| """Inference output helpers for the LibriTTS codec recipe.""" |
|
|
| from __future__ import annotations |
|
|
| import numpy as np |
|
|
|
|
| def build_output(data, model_output, idx): |
| """Format one AudioCoding roundtrip result for SCP/artifact writing. |
| |
| Args: |
| data: Dataset item (``inference: true`` mode), providing ``utt_id`` |
| and the ground-truth ``wav_path``. |
| model_output: Dict returned by |
| ``espnet2.bin.gan_codec_inference.AudioCoding.__call__``, |
| containing the resynthesized waveform under ``resyn_audio``. |
| idx: Dataset index, used as a fallback identifier. |
| |
| Returns: |
| Dict with ``utt_id``, the ground-truth wav path under ``ref``, and |
| the resynthesized waveform under ``wav`` (1-D float32 numpy array, |
| materialized as a WAV file via ``output_artifacts``). |
| |
| Raises: |
| RuntimeError: If *model_output* has no ``resyn_audio`` key. |
| |
| Examples: |
| Wired up from ``conf/inference.yaml`` as |
| ``output_fn: src.inference.build_output``; the runner calls it once |
| per test utterance: |
| ```python |
| out = build_output( |
| {"utt_id": "1089-134686-000002-000000", "wav_path": "/data/a.wav"}, |
| {"resyn_audio": torch.zeros(24000)}, |
| 0, |
| ) |
| sorted(out) # -> ['ref', 'utt_id', 'wav'] |
| out["wav"].shape # -> (24000,) |
| ``` |
| """ |
| utt_id = str(data.get("utt_id", idx)) |
| ref = str(data.get("wav_path", "")) |
| wav = model_output.get("resyn_audio") |
| if wav is None: |
| raise RuntimeError("Codec inference output does not contain 'resyn_audio'.") |
| if hasattr(wav, "detach"): |
| wav = wav.detach().cpu().numpy() |
| wav = np.asarray(wav, dtype=np.float32).reshape(-1) |
| return {"utt_id": utt_id, "ref": ref, "wav": wav} |
|
|
