Add audio-classification fp16 recipe for audeering/wav2vec2-large-robust-24-ft-age-gender

[ssss141414]

Add audio-classification fp16 recipe: audeering/wav2vec2-large-robust-24-ft-age-gender

1. Model & task

audeering/wav2vec2-large-robust-24-ft-age-gender @ audio-classification, fp16.

2. What this PR adds

• examples/recipes/audeering_wav2vec2-large-robust-24-ft-age-gender/audio-classification_fp16_config.json (new recipe)
• examples/recipes/README.md (index row + total bump)

Recipe-only. No source code changed (Effort-L0★).

3. Why it's needed (the engineering delta)

This is a custom trust_remote_code model — config.json architectures=["Model"] (an AgeGenderModel: a wav2vec2 backbone with age-regression + gender-classification heads), not a standard Wav2Vec2ForSequenceClassification.

Because of the custom head, winml build auto-config mis-resolves the task to text-classification, and the build fails:

Error: wav2vec2 doesn't support task text-classification for the onnx backend.
Supported tasks are: feature-extraction, automatic-speech-recognition,
audio-classification, audio-frame-classification, audio-xvector.

This recipe pins task=audio-classification (generated via winml config -t audio-classification --trust-remote-code), which makes the model build. winml build has no --task flag, so a checked-in recipe is the only way to override the mis-resolved task.

4. Real-engineering gate (_meta-038)

• Gate-1 (recipe vs winml config -t audio-classification autoconf): IDENTICAL.
• Gate-2 (build without recipe): FAILS (auto-config picks text-classification).

Passing Gate-2-fails-without-recipe ⇒ genuine delta, not catalog padding.

5. Build validation (CPU)

Precision	Result	Time
fp32	exit 0	210.5 s (Export 96.5 + Optimize 45.5)
fp16	exit 0	166.8 s (Export 93.8 + Optimize 41.6 + FP16 29.4)

fp16 artifact confirmed genuinely fp16: 428 FLOAT16 initializers, 0 FLOAT32.

6. Structural (L0)

IR 8, opset 17. IN input_values [1,16000] float32 (1 s @ 16 kHz, value_range [-1,1]). OUT logits [1,3] float32.

7. Perf (L1, CPU / CPUExecutionProvider)

mean 133.95 ms, P50 132.06 ms, P90 140.35 ms, throughput 7.47 samples/sec (20 iters). Model-load +1227 MB RAM, inference +43.7 MB.

8. Scope / honest limits

• CPU-only. DML / QNN / OpenVINO / VitisAI / NvTensorRTRTX not validated (host availability).
• No L2 parity vs PyTorch. The eager AgeGenderModel.forward returns an age+gender tuple, not a single logits tensor; the exported audio-classification head is a 3-logit output. A parity harness would need a custom adapter — deferred.
• No L3 eval (audio-classification default-dataset availability not confirmed).

9. Note on the exported head

The exported graph produces logits[1,3], i.e. the winml audio-classification export path, not the model's native {age scalar, gender 2-way} multi-head. Downstream consumers should expect a 3-logit classification output.

10. Requirements

--trust-remote-code is required at both config and build time (custom modeling code).

11. Testing done

Two builds (fp32 + fp16, both exit 0), fp16 initializer dtype audit, winml perf on CPU. Reproducible from the checked-in recipe via winml build -c <recipe> -m audeering/wav2vec2-large-robust-24-ft-age-gender --trust-remote-code.

12. Follow-ups (out of scope here)

• Non-CPU EP validation once hosts are available.
• L2 parity adapter for the custom multi-head model.
• Skill-side note (separate Lane-A PR): winml build has no --task; the only task override is a pre-generated recipe.

[xieofxie]

if without config, could perf run it directly?

[ssss141414]

if without config, could perf run it directly?

As PR's description, it will failed on build stage without config. And the following command will not be run (perf/ eval)...

Because of the custom head, winml build auto-config mis-resolves the task to text-classification, and the build fails:

Error: wav2vec2 doesn't support task text-classification for the onnx backend.
Supported tasks are: feature-extraction, automatic-speech-recognition,
audio-classification, audio-frame-classification, audio-xvector.

[xieofxie]

if without config, could perf run it directly?

As PR's description, it will failed on build stage without config. And the following command will not be run (perf/ eval)...

Because of the custom head, winml build auto-config mis-resolves the task to text-classification, and the build fails:

Error: wav2vec2 doesn't support task text-classification for the onnx backend.
Supported tasks are: feature-extraction, automatic-speech-recognition,
audio-classification, audio-frame-classification, audio-xvector.

same as another pr