chunkseg

A Python package for comprehensive evaluation of segmentation (chaptering) quality in audio and video content. Chunkseg implements three complementary evaluation protocols introduced in Beyond Transcripts: A Renewed Perspective on Audio Chaptering (arXiv:2602.08979):

1. Discretized time evaluation — Convert boundaries to fixed-size time chunks and apply established text segmentation metrics (Pk, WindowDiff, Boundary Similarity, GHD) plus binary classification metrics (F1, precision, recall).
2. Continuous time evaluation — Collar-based boundary F1 that matches predicted and reference boundaries within a time tolerance window (default ±3 s).
3. Title evaluation — BERTScore- and ROUGE-L-based comparison of chapter titles in two modes: Temporally Matched (TM) and Global Concatenation (GC).

By evaluating in the time domain rather than the text domain, chunkseg is transcript-invariant and enables comparisons across models that produce very different output formats.

Install

pip install chunkseg

For title evaluation (BERTScore + ROUGE-L):

pip install "chunkseg[titles]"

For forced alignment from structured transcripts without timestamps:

pip install "chunkseg[align]"

Evaluation Protocols

1. Discretized Time (Time-Chunks)

Segment boundaries are projected onto a sequence of fixed-size binary time chunks. A chunk is labelled 1 if a boundary falls within it, 0 otherwise. Standard binary classification metrics (precision, recall, F1, accuracy, specificity) and established segmentation metrics (Pk, WindowDiff, Boundary Similarity, GHD) are then computed over this sequence.

This approach is compatible with any model that produces boundary timestamps, regardless of transcript format.

2. Continuous Time (Collar-Based F1)

Predicted and reference boundary timestamps are compared directly in continuous time. A predicted boundary counts as a true positive if it falls within ±collar seconds of a reference boundary, using greedy closest-first 1-to-1 matching. Returns collar_precision, collar_recall, and collar_f1.

This metric is always computed alongside the time-chunk metrics whenever timestamps are available.

3. Title Evaluation

Chapter title quality is measured in two modes, both supporting BERTScore (BS) and ROUGE-L (RL):

Mode	Description
TM-BS / TM-RL	Temporally Matched — pair hyp/ref titles by start time within a tolerance window, score matched pairs only
GC-BS / GC-RL	Global Concatenation — join all titles with \n, compute a single score on the two concatenated strings
tm_matched	Fraction of reference titles that were matched (0–1)

Requires reference_titles (and optionally hyp_titles) in the input. Enable with --titles in the CLI or titles=True in evaluate_batch().

Usage

Python API

from chunkseg import evaluate, evaluate_batch, print_results

# Timestamps mode — no transcript or audio needed
result = evaluate(
    hypothesis=[120.5, 300.0],
    reference=[125.0, 310.0],
    duration=600.0,
    chunk_size=6.0,
    collar=3.0,
)

# Batch evaluation with aggregated metrics and bootstrap CIs
results = evaluate_batch(
    samples=[
        {"hypothesis": [120.5], "reference": [125.0], "duration": 600.0},
        {"hypothesis": [300.0], "reference": [310.0], "duration": 500.0},
    ],
    chunk_size=6.0,
    collar=3.0,
)
print_results(results)

# Title evaluation (requires chunkseg[titles])
result = evaluate(
    hypothesis=[24.2, 33.94],
    reference=[11.0, 23.0, 34.0],
    duration=50.0,
    hyp_titles=[("Set a background", 24.2), ("Clip the background", 33.94)],
    reference_titles=[("Wrap text with a span", 11.0),
                      ("Add a background", 23.0),
                      ("Clip background to text", 34.0)],
    tolerance=5.0,
)

# Standalone title scoring
from chunkseg import compute_title_scores

scores = compute_title_scores(
    hyp_titles=[("Introduction", 0.0), ("Methods", 60.0)],
    ref_titles=[("Intro", 0.0), ("Methodology", 58.0)],
    tolerance=5.0,
)
# Returns: tm_bs_f1, tm_rl_f1, gc_bs_f1, gc_rl_f1, tm_matched, ...

CLI

# Basic timestamps mode
chunkseg samples.jsonl

# With title evaluation and custom collar
chunkseg samples.jsonl --titles --tolerance 5.0 --collar 3.0

# With WER (requires reference_transcript field)
chunkseg samples.jsonl --wer

# Structured transcript mode (forced alignment)
chunkseg transcripts.jsonl --format cstart --lang eng

# Structured transcript with embedded timestamps
chunkseg transcripts.jsonl --format cstart_ts

# Save results to JSON
chunkseg samples.jsonl --titles --output results.json

Input Format

Each line of the input JSONL file must be a JSON object. The required fields depend on the evaluation mode.

