findsylls

Language-agnostic toolkit for unsupervised syllable-level speech segmentation, embedding extraction, and evaluation.

findsylls provides a full pipeline from raw audio to clustered syllable embeddings:

• Envelope computation — RMS, Hilbert, low-pass, SBS, theta, gammatone, periodicity, and neural pseudo-envelopes
• Syllable segmentation — classical peak detection, convex-hull and threshold segmentation, and neural end-to-end methods (Sylber, VG-HuBERT)
• Feature extraction — MFCC, mel spectrogram, HuBERT, Sylber, VG-HuBERT
• Syllable embedding — pooled per-syllable vectors for downstream tasks
• Unsupervised discovery — k-means, mini-batch k-means, agglomerative clustering
• Evaluation — F1 against TextGrid annotations at phone, syllable, and word granularity
• Visualization — waveform, envelope, segmentation, and feature-matrix plots

---

Install

pip install findsylls                  # core (classical methods)
pip install 'findsylls[embedding]'     # neural feature extraction (HuBERT, VG-HuBERT)
pip install 'findsylls[end2end]'       # neural segmenters (Sylber, VG-HuBERT)
pip install 'findsylls[viz]'           # plotting extras
pip install 'findsylls[storage]'       # HDF5 corpus storage
pip install 'findsylls[all]'           # everything

---

Quick Start

1 — Segment audio into syllables

from findsylls import segment_audio

# Classical: peak detection on an SBS amplitude envelope
syllables, envelope, times = segment_audio(
    "audio.wav",
    method="peakdetect",
    segmentation_kwargs={"envelope_method": "sbs"},
    return_envelope=True,
)

print(f"Found {len(syllables)} syllables")
# syllables: [(start_s, nucleus_s, end_s), ...]

Waveform (gray), SBS amplitude envelope (blue), syllable boundaries (green), and detected nuclei (red dots) for a sample utterance.

---

Module Guide

Envelope (findsylls.envelope)

The envelope module converts a raw audio waveform into a 1-D amplitude signal. All computers implement EnvelopeComputer.compute(audio, sr) → (envelope, times).

from findsylls.audio.utils import load_audio
from findsylls.envelope import (
    RMSEnvelope, HilbertEnvelope, ThetaEnvelope, SBSEnvelope,
    LowpassEnvelope, CLSAttentionEnvelope, GreedyCosineEnvelope,
)
from findsylls.plotting import plot_multiple_envelopes

audio, sr = load_audio("audio.wav")

envelopes = {}
for name, computer in [
    ("RMS",     RMSEnvelope()),
    ("Hilbert", HilbertEnvelope()),
    ("Theta",   ThetaEnvelope()),
    ("SBS",     SBSEnvelope()),
]:
    env, times = computer.compute(audio, sr)
    envelopes[name] = (env, times)

fig = plot_multiple_envelopes(audio, sr, envelopes)

Four classical envelope methods on the same utterance. SBS and Theta track syllabic rhythm most closely; Hilbert and RMS give a more continuous energy contour.

You can also call the functional dispatch directly:

from findsylls import get_amplitude_envelope

envelope, times = get_amplitude_envelope(audio, sr, method="theta")

Available envelope methods: rms, hilbert, lowpass, sbs, theta, gammatone, periodicity, cls_attention, greedy_cosine, mincut

---

Segmentation (findsylls.segmentation)

All segmenters return List[(start_s, nucleus_s, end_s)].

Classical — peak detection

from findsylls import segment_audio
from findsylls.plotting import plot_multiple_envelope_segmentations
from findsylls.audio.utils import load_audio
from findsylls.envelope import HilbertEnvelope, ThetaEnvelope, SBSEnvelope
from findsylls.segmentation import get_segmenter

audio, sr = load_audio("audio.wav")

results = {}
for name, env_method in [("Hilbert", "hilbert"), ("Theta", "theta"), ("SBS", "sbs")]:
    env_computer = {"hilbert": HilbertEnvelope, "theta": ThetaEnvelope, "sbs": SBSEnvelope}[env_method]()
    env, times = env_computer.compute(audio, sr)
    segmenter = get_segmenter("peakdetect", envelope_method=env_method)
    segments = segmenter.segment(audio=audio, sr=sr)
    results[name] = (env, times, segments)

fig = plot_multiple_envelope_segmentations(audio, sr, results)

The same audio segmented by peakdetect using three different envelope methods. Each panel shows how the chosen envelope shape influences where boundaries fall.

