recipe(Common-Voice-Gender-Detection): audio-classification cpu recipes

[codykk]

Summary

Adds verified CPU recipes (fp32 and fp16) for prithivMLmods/Common-Voice-Gender-Detection, a Wav2Vec2ForSequenceClassification model for binary gender classification (female/male) from audio. Effort L0, Goal ceiling L2, Outcome L0. Both precisions pass build and perf; eval (L3) is CLI-BLOCKED due to audio-classification task not being supported for eval.

Model metadata

What the model does: Binary gender classification (female/male) from raw audio waveforms using a fine-tuned Wav2Vec2 encoder.

Primary user stories: Classify speaker gender from audio clips for analytics, content tagging, and demographic analysis.

Supported tasks: audio-classification

Model architecture:

Wav2Vec2ForSequenceClassification
├── Feature extractor (CNN: 7 conv layers, 512 channels)
├── Feature projection (512 → 1024)
├── Encoder stack x 24
│   ├── Self-attention (16 heads, 1024 dim)
│   ├── Feed-forward (1024 → 4096 → 1024, GELU)
│   └── Residual + LayerNorm
└── Classification head (1024 → 256 → 2: female/male)

• Source/confidence: pinned checkpoint config and Wav2Vec2ForSequenceClassification source (verified).

Validation and support evidence

Baseline

• Pinned main commit: 57dc6b20
• Build PASS, Perf PASS on main
• Auto-config produces identical fp32 recipe; fp16 adds quant block with mode: "fp16"

Goal

• Effort: L0 (recipe-only)
• Goal ceiling: L2
• Outcome: L0
• Success definition: Build PASS + Perf PASS for cpu/cpu fp32 and fp16

Outcome

L0 PASS for both tuples. L2 reached (cosine similarity 1.0 for both precisions). L3 CLI-BLOCKED: audio-classification task not supported for eval.

Shipped recipes:

• examples/recipes/prithivMLmods_Common-Voice-Gender-Detection/cpu/cpu/audio-classification_fp32_config.json
• examples/recipes/prithivMLmods_Common-Voice-Gender-Detection/cpu/cpu/audio-classification_fp16_config.json

Appended finding: wav2vec2-005 to model_knowledge/wav2vec2.json.

Per-EP/device/precision results

Tier	EP / Device	Precision	Verdict	Mean	Throughput	RAM Δ	Eval metric
L0	CPUExecutionProvider / cpu	fp32	PASS (26.6s, 360.9MB)	—	—	—	—
L1	CPUExecutionProvider / cpu	fp32	PASS	49.08 ms	20.38 sps	+401.5 MB	—
L2	CPUExecutionProvider / cpu	fp32	PASS	—	—	—	cosine=1.0, max_abs=1.05e-5
L3	CPUExecutionProvider / cpu	fp32	CLI-BLOCKED	—	—	—	—
L0	CPUExecutionProvider / cpu	fp16	PASS (32.6s, 180.5MB)	—	—	—	—
L1	CPUExecutionProvider / cpu	fp16	PASS	63.25 ms	15.81 sps	+422.6 MB	—
L2	CPUExecutionProvider / cpu	fp16	PASS	—	—	—	cosine=1.0, max_abs=3.10e-3
L3	CPUExecutionProvider / cpu	fp16	CLI-BLOCKED	—	—	—	—

L3 CLI-BLOCKED: audio-classification task not supported for eval. Rules parquet unavailable on host.

Delta

• fp32 recipe: identical to baseline auto-config output
• fp16 recipe: adds quant block with mode: "fp16" and default calibration settings
• examples/recipes/README.md untouched

Analyze summary — component level and op level

Static rule analysis: rules parquet unavailable on host. Analyze data could not be produced.

Component-level summary

Artifact	Status	Notes
fp32	ANALYZE-UNAVAILABLE	Rules parquet not available on host
fp16	ANALYZE-UNAVAILABLE	Rules parquet not available on host

Op-level summary

Not available — rules parquet unavailable on host.

Reproduce commands

# Build fp32
winml build --model prithivMLmods/Common-Voice-Gender-Detection --config examples/recipes/prithivMLmods_Common-Voice-Gender-Detection/cpu/cpu/audio-classification_fp32_config.json --output $OUT/fp32

# Build fp16
winml build --model prithivMLmods/Common-Voice-Gender-Detection --config examples/recipes/prithivMLmods_Common-Voice-Gender-Detection/cpu/cpu/audio-classification_fp16_config.json --output $OUT/fp16

# Perf fp32
winml perf --model $OUT/fp32 --runs 20

# Perf fp16
winml perf --model $OUT/fp16 --runs 20