Timestamps mode (minimal)

{"hypothesis": [24.2, 33.94], "reference": [11.0, 23.0, 34.0], "duration": 50.0}

With title evaluation

{
  "hypothesis": [24.2, 33.94],
  "reference": [11.0, 23.0, 34.0],
  "duration": 50.0,
  "reference_titles": [["Wrap text with a span", 11.0],
                       ["Add a background", 23.0],
                       ["Clip background to text", 34.0]],
  "hyp_titles": [["Set a background", 24.2],
                 ["Clip the background", 33.94]]
}

Structured transcript (forced alignment)

{
  "hypothesis": "[CSTART] Intro [CEND] text... [CSTART] Main [CEND] more...",
  "reference": [125.0],
  "audio": "/path/to/audio.wav",
  "duration": 600.0
}

All fields

Field	Type	Required	Description
hypothesis	list[float] or str	Yes	Predicted boundaries (seconds) or structured transcript
reference	list[float]	Yes	Ground-truth boundary timestamps (seconds)
duration	float	Yes	Total duration (seconds)
audio	str	For alignment	Path to audio file
reference_titles	[[title, seconds], ...]	For --titles	Ground-truth chapter titles with start times
hyp_titles	[[title, seconds], ...]	Optional	Predicted titles (inferred from transcript if omitted)
reference_transcript	str	For --wer	Reference transcript text

Input Modes

Mode	hypothesis type	Audio?	Notes
Timestamps	list[float]	No	Direct boundary seconds
Transcript (alignment)	str	Yes	format="cstart" etc., requires chunkseg[align]
Transcript + timestamps (use provided)	str	No	format="cstart_ts" etc.
Transcript + timestamps (force alignment)	str	Yes	format="cstart_ts", force_alignment=True

Parser Presets

Preset	Format	Timestamps?
cstart	[CSTART] Title [CEND] text...	No
cstart_ts	[CSTART] 1:23:45 - Title [CEND] text...	Yes
newline	Sections separated by blank lines	No
markdown	# Title\nText\n## Subtitle...	No
markdown_ts	# 0:15:30 - Introduction\nText...	Yes
custom	User-provided regex	No
custom_ts	User-provided regex with (?P<timestamp>...)	Yes

Metrics

Discretized Time (Time-Chunks)

Metric	Description
f1	Harmonic mean of precision and recall (derived from aggregated P/R)
precision	TP / (TP + FP)
recall	TP / (TP + FN)
accuracy	(TP + TN) / Total
specificity	TN / (TN + FP)
pk	Beeferman's Pk (lower is better)
window_diff	WindowDiff (lower is better)
boundary_similarity	Boundary similarity score
ghd	Generalized Hamming Distance

Continuous Time (Collar-Based)

Metric	Description
collar_f1	F1 within ±collar seconds (default ±3 s)
collar_precision	Precision within collar
collar_recall	Recall within collar

Title Evaluation

Metric	Description
tm_bs_f1	BERTScore F1 on temporally matched title pairs
tm_bs_precision	BERTScore precision on matched pairs
tm_bs_recall	BERTScore recall on matched pairs
tm_rl_f1	ROUGE-L F1 on temporally matched title pairs
tm_rl_precision	ROUGE-L precision on matched pairs
tm_rl_recall	ROUGE-L recall on matched pairs
tm_matched	Fraction of reference titles matched (0–1)
gc_bs_f1	BERTScore F1 on globally concatenated titles
gc_bs_precision	BERTScore precision on concatenated titles
gc_bs_recall	BERTScore recall on concatenated titles
gc_rl_f1	ROUGE-L F1 on globally concatenated titles
gc_rl_precision	ROUGE-L precision on concatenated titles
gc_rl_recall	ROUGE-L recall on concatenated titles

All scalar metrics are reported as {mean, std, ci_lower, ci_upper} with bootstrap confidence intervals (default 100 iterations, configurable via --num-bootstrap).

CLI Reference

chunkseg <input.jsonl> [options]

Options:
  --chunk-size FLOAT    Chunk size in seconds (default: 6.0)
  --collar FLOAT        Collar size for boundary F1 (default: 3.0)
  --format STR          Parser preset for transcript mode
  --custom-pattern STR  Regex for custom/custom_ts format
  --timestamp-format STR Timestamp format for custom_ts
  --lang STR            ISO 639-3 language code for alignment (default: eng)
  --force-alignment     Derive timestamps from audio alignment
  --wer                 Compute WER (requires reference_transcript field)
  --titles              Compute title metrics (requires reference_titles field)
  --tolerance FLOAT     Time tolerance for TM matching in seconds (default: 5.0)
  --output FILE         Write results to JSON file
  --num-bootstrap INT   Bootstrap iterations for CIs (default: 100)

Dependencies

Required:

• numpy
• segeval
• nltk
• jiwer

Optional — title evaluation (chunkseg[titles]):

• bert-score
• rouge-score

Optional — transcript alignment (chunkseg[align]):

• torch >= 2.1.0
• torchaudio >= 2.1.0

Citation

If you use chunkseg in your research, please cite:

@article{retkowski2026beyond,
  title     = {Beyond Transcripts: A Renewed Perspective on Audio Chaptering},
  author    = {Retkowski, Fabian and Z{\"u}fle, Maike and Nguyen, Thai Binh and Niehues, Jan and Waibel, Alexander},
  journal   = {arXiv preprint arXiv:2602.08979},
  year      = {2026},
  url       = {https://arxiv.org/abs/2602.08979}
}

License

This project is licensed under the MIT License - see the LICENSE file for details.