recipe(wav2vec2): add jonatasgrosman/wav2vec2-large-xlsr-53-polish ASR recipes

[codykk]

Summary

Add verified CPU fp32 and fp16 recipes for jonatasgrosman/wav2vec2-large-xlsr-53-polish, a Wav2Vec2ForCTC Polish automatic speech recognition model. The recipe forces model_class=AutoModelForCTC to work around the default AutoModelForSpeechSeq2Seq loader which rejects Wav2Vec2Config.

Model metadata

• What it does: Polish automatic speech recognition (ASR) using CTC decoding over a wav2vec2-large-xlsr-53 backbone fine-tuned on Polish speech data.
• Primary user stories: User supplies Polish audio waveform to obtain character-level CTC logits for Polish speech transcription.
• Supported tasks: automatic-speech-recognition (CTC)
• Architecture: Wav2Vec2ForCTC — convolutional feature encoder → transformer encoder → linear CTC head. Output: logits [1, 49, 40] (49 CTC frames × 40 Polish character vocab).

Validation and support evidence

Tier	Tuple	Verdict	Evidence
Baseline	winml build -m jonatasgrosman/wav2vec2-large-xlsr-53-polish (no recipe)	FAIL	Unrecognized configuration class Wav2Vec2Config for AutoModelForSpeechSeq2Seq
L0 build	cpu/cpu/fp32	PASS	Exit 0, 113.3s (Export 49.8s + Optimize 56.6s), 1.2 GB
L0 build	cpu/cpu/fp16	PASS	Exit 0, 104.0s (Export 48.0s + Optimize 52.9s)
L1 perf	cpu/cpu/fp32	PASS	mean=132.68ms, P50=132.07ms, P90=138.27ms, 7.54 sps, +1262 MB RAM
L1 perf	cpu/cpu/fp16	PASS	mean=130.41ms, P50=129.56ms, P90=135.35ms, 7.67 sps, +1261 MB RAM
L2 parity	cpu/cpu/fp32	PASS	cosine=0.99999988, max_abs=1.3e-4 vs PyTorch
L2 parity	cpu/cpu/fp16	PASS	cosine=0.99999988, max_abs=1.3e-4 vs PyTorch
L3 eval	cpu/cpu/*	CLI-BLOCKED	automatic-speech-recognition not in winml eval task registry

Delta: Baseline FAIL → recipe PASS. The recipe forces loader.model_class=AutoModelForCTC (CTC loader instead of default seq2seq).

Reproduce:

winml build -m jonatasgrosman/wav2vec2-large-xlsr-53-polish -c examples/recipes/jonatasgrosman_wav2vec2-large-xlsr-53-polish/cpu/cpu/automatic-speech-recognition_fp32_config.json -o out/fp32 --ep cpu --device cpu
winml perf -m out/fp32/model.onnx --ep cpu --device cpu

Environment: winml 0.2.0, origin/main @ 67169d45, CPU host only.