nemo_curator.stages.audio.alm.alm_data_builder

ALM Data Builder Stage - Native NeMo Curator Implementation.

Creates training windows from audio segments. Follows the exact pattern from NeMo Curator: https://github.com/NVIDIA-NeMo/Curator/blob/main/nemo_curator/stages/audio/common.py

Produces identical output to SDP implementation.

Module Contents

Classes

Name	Description
`ALMDataBuilderStage`	Build ALM training windows from audio segments.
`BuilderStats`	Tracks segment loss reasons and counts during window building.

Functions

Name	Description
`_compute_speaker_durations`	-
`_get_bandwidth`	-
`_record_window_loss`	Record statistics for a rejected window.
`_truncate_segment`	Truncate a segment’s words at the given end time, returning a modified copy.

Data

MIN_SEGMENTS_PER_WINDOW

API

class nemo_curator.stages.audio.alm.alm_data_builder.ALMDataBuilderStage(
    name: str = 'alm_data_builder',
    target_window_duration: float = 120.0,
    tolerance: float = 0.1,
    min_bandwidth: int = 8000,
    min_sample_rate: int = 16000,
    min_speakers: int = 2,
    max_speakers: int = 5,
    truncation: bool = True,
    drop_fields: str = 'words',
    drop_fields_top_level: str = 'words,segments'
)

Dataclass

Bases: ProcessingStage[AudioTask, AudioTask]

Build ALM training windows from audio segments.

Filters segments by sample rate, bandwidth, speaker count, and duration to create valid training windows. Mutates the entry dict in-place, adding windows, stats, and truncation_events keys.

drop_fields

str = 'words'

drop_fields_top_level

str = 'words,segments'

max_speakers

int = 5

min_bandwidth

int = 8000

min_sample_rate

int = 16000

min_speakers

int = 2

name

str = 'alm_data_builder'

target_window_duration

float = 120.0

tolerance

float = 0.1

truncation

bool = True

nemo_curator.stages.audio.alm.alm_data_builder.ALMDataBuilderStage.__post_init__() -> None

Compute derived parameters - EXACT match to SDP.

nemo_curator.stages.audio.alm.alm_data_builder.ALMDataBuilderStage._process_single_entry(
    entry_data: dict[str, typing.Any]
) -> dict[str, typing.Any]

Process a single entry and extract valid training windows.

nemo_curator.stages.audio.alm.alm_data_builder.ALMDataBuilderStage.inputs() -> tuple[list[str], list[str]]

nemo_curator.stages.audio.alm.alm_data_builder.ALMDataBuilderStage.process(
    task: nemo_curator.tasks.AudioTask
) -> nemo_curator.tasks.AudioTask

class nemo_curator.stages.audio.alm.alm_data_builder.BuilderStats(
    total_segments: int = 0,
    total_dur: float = 0.0,
    swift_path: str = '',
    audio_sample_rate: int = 0,
    lost_bw: int = 0,
    dur_lost_bw: float = 0.0,
    lost_sr: int = 0,
    dur_lost_sr: float = 0.0,
    lost_spk: int = 0,
    dur_lost_spk: float = 0.0,
    lost_win: int = 0,
    dur_lost_win: float = 0.0,
    lost_no_spkr: int = 0,
    dur_lost_no_spkr: float = 0.0,
    lost_next_seg_bm: int = 0,
    dur_lost_next_seg_bm: float = 0.0,
    lost_win_full_data: list = list()
)

Dataclass

Tracks segment loss reasons and counts during window building.

audio_sample_rate

int = 0

dur_lost_bw

float = 0.0

dur_lost_next_seg_bm

float = 0.0

dur_lost_no_spkr

float = 0.0

dur_lost_spk

float = 0.0

dur_lost_sr

float = 0.0

dur_lost_win

float = 0.0

lost_bw

int = 0

lost_next_seg_bm

int = 0

lost_no_spkr

int = 0

lost_spk

int = 0

lost_sr

int = 0

lost_win

int = 0

lost_win_full_data

list = field(default_factory=list)

swift_path

str = ''

total_dur

float = 0.0

total_segments

int = 0

nemo_curator.stages.audio.alm.alm_data_builder.BuilderStats.to_dict() -> dict[str, typing.Any]

nemo_curator.stages.audio.alm.alm_data_builder._compute_speaker_durations(
    segments: list[dict[str, typing.Any]]
) -> dict[str, float]

nemo_curator.stages.audio.alm.alm_data_builder._get_bandwidth(
    seg: dict[str, typing.Any]
) -> int

nemo_curator.stages.audio.alm.alm_data_builder._record_window_loss(
    stat: nemo_curator.stages.audio.alm.alm_data_builder.BuilderStats,
    seg: dict[str, typing.Any],
    segments: list[dict[str, typing.Any]],
    start_idx: int,
    curr_idx: int,
    window_segs: list[dict[str, typing.Any]],
    drop_fields: set[str],
    min_bandwidth: int
) -> None

Record statistics for a rejected window.

nemo_curator.stages.audio.alm.alm_data_builder._truncate_segment(
    seg: dict[str, typing.Any],
    truncated_end: float
) -> dict[str, typing.Any]

Truncate a segment’s words at the given end time, returning a modified copy.

nemo_curator.stages.audio.alm.alm_data_builder.MIN_SEGMENTS_PER_WINDOW = 2