帧长 frame_len
帧移 frame_shift
重叠部分长度 frame_overlap

    /*
    * Cut data length of frame_overlap+buffer_shift+frame_overlap
    * If no enough data, zeros are filled in.
    * The data in fram_overlap just participate in model forward
    * All we care about is the data length of buffer_shift in middle
    * */
    int input_size = new_input_data.size();
    int frame_overlap = _opts.frame_len - _opts.frame_shift;
    int buffer_shift = _opts.sample_frequency * _opts.slice_second;
    for (int start = 0; start < input_size; start += buffer_shift) {
        int end = std::min(start + buffer_shift, input_size);
        int buffer_size = frame_overlap + (end - start) + frame_overlap;
        int start_pad = std::max(start - frame_overlap, 0);
        int end_pad = std::min(end + frame_overlap, input_size);
        int start_buffer = frame_overlap - (start - start_pad);
        int copy_buffer_len = end_pad - start_pad;
        std::vector<float> input_buffer(buffer_size, 0);
        memcpy(input_buffer.data() + start_buffer,
               new_input_data.data() + start_pad,
               copy_buffer_len * sizeof(float));
        std::vector<int64_t> input_dims = {1, (int64_t)input_buffer.size()};
        std::vector<std::vector<BaseFloat>> output_buffer;
        if (_inference_utils->run(input_buffer, input_dims, output_buffer)
            != 0) {
            IDEC_WARN << "Fail to get denoised data";
            return -1;
        }
        if (start == 0) {
            output_data.resize(output_buffer.size());
            for (size_t i = 0; i < output_data.size(); i++) {
                output_data[i].reserve(input_size);
            }
        }
        for (size_t i = 0; i < output_data.size(); i++) {
            IDEC_ASSERT(output_buffer[i].size() == buffer_size);
            output_data[i].insert(
                    output_data[i].end(),
                    output_buffer[i].data() + frame_overlap,
                    output_buffer[i].data() + frame_overlap + (end - start));
        }
    }