Skip to content

CURATOR_BRANCH_COMPARISON.md

Latest commit

 

History

History
565 lines (437 loc) · 26.1 KB

CURATOR_BRANCH_COMPARISON.md

File metadata and controls

565 lines (437 loc) · 26.1 KB

Curator Branch Comparison

Curator branch under review: Granary v2 audio standardization branch Reference branch: reference integration branch Common ancestor: 1cbb4a62 (28 commits in ref since, 15 in ours since)

Diff summary: 72 files changed, +5668 / -2124 lines

Table of Contents

  1. High-Level Philosophy Differences
  2. Features Only in Reference (Removed by Us)
  3. Features Only in Ours (Added by Us)
  4. Shared Features with Different Implementations
  5. Naming & Convention Differences
  6. Dependency Differences
  7. Test Coverage Differences
  8. Entry Point & Pipeline Assembly Differences
  9. Recommended Merge Strategy

1. High-Level Philosophy Differences

Aspect Reference integration branch Curator branch under review
Decision tracking additional_notes dict with per-stage entries via pipeline_utils.set_note() Removed entirely — stages set _skip_me but don't log reasons to a dict
Skip key name _skipme (NeMo/lhotse convention per latest PR #8) _skip_me (Python naming convention)
Multi-node scaling num_workers parameter on GPU stages + xenna_stage_spec() override Removed num_workers and xenna_stage_spec() — lets autoscaler decide
SED (Sound Event Detection) Full CNN14/PANNs inference + postprocessing pipeline Completely removed — SED is out of scope for Granary v2
Segment extraction Has segment_extractor.py in segmentation/ Has SegmentExtractionStage in io/ (605 lines, much more comprehensive)
Module organization Some I/O stages in alm/ subpackage Moved I/O stages to dedicated io/ subpackage
Sample rate key sampling_rate (NeMo convention) sample_rate (simpler, more standard)
Language support LANG_CODE_TO_NAME in shared pipeline_utils.py Inlined into qwen_omni.py, imported from there by qwen_asr.py
Lazy imports Segmentation uses __getattr__ pattern Direct imports (simpler but adds hard deps)
PnC per-language source_lang_key on PnC stage + {language} placeholder in prompts Removed — single prompt template without {language} placeholder
Hydra entry point Not present (only run_pipeline.py argparse) tutorials/audio/qwen_omni_inprocess/main.py (Hydra config-driven)

2. Features Only in Reference (Removed by Us)

2.1 Sound Event Detection (SED)

Files in reference only:

  • nemo_curator/stages/audio/inference/sed.py (274 lines)
  • nemo_curator/stages/audio/inference/sed_models/__init__.py
  • nemo_curator/stages/audio/inference/sed_models/cnn14.py (352 lines)
  • nemo_curator/stages/audio/postprocessing/sed_postprocessing.py (160 lines)
  • nemo_curator/stages/audio/postprocessing/sed_utils.py (275 lines)

What it does:

  • SEDInferenceStage — runs PANNs CNN14 model on each audio waveform, produces per-frame (T×527) AudioSet class probabilities
  • SEDPostprocessingStage — applies thresholds, minimum duration, merge gaps to produce event labels (speech, music, noise, etc.)
  • Supports saving results as .npz sidecar files

Why we removed it:

  • SED is not part of the Granary v2 transcription pipeline
  • Adds a large dependency (torchlibrosa) and GPU memory requirement
  • The run_pipeline.py in ref conditionally adds SED stages if --sed_checkpoint is provided; we removed all SED CLI args

2.2 pipeline_utils.py — Decision Tracking System

File in reference only: nemo_curator/stages/audio/pipeline_utils.py (53 lines)

What it provides:

def set_note(task_data: dict, stage_name: str, value: str, notes_key: str = "additional_notes") -> None:
    """Write a stage decision note into the additional_notes dict."""
    notes = task_data.get(notes_key)
    if not isinstance(notes, dict):
        notes = {}
        task_data[notes_key] = notes
    notes[stage_name] = value

Plus LANG_CODE_TO_NAME dict (language codes → English names).

How reference uses it: Every stage (WhisperHallucination, PnC, ITN, InitializeFields) calls set_note() to record WHY a decision was made:

  • "skipped (flagged)", "skipped (empty)", "applied (modified)", "applied (unchanged)", "hallucination (unique_words, long_word)", "fallback (length_mismatch)", etc.

