nemo_curator.stages.audio.tagging.split

Audio Splitting and Joining Stages.

Module Contents

Classes

Name	Description
`JoinSplitAudioMetadataStage`	Stage for joining metadata of previously split audio files.
`SplitASRAlignJoinStage`	Composite stage: Split long audio -> ASR align -> Join results.
`SplitLongAudioStage`	Stage that splits long audio files into smaller segments.

API

class nemo_curator.stages.audio.tagging.split.JoinSplitAudioMetadataStage(
    name: str = 'JoinSplitAudioMetadata'
)

Dataclass

Bases: ProcessingStage[AudioTask, AudioTask]

Stage for joining metadata of previously split audio files.

Combines the metadata (transcripts and alignments) of audio files that were previously split by SplitLongAudioStage. Adjusts timestamps and concatenates transcripts to recreate the original audio’s metadata.

name

str = 'JoinSplitAudioMetadata'

nemo_curator.stages.audio.tagging.split.JoinSplitAudioMetadataStage._join_split_metadata(
    meta_entry: dict
) -> None

Join metadata from split audio files.

nemo_curator.stages.audio.tagging.split.JoinSplitAudioMetadataStage.inputs() -> tuple[list[str], list[str]]

nemo_curator.stages.audio.tagging.split.JoinSplitAudioMetadataStage.outputs() -> tuple[list[str], list[str]]

nemo_curator.stages.audio.tagging.split.JoinSplitAudioMetadataStage.process(
    task: nemo_curator.tasks.AudioTask
) -> nemo_curator.tasks.AudioTask

Process entries and join split audio metadata.

This stage collects all entries and processes meta-entries to join split audio files back together.

class nemo_curator.stages.audio.tagging.split.SplitASRAlignJoinStage(
    suggested_max_len: float = 3600.0,
    min_len: float = 1.0,
    model_name: str = 'nvidia/parakeet-tdt_ctc-1.1b',
    model_path: str | None = None,
    is_fastconformer: bool = True,
    decoder_type: str = 'rnnt',
    max_len: float = 40.0,
    batch_size: int = 100,
    transcribe_batch_size: int = 32,
    split_batch_size: int = 5000,
    num_workers: int = 10,
    infer_segment_only: bool = False,
    compute_timestamps: bool = True,
    timestamp_type: str = 'word',
    text_key: str = 'text',
    words_key: str = 'words',
    disable_word_confidence: bool = False,
    segments_key: str = 'segments',
    name: str = 'SplitASRAlignJoin'
)

Dataclass

Bases: CompositeStage[AudioTask, AudioTask]

Composite stage: Split long audio -> ASR align -> Join results.

Decomposes into three sequential stages that always run together:

SplitLongAudioStage — splits audio exceeding suggested_max_len
NeMoASRAlignerStage — transcribes and aligns each chunk
JoinSplitAudioMetadataStage — merges transcripts back into original entries

Parameters:

suggested_max_len

floatDefaults to 3600.0

Target max length for audio segments (seconds).

min_len

floatDefaults to 1.0

Minimum length for any split segment (also used by ASR).

max_len

floatDefaults to 40.0

Maximum length of audio segments for ASR processing (seconds).

model_name

strDefaults to 'nvidia/parakeet-tdt_ctc-1.1b'

Pretrained NeMo ASR model name.

model_path

str | NoneDefaults to None

Local model file path (overrides model_name if set).

is_fastconformer

boolDefaults to True

Whether the model encoder is FastConformer.

decoder_type

strDefaults to 'rnnt'

Decoder type — "ctc" or "rnnt".

batch_size

intDefaults to 100

Entries per processing chunk in ASR.

transcribe_batch_size

intDefaults to 32

Batch size passed to the ASR model’s transcribe call.

split_batch_size

intDefaults to 5000

Max entries/paths per batch when chunking segments.

num_workers

intDefaults to 10

Data-loading workers for ASR inference.

infer_segment_only

boolDefaults to False

If True, run ASR only on individual segments rather than full audio / meta-entries.

compute_timestamps

boolDefaults to True

Whether to compute word-level timestamps.

timestamp_type

strDefaults to 'word'

Timestamp granularity ("word" or "char").

text_key

strDefaults to 'text'

Output key for predicted text.

words_key

strDefaults to 'words'

Output key for word-level alignments.

disable_word_confidence

boolDefaults to False

Whether to disable word confidence scores.

segments_key

strDefaults to 'segments'

Key for the segments list in each manifest entry.