Preset segmenters (paper-replication configurations)

Preset classes replicate the exact configurations from published papers. Each carries a REFERENCE attribute and a cite() method — see Preset Citations below.

Preset	Paper	Approach
SBSPeakdetectSegmenter	my personal config	spectral band subtraction envelope + Billauer peak detection
ThetaOscillatorSegmenter	Räsänen et al. 2018	gammatone filterbank → damped-oscillator sonority
EnergyPeriodicitySegmenter	Xie & Niyogi 2006	periodicity-gated regions + energy convex-hull nucleus picking
RhythmGuidedSegmenter	Zhang & Glass 2009	rhythm-guided peak recovery of closely-spaced nuclei
SylberSegmenter	Cho et al. 2025	greedy cosine merging on Sylber HuBERT features
VGHubertMinCutSegmenter	Peng et al. 2023	self-similarity MinCut on VG-HuBERT features
VGHubertCLSSegmenter	Peng & Harwath 2022	CLS-attention thresholding on VG-HuBERT

from findsylls.segmentation.presets import (
    ThetaOscillatorSegmenter,     # Räsänen et al. 2018 — gammatone + oscillator (no GPU)
    EnergyPeriodicitySegmenter,   # Xie & Niyogi 2006 — periodicity + energy convex-hull (no GPU)
    RhythmGuidedSegmenter,        # Zhang & Glass 2009 — rhythm-guided peak recovery (no GPU)
    SylberSegmenter,              # Cho et al. 2025 — greedy cosine on Sylber HuBERT
    VGHubertMinCutSegmenter,      # Peng et al. 2023 — SSM MinCut on VG-HuBERT
    VGHubertCLSSegmenter,         # Peng & Harwath 2022 — CLS attention on VG-HuBERT
)
from findsylls.audio.utils import load_audio

audio, sr = load_audio("audio.wav")

# Theta oscillator (no model download, paper defaults: f=5, Q=0.5, N=8)
theta = ThetaOscillatorSegmenter()
syllables = theta.segment(audio, sr=sr)

# Xie & Niyogi periodicity+energy, and Zhang & Glass rhythm-guided (no model download)
nuclei = EnergyPeriodicitySegmenter().segment(audio, sr=sr)
nuclei = RhythmGuidedSegmenter().segment(audio, sr=sr)

# Sylber (requires findsylls[end2end])
sylber = SylberSegmenter()
syllables = sylber.segment(audio, sr=sr)

# VG-HuBERT MinCut (syllable mode, layer 8; requires findsylls[end2end])
vgh_mincut = VGHubertMinCutSegmenter(mode="syllable")
syllables = vgh_mincut.segment(audio, sr=sr)

# VG-HuBERT CLS attention (word mode, layer 9; requires findsylls[end2end])
vgh_cls = VGHubertCLSSegmenter(mode="word")
words = vgh_cls.segment(audio, sr=sr)

Speech Activity Detection and utterance boundaries

Real recordings often contain silence between utterances. Without SAD, peakdetect can place spurious boundaries in silent regions; neural segmenters waste computation on silence. Pass sad= to any segmenter to restrict segmentation to detected speech regions.

from findsylls.segmentation.presets import (
    SBSPeakdetectSegmenter,
    ThetaOscillatorSegmenter,
    SylberSegmenter,
)
from findsylls.audio.utils import load_audio

audio, sr = load_audio("recording_with_pauses.wav")

# SBS envelope + peak detection with energy VAD — fast, no model download
sbs = SBSPeakdetectSegmenter(sad="energy")
syllables = sbs.segment(audio, sr)