Why we removed it:

  • Adds per-sample overhead (dict creation + string assignment for every task)
  • Output JSONL already has _skip_me field which indicates skipped status
  • For debugging, log messages provide the same info without polluting data
  • Keeping output manifests clean (fewer keys = smaller JSONL)

2.3 num_workers + xenna_stage_spec() — Manual Worker Scaling

In reference, every GPU stage has:

num_workers: int | None = None

def xenna_stage_spec(self) -> dict[str, Any]:
    spec: dict[str, Any] = {}
    if self.num_workers is not None:
        spec["num_workers"] = self.num_workers
    return spec

Plus CLI args: --omni_num_workers, --asr_num_workers, --pnc_num_workers, --itn_num_workers

In ours: Completely removed. The Xenna autoscaler handles worker count based on available GPUs and resource requests.

Why we removed it:

  • Manual worker counts are fragile (wrong value = underutilization or OOM)
  • Autoscaler adapts to different hardware (L40S vs A100 vs H100)
  • Removes 4 CLI args and 4 repeated method definitions

2.4 autoscale_interval_s — Scaling Interval Parameter

In reference: --autoscale_interval_s arg (default 180s) passed to XennaExecutor config.

In ours: Removed — uses Xenna default.

2.5 pnc_source_lang_key — Per-Sample Language for PnC

In reference: PnC stage accepts source_lang_key and passes languages to QwenTextLLM.generate(). The prompt template uses {language} placeholder for per-sample language resolution.

In ours: Restored. PnC still accepts source_lang_key, passes per-sample languages to QwenTextLLM.generate(), and the default PnC prompt now includes {language} in the same places as the reference branch.

Impact: Current code now supports multilingual PnC where the prompt says e.g. "Add punctuation to this Italian text: {text}", while keeping the local _skip_me convention and worker lifecycle improvements.

2.6 Granary v1 _skipme Preservation

In reference (PR #8):

def _init_task(self, task: AudioTask) -> None:
    v1_skipme = task.data.get(self.skip_me_key, "")
    notes = task.data.get(self.notes_key, {})
    if v1_skipme:
        notes["v1_skipme"] = v1_skipme
    task.data[self.notes_key] = notes

InitializeFields preserves any existing _skipme value into additional_notes["v1_skipme"] before resetting.

In ours: Just sets _skip_me = "" unconditionally. No preservation.

2.7 Extra drop_keys in InitializeFields

In reference: _DEFAULT_DROP_KEYS includes "selected_transcript", "taskname", "orig_text", "orig_answer" (4 extra keys).

In ours: Only drops "pnc", "itn", "timestamp".

2.8 segment_extractor.py (in segmentation/)

In reference: nemo_curator/stages/audio/segmentation/segment_extractor.py (86 lines) — simpler segment extraction.

In ours: Removed from segmentation, replaced by comprehensive io/extract_segments.py (605 lines).

3. Features Only in Ours (Added by Us)

3.1 Hydra Entry Point (tutorials/audio/qwen_omni_inprocess/main.py)

321 lines. A Hydra-based entry point that wraps the same pipeline as run_pipeline.py but accepts configuration via YAML + overrides instead of argparse.

Key features:

  • build_granary_v2_pipeline(cfg: DictConfig) — constructs the full stage chain from config
  • _prefetch_models() — parallel HuggingFace model download with ThreadPoolExecutor
  • Hydra config at tutorials/audio/qwen_omni_inprocess/qwen_omni_inprocess.yaml (118 lines)

Why added: NvLLMOps Kratos run_curator.py invokes the pipeline via:

subprocess.run([python, "tutorials/audio/{pipeline}/main.py",
    "--config-path=...", "--config-name=...",
    "workspace_dir=...", "input_manifest=..."])

Hydra overrides are cleaner than building long argparse command lines.

3.2 SegmentExtractionStage (io/extract_segments.py)

605 lines. Comprehensive segment extraction that auto-detects pipeline combo:

  • Combo 2: No VAD / VAD only — extracts by original_start_ms/original_end_ms
  • Combo 3: Speaker diarization — extracts per-speaker intervals from diar_segments
  • Combo 4: VAD + speaker — extracts speaker-segment timestamps

Supports: WAV/FLAC output, parallel processing, manifest rewriting with new paths.