batch_size

int = 100

compute_timestamps

bool = True

decoder_type

str = 'rnnt'

disable_word_confidence

bool = False

infer_segment_only

bool = False

is_fastconformer

bool = True

max_len

float = 40.0

min_len

float = 1.0

model_name

str = 'nvidia/parakeet-tdt_ctc-1.1b'

model_path

str | None = None

name

str = 'SplitASRAlignJoin'

num_workers

int = 10

segments_key

str = 'segments'

split_batch_size

int = 5000

suggested_max_len

float = 3600.0

text_key

str = 'text'

timestamp_type

str = 'word'

transcribe_batch_size

int = 32

words_key

str = 'words'

nemo_curator.stages.audio.tagging.split.SplitASRAlignJoinStage.__post_init__() -> None

nemo_curator.stages.audio.tagging.split.SplitASRAlignJoinStage.decompose() -> list[nemo_curator.stages.base.ProcessingStage]

class nemo_curator.stages.audio.tagging.split.SplitLongAudioStage(
    suggested_max_len: float = 3600.0,
    min_len: float = 1.0,
    name: str = 'SplitLongAudio'
)

Dataclass

Bases: ProcessingStage[AudioTask, AudioTask]

Stage that splits long audio files into smaller segments.

Processes audio files that exceed a specified maximum length by splitting them at natural pauses to maintain speech coherence.

Parameters:

suggested_max_len

floatDefaults to 3600.0

Target maximum length for audio segments in seconds

min_len

floatDefaults to 1.0

Minimum length for any split segment

min_len

float = 1.0

name

str = 'SplitLongAudio'

suggested_max_len

float = 3600.0

nemo_curator.stages.audio.tagging.split.SplitLongAudioStage._build_split_metadata(
    audio_item_id: str,
    split_filepaths: list[str],
    split_durations: list[float],
    fallback: bool = False
) -> list[dict]

staticmethod

Build per-split metadata dicts from filepaths and durations.

nemo_curator.stages.audio.tagging.split.SplitLongAudioStage._do_split(
    task: nemo_curator.tasks.AudioTask
) -> nemo_curator.tasks.AudioTask

Core splitting logic, separated to keep statement count within limits.

nemo_curator.stages.audio.tagging.split.SplitLongAudioStage.get_split_points(
    metadata: dict
) -> list[float]

Get the split points for the audio file based on segments.

nemo_curator.stages.audio.tagging.split.SplitLongAudioStage.inputs() -> tuple[list[str], list[str]]

nemo_curator.stages.audio.tagging.split.SplitLongAudioStage.outputs() -> tuple[list[str], list[str]]

nemo_curator.stages.audio.tagging.split.SplitLongAudioStage.process(
    task: nemo_curator.tasks.AudioTask
) -> nemo_curator.tasks.AudioTask

Process entry to split long audio files.

Audio Splitting and Joining Stages.

Module Contents

Classes

Name	Description
`JoinSplitAudioMetadataStage`	Stage for joining metadata of previously split audio files.
`SplitASRAlignJoinStage`	Composite stage: Split long audio -> ASR align -> Join results.
`SplitLongAudioStage`	Stage that splits long audio files into smaller segments.

API

class nemo_curator.stages.audio.tagging.split.JoinSplitAudioMetadataStage(
    name: str = 'JoinSplitAudioMetadata'
)

Dataclass

Bases: ProcessingStage[AudioTask, AudioTask]

Stage for joining metadata of previously split audio files.

name

str = 'JoinSplitAudioMetadata'

nemo_curator.stages.audio.tagging.split.JoinSplitAudioMetadataStage._join_split_metadata(
    meta_entry: dict
) -> None

Join metadata from split audio files.

nemo_curator.stages.audio.tagging.split.JoinSplitAudioMetadataStage.inputs() -> tuple[list[str], list[str]]

nemo_curator.stages.audio.tagging.split.JoinSplitAudioMetadataStage.outputs() -> tuple[list[str], list[str]]

nemo_curator.stages.audio.tagging.split.JoinSplitAudioMetadataStage.process(
    task: nemo_curator.tasks.AudioTask
) -> nemo_curator.tasks.AudioTask

Process entries and join split audio metadata.

This stage collects all entries and processes meta-entries to join split audio files back together.

class nemo_curator.stages.audio.tagging.split.SplitASRAlignJoinStage(
    suggested_max_len: float = 3600.0,
    min_len: float = 1.0,
    model_name: str = 'nvidia/parakeet-tdt_ctc-1.1b',
    model_path: str | None = None,
    is_fastconformer: bool = True,
    decoder_type: str = 'rnnt',
    max_len: float = 40.0,
    batch_size: int = 100,
    transcribe_batch_size: int = 32,
    split_batch_size: int = 5000,
    num_workers: int = 10,
    infer_segment_only: bool = False,
    compute_timestamps: bool = True,
    timestamp_type: str = 'word',
    text_key: str = 'text',
    words_key: str = 'words',
    disable_word_confidence: bool = False,
    segments_key: str = 'segments',
    name: str = 'SplitASRAlignJoin'
)

Dataclass

Bases: CompositeStage[AudioTask, AudioTask]

Composite stage: Split long audio -> ASR align -> Join results.

