| """Inference model wrapper and output formatting for the MuST-C en-de ST recipe. |
| |
| ``conf/inference.yaml`` uses both halves of this module:: |
| |
| model: |
| _target_: src.inference.Speech2TextShortSafe |
| output_fn: src.inference.build_output |
| |
| ESPnet3 decodes one sample at a time |
| (``espnet3/systems/base/inference_runner.py``: ``data = dataset[idx]``) and |
| writes one SCP file per key ``build_output`` returns, under |
| ``${inference_dir}/<test_name>/`` -- so the keys below become ``hyp.scp`` and |
| ``ref.scp``, which ``conf/metrics.yaml`` scores. |
| """ |
|
|
| from espnet2.bin.st_inference import Speech2Text |
| from espnet2.legacy.nets.beam_search import Hypothesis |
| from espnet2.legacy.nets.pytorch_backend.transformer.subsampling import ( |
| TooShortUttError, |
| ) |
|
|
|
|
| class Speech2TextShortSafe(Speech2Text): |
| """``Speech2Text`` that survives utterances too short to subsample. |
| |
| egs2 does not apply the 0.1-20 s filter to test sets (``st.sh``: "NOT |
| applying to test_sets to keep original data"), so tst-COMMON contains 54 of |
| 2,641 utterances below the 7 frames a conv2d-subsampling conformer needs -- |
| the shortest is 0.050 s. Decoding one raises ``TooShortUttError``. |
| |
| espnet2 already handles this, but in its CLI loop |
| (``espnet2/bin/st_inference.py``), not inside ``Speech2Text.__call__``; |
| espnet3 calls the class directly, so a single 0.05 s utterance otherwise |
| aborts the whole test set. |
| |
| The fallback mirrors egs2 exactly: emit ``" "`` and keep the utterance, so |
| its reference is still scored (as a miss) rather than dropped -- dropping |
| would inflate BLEU by removing the hardest items from the denominator. |
| """ |
|
|
| def __call__(self, *args, **kwargs): |
| """Decode one utterance, tolerating one too short to subsample. |
| |
| Forwarded to ``espnet2.bin.st_inference.Speech2Text``, whose only |
| argument is ``speech``. |
| |
| Args: |
| speech: input speech of shape (nsamples,) or (nsamples, nchannels=1) |
| |
| Returns: |
| n-best list of (text, token, token_int, hyp) |
| """ |
| try: |
| return super().__call__(*args, **kwargs) |
| except TooShortUttError: |
| hyp = Hypothesis(score=0.0, scores={}, states={}, yseq=[]) |
| return [(" ", ["<space>"], [2], hyp)] |
|
|
|
|
| def build_output(data, model_output, idx): |
| """Turn one decoding result into the dict ESPnet3 writes to SCP. |
| |
| Args: |
| data: The raw dataset sample. ``conf/inference.yaml``'s test entries set |
| ``return_utt_id: true``, so this carries a real MuST-C id. |
| model_output: n-best list; ``[0][0]`` is the best hypothesis text. |
| idx: Index of the sample within its test set. |
| |
| Returns: |
| dict with ``utt_id``, ``hyp`` and ``ref``. |
| """ |
| return { |
| "utt_id": data.get("utt_id", str(idx)), |
| "hyp": model_output[0][0], |
| "ref": data.get("text", ""), |
| } |
|
|