3.3 Speaker Diarization Stage (inference/speaker_diarization/pyannote.py)

304 lines. GPU-accelerated speaker diarization using Pyannote:

  • Wraps pyannote/speaker-diarization-3.1 model
  • Produces per-sample diar_segments list with speaker labels and timestamps
  • Configurable min/max speakers, min segment duration

3.4 WhisperX VAD Stage (inference/vad/whisperx_vad.py)

188 lines. Voice Activity Detection using WhisperX's silero-based VAD:

  • Produces VAD segments with start/end timestamps
  • Configurable threshold, min speech duration, min silence duration

3.5 NeMo ASR Stage (inference/asr/asr_nemo.py)

130 lines. NeMo framework ASR inference stage:

  • Wraps pretrained NeMo ASR models (Conformer, etc.)
  • GPU-accelerated batch inference
  • Alternative to Qwen-based ASR for certain pipelines

3.6 ALMManifestReader / ALMManifestWriter (moved to io/)

Moved from alm/ to io/ subpackage for better organization. The io/ package now contains all I/O stages:

  • alm_manifest_reader.py (105 lines)
  • alm_manifest_writer.py (84 lines)
  • nemo_tarred_reader.py
  • sharded_manifest_writer.py (118 lines)
  • extract_segments.py (605 lines)

3.7 ManifestReader / ManifestReaderStage / ManifestWriterStage (in common.py)

Generic manifest I/O stages added to common.py:

  • ManifestReaderStage — reads JSONL files line-by-line via fsspec
  • ManifestReader — CompositeStage wrapping FilePartitioning + reading
  • ManifestWriterStage — appends AudioTasks to JSONL output

3.8 ShardedManifestWriterStage (moved to io/)

118 lines. Was in alm/sharded_manifest_writer.py, moved to io/sharded_manifest_writer.py.

3.9 PIPELINE_DEEP_DIVE.md

1161 lines. Comprehensive documentation of the entire Granary v2 pipeline:

  • Stage-by-stage breakdown with input/output schemas
  • GPU resource requirements per stage
  • Data flow diagrams
  • Configuration reference

3.10 Comprehensive Test Suite

New test files (all tests/stages/audio/):

  • inference/test_qwen_asr.py (121 lines)
  • inference/test_qwen_omni.py (131 lines)
  • io/test_extract_segments.py (157 lines)
  • io/test_nemo_tarred_reader.py (139 lines)
  • io/test_sharded_manifest_writer.py (97 lines)
  • text_filtering/test_abbreviation_concat.py (102 lines)
  • text_filtering/test_disfluency_wer_guard.py (100 lines)
  • text_filtering/test_itn_restoration.py (135 lines)
  • text_filtering/test_pnc_content_guard.py (115 lines)
  • text_filtering/test_pnc_restoration.py (110 lines)
  • text_filtering/test_select_best_prediction.py (106 lines)

3.12 granary_v2_postprocessing/README.md and requirements.txt

Added documentation and dependency listing for the tutorial.

4. Shared Features with Different Implementations

4.1 InferenceQwenOmniStage

Aspect Reference Ours
sample_rate_key "sampling_rate" "sample_rate"
skip_me_key field Present (sets "empty_audio" on empty waveforms) Removed — doesn't flag empty audio
num_workers Present Removed
xenna_stage_spec() Returns {"num_workers": N} Removed
setup_on_node() Creates full model + calls setup() Only does snapshot_download() (weight caching)
process_batch() input validation No explicit validation Calls self.validate_input(task)
Empty audio handling Returns skipped_indices, sets _skipme="empty_audio" Model returns empty string, no flagging
generate() return (pred_texts, disfluency_texts, skipped_indices) (pred_texts, disfluency_texts)

4.2 InferenceQwenASRStage

Aspect Reference Ours
sample_rate_key "sampling_rate" "sample_rate"
gpu_memory_utilization default 0.7 0.95
num_workers Present Removed
setup_on_node() Creates full model Only snapshot_download()
Run-index selection Inline in process_batch Extracted to _select_run_indices() method
Language resolution Inline in process_batch Extracted to _resolve_languages() method
Input validation None self.validate_input(task) per task