# Theta oscillator with energy VAD
theta = ThetaOscillatorSegmenter(sad="energy")
syllables = theta.segment(audio, sr)

# Silero VAD — more accurate on noisy audio (requires findsylls[end2end])
sylber = SylberSegmenter(sad="silero")
syllables = sylber.segment(audio, sr)

sad="energy" uses a simple energy threshold; sad="silero" uses a small neural VAD model. Both chunk the audio into speech regions, run the segmenter on each chunk independently, and reassemble the results with global timestamps.

add_utterance_boundaries controls whether the segmenter inserts boundary markers at the onset and offset of each speech region. For envelope-based segmenters (SBS, Theta) this is done in-algorithm — a valley is injected at both edges before peak detection, so the algorithm can produce segments that cover the full speech region even when there is no natural acoustic valley near the edge. Default is True for all segmenters.

# SAD off, boundaries on (default): covers the full single audio clip even
# when the first or last peak has no natural valley on its outer side.
sbs = SBSPeakdetectSegmenter(sad=None, add_utterance_boundaries=True)

# SAD + boundaries (recommended for multi-event recordings): each detected
# speech region gets boundary valleys, so no syllable is dropped at a region
# edge due to a missing valley.
sbs = SBSPeakdetectSegmenter(sad="energy", add_utterance_boundaries=True)

# Disable boundaries if passing pre-chunked single-utterance audio and you
# want strict valley-only boundaries with no edge injection.
sbs = SBSPeakdetectSegmenter(add_utterance_boundaries=False)

For neural segmenters (SylberSegmenter, VGHubertMinCutSegmenter, VGHubertCLSSegmenter) add_utterance_boundaries is accepted for API consistency but has no effect — those algorithms already produce contiguous segmentation of whatever chunk they receive.

SBS (left) and Theta (right) on a Kono recording with 6 speech events separated by silence (~12.4 s total). Top row: no SAD, no boundaries — both methods miss entire utterances where no valley bridges the silence gap. Middle row: no SAD, boundaries enabled — adds coverage only at the very start and end of the full recording, not between utterances. Bottom row: SAD + boundaries (recommended) — each speech region is segmented independently with boundary valleys at its edges; all speech events are captured.

Generic dispatch

from findsylls.segmentation import get_segmenter, list_segmenters, list_segmenter_presets

print(list_segmenters())
# ['peakdetect', 'convexhull', 'threshold', 'cls_attention', 'mincut', 'greedy_cosine']

print(list_segmenter_presets())
# {'theta_oscillator': ThetaOscillatorSegmenter, 'sylber': SylberSegmenter, ...}

segmenter = get_segmenter("mincut")
syllables = segmenter.segment(audio, sr=sr)

---

Feature Extraction (findsylls.features)

Feature extractors implement FeatureExtractor.extract(audio, sr) → np.ndarray (shape: [T, D]).

from findsylls.audio.utils import load_audio
from findsylls.features import MFCCExtractor, MelSpectrogramExtractor, HuBERTExtractor
from findsylls.plotting import plot_multiple_feature_matrices
import numpy as np

audio, sr = load_audio("audio.wav")

mfcc    = MFCCExtractor(n_mfcc=13)
melspec = MelSpectrogramExtractor(n_mels=64)

mfcc_feat = mfcc.extract(audio, sr)
mel_feat  = melspec.extract(audio, sr)

feature_results = {
    "MFCC (13 coeffs)":        (mfcc_feat,  np.linspace(0, len(audio)/sr, mfcc_feat.shape[0])),
    "Mel Spectrogram (64 bins)": (mel_feat, np.linspace(0, len(audio)/sr, mel_feat.shape[0])),
}

fig = plot_multiple_feature_matrices(audio, sr, feature_results)

MFCC and mel spectrogram feature matrices for the same utterance. Color encodes feature value; brighter = higher activation.