Decomposes into three sequential stages that always run together:

SplitLongAudioStage — splits audio exceeding suggested_max_len
NeMoASRAlignerStage — transcribes and aligns each chunk
JoinSplitAudioMetadataStage — merges transcripts back into original entries

Parameters:

suggested_max_len

floatDefaults to 3600.0

Target max length for audio segments (seconds).

min_len

floatDefaults to 1.0

Minimum length for any split segment (also used by ASR).

max_len

floatDefaults to 40.0

Maximum length of audio segments for ASR processing (seconds).

model_name

strDefaults to 'nvidia/parakeet-tdt_ctc-1.1b'

Pretrained NeMo ASR model name.

model_path

str | NoneDefaults to None

Local model file path (overrides model_name if set).

is_fastconformer

boolDefaults to True

Whether the model encoder is FastConformer.

decoder_type

strDefaults to 'rnnt'

Decoder type — "ctc" or "rnnt".

batch_size

intDefaults to 100

Entries per processing chunk in ASR.

transcribe_batch_size

intDefaults to 32

Batch size passed to the ASR model’s transcribe call.

split_batch_size

intDefaults to 5000

Max entries/paths per batch when chunking segments.

num_workers

intDefaults to 10

Data-loading workers for ASR inference.

infer_segment_only

boolDefaults to False

If True, run ASR only on individual segments rather than full audio / meta-entries.

compute_timestamps

boolDefaults to True

Whether to compute word-level timestamps.

timestamp_type

strDefaults to 'word'

Timestamp granularity ("word" or "char").

text_key

strDefaults to 'text'

Output key for predicted text.

words_key

strDefaults to 'words'

Output key for word-level alignments.

disable_word_confidence

boolDefaults to False

Whether to disable word confidence scores.

segments_key

strDefaults to 'segments'

Key for the segments list in each manifest entry.

batch_size

int = 100

compute_timestamps

bool = True

decoder_type

str = 'rnnt'

disable_word_confidence

bool = False

infer_segment_only

bool = False

is_fastconformer

bool = True

max_len

float = 40.0

min_len

float = 1.0

model_name

str = 'nvidia/parakeet-tdt_ctc-1.1b'

model_path

str | None = None

name

str = 'SplitASRAlignJoin'

num_workers

int = 10

segments_key

str = 'segments'

split_batch_size

int = 5000

suggested_max_len

float = 3600.0

text_key

str = 'text'

timestamp_type

str = 'word'

transcribe_batch_size

int = 32

words_key

str = 'words'

nemo_curator.stages.audio.tagging.split.SplitASRAlignJoinStage.__post_init__() -> None

nemo_curator.stages.audio.tagging.split.SplitASRAlignJoinStage.decompose() -> list[nemo_curator.stages.base.ProcessingStage]

class nemo_curator.stages.audio.tagging.split.SplitLongAudioStage(
    suggested_max_len: float = 3600.0,
    min_len: float = 1.0,
    name: str = 'SplitLongAudio'
)

Dataclass

Bases: ProcessingStage[AudioTask, AudioTask]

Stage that splits long audio files into smaller segments.

Processes audio files that exceed a specified maximum length by splitting them at natural pauses to maintain speech coherence.

Parameters:

suggested_max_len

floatDefaults to 3600.0

Target maximum length for audio segments in seconds

min_len

floatDefaults to 1.0

Minimum length for any split segment

min_len

float = 1.0

name

str = 'SplitLongAudio'

suggested_max_len

float = 3600.0

nemo_curator.stages.audio.tagging.split.SplitLongAudioStage._build_split_metadata(
    audio_item_id: str,
    split_filepaths: list[str],
    split_durations: list[float],
    fallback: bool = False
) -> list[dict]

staticmethod

Build per-split metadata dicts from filepaths and durations.

nemo_curator.stages.audio.tagging.split.SplitLongAudioStage._do_split(
    task: nemo_curator.tasks.AudioTask
) -> nemo_curator.tasks.AudioTask

Core splitting logic, separated to keep statement count within limits.

nemo_curator.stages.audio.tagging.split.SplitLongAudioStage.get_split_points(
    metadata: dict
) -> list[float]

Get the split points for the audio file based on segments.

nemo_curator.stages.audio.tagging.split.SplitLongAudioStage.inputs() -> tuple[list[str], list[str]]

nemo_curator.stages.audio.tagging.split.SplitLongAudioStage.outputs() -> tuple[list[str], list[str]]

nemo_curator.stages.audio.tagging.split.SplitLongAudioStage.process(
    task: nemo_curator.tasks.AudioTask
) -> nemo_curator.tasks.AudioTask

Process entry to split long audio files.