4.3 PnCRestorationStage

Aspect Reference Ours
skip_me_key "_skipme" "_skip_me"
notes_key Present (writes decision notes) Removed
source_lang_key Present (passes languages to model) Removed
num_workers Present Removed
xenna_stage_spec() Present Removed
Batch processing Calls self._model.generate(eligible_texts, languages=eligible_langs) once Chunks into self.batch_size sub-batches with for-loop
Decision tracking set_note() for "applied (modified)", "applied (unchanged)", "kept as-is (incomplete)" None
Input validation None self.validate_input(task) per task
_partition_tasks() Inline logic Extracted to dedicated method

4.4 ITNRestorationStage

Aspect Reference Ours
skip_me_key "_skipme" "_skip_me"
notes_key Present Removed
num_workers Present Removed
max_output_tokens default 512 4096
max_model_len default 2048 4096
max_num_seqs default 64 16
process() Delegates to process_batch Raises NotImplementedError
Batch logic Inline in process_batch Extracted to _collect_prompts() and _apply_output() methods
Decision tracking set_note() calls None
Input validation None self.validate_input(task) per task

4.5 WhisperHallucinationStage

Aspect Reference Ours
skip_me_key "_skipme" "_skip_me"
name field position After other fields First field (Python dataclass ordering)
Decision tracking set_note() for "hallucination (...)", "recovered", etc. None
Phrase file parsing Reads lines as-is ({line.strip() for line in f}) Parses trailing integers (removes count suffixes from phrase lists)

4.6 InitializeFieldsStage

Aspect Reference Ours
Skip key "_skipme" "_skip_me"
Notes key Present, preserves v1 _skipme into notes Removed
Extra drop keys "selected_transcript", "taskname", "orig_text", "orig_answer" Not present
Internal method _init_task() called by both process() and process_batch() Logic duplicated inline in both methods
name field position After other fields First field

4.7 NemoTarShardReaderStage

Aspect Reference Ours
Sample rate key output Both "sampling_rate" and "sample_rate" Only "sample_rate"
outputs() Includes "sampling_rate" Only "sample_rate", "corpus", "num_channels"
Tar file handling No try/finally (tar.close() at end) try/finally ensures tar.close() even on exception
RayStageSpecKeys import Try/except ModuleNotFoundError fallback Direct import from nemo_curator.backends.utils

4.8 QwenTextLLM (model wrapper)

Aspect Reference Ours
_prepare_single() Accepts language parameter, formats {text} + {language} Same
_prepare_batch() Accepts `languages: list[str] None`
generate() Accepts `languages: list[str] None`
Batch mapping pool.map(_prepare_single, texts, templates, langs) Same, plus output-token metrics

4.9 QwenOmni (model wrapper, models/qwen_omni.py)

Aspect Reference Ours
generate() return (pred_texts, disfluency_texts, skipped_indices) (pred_texts, disfluency_texts)
Empty audio handling Tracks skipped_indices set, returns it Returns empty string for empty audio, no tracking

4.10 QwenASR (model wrapper, models/qwen_asr.py)

Aspect Reference Ours
Main difference Minor - mostly same Same core logic, slightly different error handling

5. Naming & Convention Differences

Convention Reference Ours
Skip key _skipme _skip_me
Sample rate key sampling_rate sample_rate
Dataclass field ordering name after functional fields name as first field
Decision tracking additional_notes dict per task Not tracked
Language lookup location pipeline_utils.py (shared module) Inlined in qwen_omni.py
Segmentation imports Lazy via __getattr__ Direct imports
process() on batch-only stages Some delegate to process_batch All raise NotImplementedError
Input validation Not enforced validate_input() check at start of process_batch()

6. Dependency Differences

pyproject.toml — audio_cuda12 extra

Package Reference Ours
vllm Not in extras (runtime install) Added to audio_cuda12
qwen-omni-utils Not in extras Added to audio_cuda12
qwen-asr Not in extras Added to audio_cuda12
fasttext Not in extras fasttext==0.9.3 added to audio_cuda12
transformers constraint >=4.56.0,<5.0 (comment: "breaks with hf_hub<1.0") >=4.56.0,<5.0 (comment: "upper-bounded by nemo-toolkit[asr] constraint")

Impact: Our branch tries to make uv sync --extra audio_cuda12 install everything needed for the Granary v2 pipeline. Reference expects runtime pip install commands (as in the Slurm sbatch script).

