Add ASR fp16 recipe for MahmoudAshraf/mms-300m-1130-forced-aligner

[ssss141414]

Add ASR fp16 recipe: MahmoudAshraf/mms-300m-1130-forced-aligner

1. Model & task

MahmoudAshraf/mms-300m-1130-forced-aligner @ automatic-speech-recognition, fp16. A CTC forced-alignment model built on facebook/mms-300m.

2. What this PR adds

• examples/recipes/MahmoudAshraf_mms-300m-1130-forced-aligner/automatic-speech-recognition_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 standard Wav2Vec2ForCTC model (architectures=["Wav2Vec2ForCTC"], vocab_size=31), not trust_remote_code. But winml build auto-config resolves automatic-speech-recognition to a seq2seq loader and fails:

Error: Unrecognized configuration class Wav2Vec2Config for this kind of
AutoModel: AutoModelForSpeechSeq2Seq. Model type should be one of DiaConfig,
... SpeechEncoderDecoderConfig, ... WhisperConfig.

winml's ASR auto-loader assumes seq2seq (Whisper/SpeechT5/SeamlessM4T family), but wav2vec2 ASR is a CTC head (single encoder + linear projection), which must be loaded via AutoModelForCTC. This recipe pins loader.model_class = AutoModelForCTC (generated via winml config --model-class AutoModelForCTC), making the model build. winml build has no --model-class/--task flag, so a checked-in recipe is the only override.

4. Real-engineering gate (_meta-038)

• Gate-1 (recipe vs winml config -t automatic-speech-recognition --model-class AutoModelForCTC autoconf): IDENTICAL.
• Gate-2 (build without recipe): FAILS (seq2seq loader rejects Wav2Vec2Config).

5. Build validation (CPU)

fp16 build exit 0 in 250.6 s (Export 93.9 + Optimize 42.9 + FP16 29.8). Final artifact 602 MB. fp16 confirmed: 426 FLOAT16 initializers, 0 FLOAT32.

6. Structural (L0)

IN input_values [1,16000] float32 (1 s @ 16 kHz, value_range [-1,1]). OUT logits [1,49,31] — 49 CTC frames × 31-token vocab (correct CTC alignment output; frame count scales with audio length).

7. Perf (L1, CPU / CPUExecutionProvider)

mean 134.84 ms, P50 133.91 ms, P90 139.96 ms, throughput 7.42 samples/sec (20 iters).

8. Scope / honest limits

• CPU-only. DML / QNN / OpenVINO / VitisAI / NvTensorRTRTX not validated (host availability). The README "all 10 buckets" bar is not independently verified here.
• No L2 parity vs PyTorch (CTC logits/alignment comparison deferred).
• No L3 eval (ASR default-dataset + WER metric not confirmed).

9. Output is frame logits, not text

The graph emits CTC frame logits [1, T, 31]; CTC collapse / forced-alignment decoding happens downstream. Feature extraction (waveform → input_values) is also outside the graph.

10. Requirements

Recipe pins model_class=AutoModelForCTC; no --trust-remote-code needed (standard architecture).

11. Testing done

fp16 build (exit 0), fp16 initializer dtype audit, winml perf on CPU. Reproducible via winml build -c <recipe> -m MahmoudAshraf/mms-300m-1130-forced-aligner.

12. Follow-ups (out of scope here)

• Non-CPU EP validation once hosts are available.
• L2 CTC-alignment parity harness.
• Skill-side note (separate Lane-A PR): winml's ASR auto-loader is seq2seq-only; CTC-family models (*ForCTC) need AutoModelForCTC. A future auto-detect of *ForCTC architectures would remove the need for this recipe override.

[xieofxie]

if without the config, could perf run it directly?

[ssss141414]

if without the 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)...

This is a standard Wav2Vec2ForCTC model (architectures=["Wav2Vec2ForCTC"], vocab_size=31), not trust_remote_code. But winml build auto-config resolves automatic-speech-recognition to a seq2seq loader and fails:

Error: Unrecognized configuration class Wav2Vec2Config for this kind of
AutoModel: AutoModelForSpeechSeq2Seq. Model type should be one of DiaConfig,
... SpeechEncoderDecoderConfig, ... WhisperConfig.

[xieofxie]

if without the 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)...

This is a standard Wav2Vec2ForCTC model (architectures=["Wav2Vec2ForCTC"], vocab_size=31), not trust_remote_code. But winml build auto-config resolves automatic-speech-recognition to a seq2seq loader and fails:

Error: Unrecognized configuration class Wav2Vec2Config for this kind of
AutoModel: AutoModelForSpeechSeq2Seq. Model type should be one of DiaConfig,
... SpeechEncoderDecoderConfig, ... WhisperConfig.

if it is the case, should we update the code to make it work?