cpu/audio-classification_w8a8_config.json

2. README row

Added jakeBland/wav2vec-vm-finetune | audio-classification to examples/recipes/README.md.

3. Build output dir

temp/wav2vec_vm_fp16/, temp/wav2vec_vm_w8a16/, temp/wav2vec_vm_w8a8/

4. Build log

All three variants: ✅ Build complete in ~70-77s

5. Appended findings

No methodology friction observed. Model uses standard wav2vec2 architecture fully covered by winml's existing Wav2Vec2OnnxConfig.

6. Optimum-coverage probe

{vendor: [], after_winml: ['audio-classification', 'audio-frame-classification', 'audio-xvector', 'automatic-speech-recognition', 'feature-extraction'], added_by_winml: [...all...], verdict: 'WINML-ONLY'}

7. Claimed (Effort, Goal, Outcome)

• Effort: L0★ (new recipe, no source edits)
• Goal ceiling: L1 (build + perf)
• Outcome: L0 (recipe + README row)

8. Goal-ladder verdict table

Tier	Verdict	Evidence
L0	PASS	All three recipes build successfully (fp16: 69.8s, w8a16: 77.0s, w8a8: 72.6s)
L1	PASS	All three run on CPU EP without crash (see perf table below)

9. Methodology-evolution declaration

No methodology friction observed.

10. Perf & eval data

Baseline (fp32, no recipe, origin/main @ 2a2ff548):

EP / Device	Precision	Verdict	Mean	p50	Throughput	RAM Δ	Model Size
CPUExecutionProvider / cpu	fp32	PASS	87.69 ms	87.57 ms	11.40 samples/s	+1268 MB	1.2 GB

Recipe variants:

EP / Device	Precision	Verdict	Mean	p50	Throughput	RAM Δ	Model Size
CPUExecutionProvider / cpu	fp16	PASS	91.09 ms	90.82 ms	10.98 samples/s	+1279 MB	602 MB
CPUExecutionProvider / cpu	w8a16	PASS	116.09 ms	114.76 ms	8.61 samples/s	+464 MB	303 MB
CPUExecutionProvider / cpu	w8a8	PASS	42.12 ms	41.67 ms	23.74 samples/s	+373 MB	303 MB

Delta vs baseline:

• w8a8: 2.1x faster, 75% smaller model, 71% less RAM
• w8a16: 75% smaller model, 63% less RAM (latency higher due to dequantize overhead on CPU)
• fp16: 50% smaller model (latency similar on CPU which doesn't natively accelerate fp16)

11. Component / op-level data

N/A — runtime check rules not available on this host (no parquet files).

12. Reproducible commands

# Baseline (fp32, no recipe)
winml build -m jakeBland/wav2vec-vm-finetune -o temp/baseline_wav2vec_vm --ep cpu --device cpu --no-analyze --no-optimize --no-quant --no-compile --rebuild
winml perf -m temp/baseline_wav2vec_vm/model.onnx --device cpu --ep cpu

# fp16
winml build -c examples/recipes/jakeBland_wav2vec-vm-finetune/cpu/cpu/audio-classification_fp16_config.json -m jakeBland/wav2vec-vm-finetune -o temp/wav2vec_vm_fp16/ --rebuild
winml perf -m temp/wav2vec_vm_fp16/model.onnx --device cpu --ep cpu

# w8a16
winml build -c examples/recipes/jakeBland_wav2vec-vm-finetune/cpu/cpu/audio-classification_w8a16_config.json -m jakeBland/wav2vec-vm-finetune -o temp/wav2vec_vm_w8a16/ --rebuild
winml perf -m temp/wav2vec_vm_w8a16/model.onnx --device cpu --ep cpu

# w8a8
winml build -c examples/recipes/jakeBland_wav2vec-vm-finetune/cpu/cpu/audio-classification_w8a8_config.json -m jakeBland/wav2vec-vm-finetune -o temp/wav2vec_vm_w8a8/ --rebuild
winml perf -m temp/wav2vec_vm_w8a8/model.onnx --device cpu --ep cpu