7. Test Coverage Differences

Test File Reference Ours
test_qwen_asr.py Not present 121 lines
test_qwen_omni.py Not present 131 lines
test_extract_segments.py Not present 157 lines
test_nemo_tarred_reader.py Not present 139 lines
test_sharded_manifest_writer.py Not present 97 lines
test_abbreviation_concat.py Not present 102 lines
test_disfluency_wer_guard.py Not present 100 lines
test_itn_restoration.py Not present 135 lines
test_pnc_content_guard.py Not present 115 lines
test_pnc_restoration.py Not present 110 lines
test_select_best_prediction.py Not present 106 lines
test_initialize_fields.py Present (modified) Modified to match _skip_me naming
test_fasttext_lid.py Present Modified (75 lines of changes)
test_regex_substitution.py Present Modified (91 lines of changes)
test_whisper_hallucination.py Present Modified (52 lines of changes)
test_common.py Present Modified (10 lines of changes)

8. Entry Point & Pipeline Assembly Differences

Reference: examples/audio/qwen_omni_inprocess/run_pipeline.py

  • Pure argparse-based CLI
  • Optionally includes SED stages (if --sed_checkpoint given)
  • InferenceQwenOmniStage is appended after optional SED block
  • Has --autoscale_interval_s and --*_num_workers args
  • Has --pnc_source_lang_key arg
  • itn_max_output_tokens default: 512
  • XennaExecutor config includes autoscale_interval_s

Ours: examples/audio/qwen_omni_inprocess/run_pipeline.py

  • Same argparse CLI but:
    • No SED args or stages
    • No num_workers args
    • No autoscale_interval_s
    • No pnc_source_lang_key
    • itn_max_output_tokens default: 4096
    • InferenceQwenOmniStage is part of the initial stages list (not appended after conditional block)

Ours (additional): tutorials/audio/qwen_omni_inprocess/main.py

  • Hydra-based entry point (does NOT exist in reference)
  • Uses build_granary_v2_pipeline(cfg) function
  • _prefetch_models() for parallel model downloads
  • Designed for NvLLMOps Kratos integration

9. Recommended Merge Strategy

Must bring from reference into ours:

  1. _skipme → _skipme — Reference's PR #8 standardized to _skipme (NeMo/lhotse convention). We should adopt this and drop _skip_me.

  2. sampling_rate key — NeMo convention is sampling_rate. Since NemoTarredAudioReader feeds into stages that then pass data downstream, we should match NeMo's convention to avoid key mismatches with existing NeMo manifests.

  3. additional_notes tracking — Valuable for debugging production runs. Consider keeping it as optional (controlled by a flag) rather than hardcoded.

  4. Extra drop_keys — "selected_transcript", "taskname", "orig_text", "orig_answer" are common in NeMo manifests and should be dropped.

  5. Granary v1 _skipme preservation — Important for incremental processing where data may already have v1 flags.

Keep from ours (do NOT revert):

  1. Hydra entry point (main.py) — Required for NvLLMOps Kratos integration
  2. SegmentExtractionStage in io/ — More comprehensive than reference's simpler version
  3. Speaker diarization + WhisperX VAD stages — New capability not in reference
  4. NeMo ASR stage — Additional inference backend
  5. Comprehensive test suite — 11 new test files
  6. PIPELINE_DEEP_DIVE.md — Documentation
  7. validate_input() calls — Better error messages
  8. Extracted helper methods (_partition_tasks, _collect_prompts, _apply_output, _select_run_indices, _resolve_languages) — Better code organization
  9. try/finally for tar handling — Prevents resource leaks
  10. Dependencies in pyproject.toml — Makes uv sync work without runtime pip

Items to discuss:

  1. SED stages — Do we want them? They're useful for audio quality assessment but add complexity and dependencies.
  2. num_workers / manual scaling — Do we need manual override for multi-node? Autoscaler may not always be optimal for bursty GPU workloads.
  3. Per-language PnC — Resolved in current code. The default PnC prompt includes {language} and the stage passes per-sample language names into QwenTextLLM.generate().
  4. phrases.txt format — Reference added a phrases.txt (65 lines) with trailing integers (counts?). Our parser handles this via rsplit; reference reads raw lines. Need to verify format compatibility.