Available extractors: mfcc, melspectrogram, hubert, sylber, vghubert

from findsylls.features import get_extractor

extractor = get_extractor("hubert")          # vanilla HuBERT base (layer 9)
features  = extractor.extract(audio, sr)     # shape: [T, 768]

---

Embedding (findsylls.embedding)

Embedding wraps feature extraction + segmentation + pooling into a single call.

Single file

from findsylls import embed_audio

embeddings, metadata = embed_audio(
    "audio.wav",
    segmentation="peakdetect",
    features="mfcc",
    pooling="mean",                          # mean | max | median | onc
    segmentation_kwargs={"envelope_method": "hilbert"},
    return_metadata=True,
)

print(embeddings.shape)                      # (n_syllables, 13)
print(metadata["num_syllables"])
print(metadata["boundaries"])                # [(start, end), ...]

Corpus

from findsylls import embed_corpus, save_embeddings

results = embed_corpus(
    audio_files=["a.wav", "b.wav", "c.wav"],
    segmentation="peakdetect",
    features="mfcc",
    pooling="mean",
    segmentation_kwargs={"envelope_method": "hilbert"},
    n_jobs=4,
)

save_embeddings(results, "embeddings.npz")

Storage-backed corpus (large datasets)

For datasets that don't fit in RAM, write embeddings directly to disk:

from findsylls.embedding import embed_corpus_to_storage

bundle = embed_corpus_to_storage(
    audio_files=["a.wav", "b.wav", ...],
    output_dir="./embeddings",
    segmentation="peakdetect",
    features="mfcc",
    pooling="mean",
    segmentation_kwargs={"envelope_method": "hilbert"},
)

print(f"Embedded {bundle['num_success']}/{bundle['num_files']} files")
# Writes: ./embeddings/embedding_manifest.csv + ./embeddings/000000_*.npz

Preset-based embedding

from findsylls.embedding import EmbeddingPipeline

pipeline = EmbeddingPipeline(preset="sylber", pooling="mean")
embeddings, metadata = pipeline.embed_audio("audio.wav", return_metadata=True)

Available pooling methods: mean, max, median, onc

---

Discovery (findsylls.discovery)

Discovery clusters syllable embeddings into unsupervised categories.

from findsylls import embed_corpus, save_embeddings
from findsylls.discovery import DiscoveryPipeline
import numpy as np

# Embed a corpus
results = embed_corpus(audio_files=["a.wav", "b.wav", "c.wav"],
                       segmentation="peakdetect", features="mfcc", pooling="mean",
                       segmentation_kwargs={"envelope_method": "hilbert"})
embeddings = np.vstack([r["embeddings"] for r in results if r.get("success")])

# Cluster
pipeline = DiscoveryPipeline(method="kmeans", model_kwargs={"n_clusters": 50})
result   = pipeline.discover(embeddings)

print(result.labels)                          # cluster assignment per syllable
print(result.fit_metrics["silhouette"])
print(result.fit_metrics["davies_bouldin"])

Streaming clustering (corpus too large for RAM)

from findsylls.embedding import embed_corpus_to_storage
from findsylls.discovery import DiscoveryPipeline

bundle = embed_corpus_to_storage(audio_files=[...], output_dir="./embeddings",
                                  segmentation="peakdetect", features="mfcc", pooling="mean",
                                  segmentation_kwargs={"envelope_method": "hilbert"})

pipeline = DiscoveryPipeline(method="minibatch_kmeans", model_kwargs={"n_clusters": 50})
result   = pipeline.discover_from_storage(manifest_path=bundle["manifest_path"])

Memory comparison:

Approach	~500K syllables × 768-D
embed_corpus + vstack + KMeans	~10 GB RAM
embed_corpus_to_storage + discover_from_storage	~500 MB RAM

Available methods: kmeans, minibatch_kmeans, agglomerative

---

Full Corpus Workflow (findsylls.pipeline)

