| 123456789101112131415161718192021222324252627282930313233343536373839404142434445464748 |
- # Copyright (c) Alibaba, Inc. and its affiliates.
- from typing import Any, Dict
- import numpy as np
- import torch
- from modelscope.metainfo import Pipelines
- from modelscope.outputs import OutputKeys
- from modelscope.pipelines.base import Input, Pipeline
- from modelscope.pipelines.builder import PIPELINES
- from modelscope.utils.constant import Tasks
- @PIPELINES.register_module(
- Tasks.voice_conversion, module_name=Pipelines.voice_conversion)
- class VCPipeline(Pipeline):
- r"""ANS (Acoustic Noise Suppression) Inference Pipeline .
- When invoke the class with pipeline.__call__(), it accept only one
- parameter:
- inputs(str): the path of wav file
- """
- SAMPLE_RATE = 16000
- def __init__(self, model, **kwargs):
- """
- use `model` and `preprocessor` to create a kws pipeline for prediction
- Args:
- model: model id on modelscope hub.
- """
- super().__init__(model=model, **kwargs)
- self.model.eval()
- self.stream_mode = kwargs.get('stream_mode', False)
- def preprocess(self, inputs: Input, **preprocess_params) -> Dict[str, Any]:
- return inputs
- def forward(self, inputs: Dict[str, Any],
- **forward_params) -> Dict[str, Any]:
- with torch.no_grad():
- outputs = self.model(inputs)
- outputs *= 32768.
- outputs = np.array(outputs, 'int16').tobytes()
- return {OutputKeys.OUTPUT_PCM: outputs}
- def postprocess(self, inputs: Dict[str, Any], **kwargs) -> Dict[str, Any]:
- return inputs
|