Appendix: File-by-File Quick Reference

Files only in reference (we deleted):

nemo_curator/stages/audio/inference/sed.py
nemo_curator/stages/audio/inference/sed_models/__init__.py
nemo_curator/stages/audio/inference/sed_models/cnn14.py
nemo_curator/stages/audio/postprocessing/sed_postprocessing.py
nemo_curator/stages/audio/postprocessing/sed_utils.py
nemo_curator/stages/audio/pipeline_utils.py
nemo_curator/stages/audio/segmentation/segment_extractor.py
tutorials/audio/granary_v2_postprocessing/phrases.txt

Files only in ours (we added):

nemo_curator/stages/audio/inference/asr/__init__.py
nemo_curator/stages/audio/inference/asr/asr_nemo.py
nemo_curator/stages/audio/inference/speaker_diarization/__init__.py
nemo_curator/stages/audio/inference/speaker_diarization/pyannote.py
nemo_curator/stages/audio/inference/vad/__init__.py
nemo_curator/stages/audio/inference/vad/whisperx_vad.py
nemo_curator/stages/audio/io/alm_manifest_reader.py
nemo_curator/stages/audio/io/alm_manifest_writer.py
nemo_curator/stages/audio/io/extract_segments.py
nemo_curator/stages/audio/io/sharded_manifest_writer.py
tests/stages/audio/inference/test_qwen_asr.py
tests/stages/audio/inference/test_qwen_omni.py
tests/stages/audio/io/test_extract_segments.py
tests/stages/audio/io/test_nemo_tarred_reader.py
tests/stages/audio/io/test_sharded_manifest_writer.py
tests/stages/audio/text_filtering/test_abbreviation_concat.py
tests/stages/audio/text_filtering/test_disfluency_wer_guard.py
tests/stages/audio/text_filtering/test_itn_restoration.py
tests/stages/audio/text_filtering/test_pnc_content_guard.py
tests/stages/audio/text_filtering/test_pnc_restoration.py
tests/stages/audio/text_filtering/test_select_best_prediction.py
tutorials/audio/granary_v2_postprocessing/README.md
tutorials/audio/granary_v2_postprocessing/requirements.txt
tutorials/audio/qwen_omni_inprocess/PIPELINE_DEEP_DIVE.md
tutorials/audio/qwen_omni_inprocess/main.py
tutorials/audio/qwen_omni_inprocess/qwen_omni_inprocess.yaml

Files modified in both (diverged):

examples/audio/qwen_omni_inprocess/run_pipeline.py
nemo_curator/models/qwen_asr.py
nemo_curator/models/qwen_omni.py
nemo_curator/models/qwen_text_llm.py
nemo_curator/stages/audio/__init__.py
nemo_curator/stages/audio/alm/alm_manifest_reader.py
nemo_curator/stages/audio/alm/alm_manifest_writer.py
nemo_curator/stages/audio/alm/sharded_manifest_writer.py
nemo_curator/stages/audio/common.py
nemo_curator/stages/audio/inference/__init__.py
nemo_curator/stages/audio/inference/qwen_asr.py
nemo_curator/stages/audio/inference/qwen_omni.py
nemo_curator/stages/audio/io/nemo_tarred_reader.py
nemo_curator/stages/audio/metrics/get_wer.py
nemo_curator/stages/audio/segmentation/__init__.py
nemo_curator/stages/audio/segmentation/speaker_separation.py
nemo_curator/stages/audio/segmentation/vad_segmentation.py
nemo_curator/stages/audio/text_filtering/abbreviation_concat.py
nemo_curator/stages/audio/text_filtering/disfluency_wer_guard.py
nemo_curator/stages/audio/text_filtering/fasttext_lid.py
nemo_curator/stages/audio/text_filtering/finalize_fields.py
nemo_curator/stages/audio/text_filtering/initialize_fields.py
nemo_curator/stages/audio/text_filtering/itn_restoration.py
nemo_curator/stages/audio/text_filtering/pnc_content_guard.py
nemo_curator/stages/audio/text_filtering/pnc_restoration.py
nemo_curator/stages/audio/text_filtering/prompts/pnc_prompt.md
nemo_curator/stages/audio/text_filtering/regex_substitution.py
nemo_curator/stages/audio/text_filtering/select_best_prediction.py
nemo_curator/stages/audio/text_filtering/whisper_hallucination.py
pyproject.toml
tutorials/audio/README.md
tutorials/audio/granary_v2_postprocessing/post_processsing_pipeline.py