FindSyllsOrchestrator and discover_corpus run the entire pipeline — embed, discover, build manifests — in one call:

from findsylls import discover_corpus

result = discover_corpus(
    audio_files="data/**/*.wav",
    output_dir="./output",
    segmentation_method="peakdetect",
    features_method="mfcc",
    pooling_method="mean",
    discovery_method="kmeans",
    segmentation_kwargs={"envelope_method": "hilbert"},
)

print(result["corpus_manifest"])             # joined DataFrame
print(result["discovery_manifest_path"])
print(result["discovery_metrics"])

Or use the class directly:

from findsylls.pipeline.orchestrator import FindSyllsOrchestrator

orch = FindSyllsOrchestrator()

# Single file: segment + embed
embeddings, metadata = orch.segment_and_embed_audio(
    "audio.wav",
    segmentation_method="peakdetect",
    features_method="mfcc",
    pooling_method="mean",
    segmentation_kwargs={"envelope_method": "hilbert"},
)

---

Evaluation (findsylls.evaluation)

Evaluate segmentation against TextGrid annotations

from findsylls import segment_audio, evaluate_segmentation

syllables, _, _ = segment_audio(
    "audio.wav",
    method="peakdetect",
    segmentation_kwargs={"envelope_method": "hilbert"},
)

peaks = [nucleus for _, nucleus, _ in syllables]
spans = [(start, end) for start, _, end in syllables]

metrics = evaluate_segmentation(
    peaks=peaks,
    spans=spans,
    textgrid_path="annotations.TextGrid",
    tiers={"phone": 2, "syllable": 1, "word": 0},
)

# Keys: nuclei, syllable_boundaries, syllable_spans, word_boundaries, word_spans
print(metrics["syllable_boundaries"])
# {'TP': 12, 'Ins': 2, 'Del': 1, 'Sub': 0, 'Precision': ..., 'Recall': ..., 'F1': ...}

Batch evaluation over a corpus

from findsylls import run_evaluation

df = run_evaluation(
    textgrid_paths="data/**/*.TextGrid",
    wav_paths="data/**/*.wav",
    tiers={"phone": 2, "syllable": 1, "word": 0},
    method="peakdetect",
    segmentation_kwargs={"envelope_method": "hilbert"},
)

print(df.groupby("method")[["syllable_boundaries_f1", "word_spans_f1"]].mean())

Discovery label metrics

Connect cluster assignments to ground-truth TextGrid labels:

from findsylls.evaluation import (
    attach_textgrid_labels_to_manifest,
    compute_discovery_label_metrics,
)

labeled = attach_textgrid_labels_to_manifest(
    manifest=corpus_manifest,
    file_manifest=file_manifest_df,
    wav_paths=["a.wav", "b.wav"],
    textgrid_paths=["a.TextGrid", "b.TextGrid"],
    textgrid_tier_index=2,                       # phone tier
)

metrics = compute_discovery_label_metrics(labeled)
print(f"Cluster purity:  {metrics['cluster_purity']:.3f}")
print(f"Label purity:    {metrics['label_purity']:.3f}")
print(f"Normalized MI:   {metrics['label_norm_mutual_info']:.3f}")
print(f"Macro F1:        {metrics['macro_f1']:.3f}")

Visualize evaluation results

from findsylls import plot_segmentation_result

# df = output of run_evaluation(), file_id = stem of the audio file
fig, ax = plot_segmentation_result(
    df,
    file_id="SP20_117",
    envelope_fn="sbs",
    syll_tier=1,
    phone_tier=2,
    word_tier=0,
)

---

Preset System (findsylls.presets)

Named presets bundle segmentation + feature + pooling configurations from published papers:

from findsylls import get_preset, resolve_preset, list_presets

print(list_presets())
# ['sylber', 'vg_hubert_cls', 'vg_hubert_mincut']

cfg = get_preset("sylber")
# {'segmentation': 'greedy_cosine', 'features': 'sylber', 'pooling': 'mean', ...}

# Merge a preset with user overrides
cfg = resolve_preset("sylber", pooling="onc")

# Use directly with EmbeddingPipeline
from findsylls.embedding import EmbeddingPipeline
pipeline = EmbeddingPipeline(preset="sylber", pooling="mean")

