vllm_omni.diffusion.models.soulx_singer.modules.preprocess.stack ¶

Lazy-loaded SoulX preprocess model tree.

logger `module-attribute` ¶

logger = init_logger(__name__)

Unified nn.Module tree for SoulX-Singer audio preprocess.

device: device

device_str = device

hop_size = hop_size

target_sr = target_sr

verbose = verbose

weights = weights

ensure_lyric()

ensure_rmvpe()

ensure_rosvot()

ensure_segmenter()

ensure_vocal_sep()

extract_f0(
    vocal, sample_rate: int, *, f0_path: str | None = None
) -> ndarray

Extract F0 on the mel grid; ndarray inputs are spooled to a temp wav.

extract_vocal(
    audio: str | tuple[ndarray, int],
) -> tuple[ndarray, int]