| """Inference output formatting for the ACE-Opencpop SVS recipe.""" | |
| def build_output(data, model_output, idx): | |
| """Turn one ``SingingGenerate`` result into the record the infer stage writes. | |
| ``conf/inference.yaml`` points ``output_fn`` here. ``wav`` is written as a | |
| wav file, while ``text`` and ``ref`` (the recorded singing) go to SCP files | |
| for the measure stage. | |
| Args: | |
| data: Dataset sample built with ``inference: true``. | |
| model_output: ``SingingGenerate`` output, whose ``wav`` is the | |
| synthesized singing. | |
| idx: Index of the sample within its test set (unused). | |
| Returns: | |
| Dict with ``utt_id``, ``text``, ``ref`` and ``wav``. | |
| """ | |
| return { | |
| "utt_id": data["utt_id"], | |
| "text": data["raw_text"], | |
| "ref": data["wav_path"], | |
| "wav": model_output["wav"].cpu().numpy(), | |
| } | |