---

CLI

# Segment audio into syllable boundaries
findsylls segment audio.wav --envelope hilbert --method peakdetect --out syllables.json

# Batch evaluation against TextGrid annotations
findsylls evaluate "data/**/*.wav" "data/**/*.TextGrid" \
  --phone-tier 2 --syllable-tier 1 --word-tier 0 \
  --envelope hilbert --method peakdetect \
  --out results.csv --aggregate summary.csv

---

Methods Reference

Envelope methods

rms · hilbert · lowpass · sbs · theta · gammatone · periodicity · cls_attention · greedy_cosine · mincut

Segmentation methods (dispatch strings)

peakdetect · convexhull · threshold · cls_attention · mincut · greedy_cosine

Preset segmenters (paper-replication classes)

SBSPeakdetectSegmenter · ThetaOscillatorSegmenter · EnergyPeriodicitySegmenter · RhythmGuidedSegmenter · SylberSegmenter · VGHubertMinCutSegmenter · VGHubertCLSSegmenter

Feature extractors

mfcc · melspectrogram · hubert · sylber · vghubert

Pooling methods

mean · max · median · onc

Discovery methods

kmeans · minibatch_kmeans · agglomerative

---

Preset Citations

Every preset segmenter ships with the full citation for its source paper. Access it programmatically without loading any model:

from findsylls.segmentation.presets import list_segmenter_presets

for name, cls in list_segmenter_presets().items():
    print(f"[{name}]")
    print(cls.REFERENCE)
    print()

Or on an instance (useful when you already have the object):

seg = ThetaOscillatorSegmenter()
seg.cite()

---

Theta Oscillator — Räsänen, Doyle & Frank (2018)

Räsänen, O., Doyle, G., & Frank, M. C. (2018). "Pre-linguistic segmentation of speech into syllable-like units." Cognition, 171, 130–150. https://doi.org/10.1016/j.cognition.2017.11.003

MATLAB implementation: https://github.com/orasanen/thetaOscillator

Energy + Periodicity — Xie & Niyogi (2006)

Xie, Z., & Niyogi, P. (2006). "Robust Acoustic-Based Syllable Detection." Interspeech 2006. https://doi.org/10.21437/Interspeech.2006-440

Rhythm-Guided — Zhang & Glass (2009)

Zhang, Y., & Glass, J. R. (2009). "Speech rhythm guided syllable nuclei detection." ICASSP 2009, 3797–3800. https://doi.org/10.1109/ICASSP.2009.4960454

Sylber — Cho et al. (2025)

Cho, C. J., Lee, N., Gupta, A., Agarwal, D., Chen, E., Black, A. W., & Anumanchipalli, G. K. (2025). "Sylber: Syllabic Embedding Representation of Speech from Raw Audio." ICLR 2025. https://arxiv.org/abs/2410.07168

VG-HuBERT MinCut — Peng et al. (2023)

Peng, P., Shang, Z., Harwath, D., & others (2023). "Syllable Discovery and Cross-Lingual Generalization in a Visually Grounded, Self-Supervised Speech Model." Interspeech 2023. https://doi.org/10.21437/Interspeech.2023-1430

Code: https://github.com/jasonppy/syllable-discovery

VG-HuBERT CLS Attention — Peng & Harwath (2022)

Peng, P., & Harwath, D. (2022). "Self-Supervised Representation Learning for Speech Using Visual Grounding and Masked Language Modeling." Interspeech 2022. https://doi.org/10.21437/Interspeech.2022-10631

Code: https://github.com/jasonppy/word-discovery

---

Citation

@misc{martinez2026findsyllslanguageagnostictoolkitsyllablelevel,
  title={findsylls: A Language-Agnostic Toolkit for Syllable-Level Speech Tokenization and Embedding},
  author={Héctor Javier Vázquez Martínez},
  year={2026},
  eprint={2603.26292},
  archivePrefix={arXiv},
  primaryClass={cs.CL},
  url={https://arxiv.org/abs/2603.26292},
}

License

MIT. See LICENSE.