signlangtk 0.1.4

Sign Language Toolkit for sign language research

Sign Language Toolkit (SLTK)

A research toolkit for sign language video analysis. SLTK provides a complete pipeline from raw video to rich, multi-tier ELAN annotation files: 3D hand reconstruction, full-body pose, face tracking, automatic sign segmentation, gloss spotting, and non-manual signal detection.

Video (.mp4)
    │
    ├─ WiLoR ──────► 3D hand keypoints + MANO rotations   (_wilor.h5)
    ├─ NLF ────────► Full-body SMPL-X pose                 (_nlf.h5)
    ├─ TEASER ─────► FLAME face parameters                 (_teaser.h5)
    │
    ├─ Segmenter ──► Sign boundaries (BIO labels)
    ├─ SignRep ────► Gloss spotting (dictionary matching)
    ├─ NMS ────────► Blinks, nods, shakes, mouth, gaze
    │
    └─ All results ► Multi-tier ELAN file (.eaf)

Installation

# Core library (ELAN I/O, CLI, corpus tools)
pip install signlangtk

# With web interface
pip install "signlangtk[api]"

# With hand extraction (WiLoR) — requires CUDA
pip install "signlangtk[wilor]"

# Full body (NLF) + face (TEASER)
pip install "signlangtk[nlf,teaser]"

# Everything (excludes rtmpose/smplfx which need mmcv via mim)
pip install "signlangtk[all]"

The PyPI package is signlangtk, the Python import is sltk:

import sltk
from sltk.extraction.wilor import WiLoRExtractor
from sltk.extraction.nlf import NLFExtractor
from sltk.extraction.teaser import TeaserExtractor

Requires Python 3.10+. For development from source:

git clone https://github.com/ed-fish/Sign-Language-Toolkit.git
cd Sign-Language-Toolkit
pip install -e ".[dev,api]"

---

Quick Start: Video to Multi-Tier ELAN

This end-to-end example takes a single video and produces an ELAN file with sign boundaries, spotted glosses, and non-manual signals on separate tiers.

from sltk.extraction.wilor import WiLoRExtractor
from sltk.extraction.teaser import TeaserExtractor
from sltk.segmentation.runner import get_runner
from sltk.segmentation.h5_loader import h5_to_features
from sltk.segmentation.postprocess import extract_segments
from sltk.nms.runner import detect_nms, export_results
from sltk.io.elan_roundtrip import ElanDocument

VIDEO = "recording.mp4"
FPS = 25.0

# ── Step 1: Extract poses ───────────────────────────────────────────
# Hands (WiLoR → MANO 3D)
with WiLoRExtractor() as ext:
    ext.load_model()
    ext.extract_from_video(VIDEO, "recording_wilor.h5")

# Face (TEASER → FLAME)
with TeaserExtractor() as ext:
    ext.load_model()
    ext.extract_from_video(VIDEO, "recording_teaser.h5")

# ── Step 2: Segment signs ──────────────────────────────────────────
features = h5_to_features("recording_wilor.h5")   # (T, 192)
runner = get_runner()
labels = runner.predict(features)                  # 0=OUT, 1=IN, 2=BEGIN
segments = extract_segments(labels)                # [(start, end), ...]

# ── Step 3: Detect non-manual signals ──────────────────────────────
blinks, nms_events, quality = detect_nms(
    "recording_teaser.h5",
    detectors={"all"},
)

# ── Step 4: Build multi-tier ELAN file ─────────────────────────────
doc = ElanDocument.new(video_path=VIDEO)

# Add segmentation tier
doc.add_tier("Segmentation")
for start_frame, end_frame in segments:
    doc.add_segment("Segmentation", start_frame / FPS, end_frame / FPS, "SIGN")

# Add NMS tiers (blinks, head movements, mouth, eyebrows, gaze)
for tier_name in ["BLINK", "HEAD-NOD", "HEAD-SHAKE", "HEAD-TILT",
                  "MOUTH-MOVEMENT", "EYEBROW-RAISE", "EYE-GAZE"]:
    doc.add_tier(tier_name)

for b in blinks:
    doc.add_segment("BLINK", b.start_frame / FPS, b.end_frame / FPS, "blink")

for ev in nms_events:
    doc.add_segment(ev.tier, ev.start_frame / FPS, ev.end_frame / FPS, ev.label)

doc.save("recording_full.eaf")

Open recording_full.eaf in ELAN to see all tiers aligned with the video.

---

Extraction: Hands, Body, and Face

All extractors share the same interface: load_model(), extract_from_video(), process_batch(). Weights are auto-downloaded from HuggingFace Hub on first use.

WiLoR — 3D Hand Reconstruction

Produces 21 keypoints per hand + MANO rotation matrices per frame.

from sltk.extraction.wilor import WiLoRExtractor, WiLoRConfig

config = WiLoRConfig(
    device="cuda:0",
    img_batch_size=128,
    rescale_factor=2.0,
)
with WiLoRExtractor(config) as ext:
    ext.load_model()
    result = ext.extract_from_video("video.mp4", "video_wilor.h5")

Output H5 structure:

video_wilor.h5
├── attrs: fps, num_frames, resolution
├── frame_idx      (num_frames, 2)           # sparse: (start_idx, count)
├── kpts_3d        (num_detections, 21, 3)   # 3D hand keypoints
├── right          (num_detections,)          # True = right hand
└── mano/
    ├── hand_pose      (num_detections, 15, 3, 3)  # joint rotations
    └── global_orient  (num_detections, 1, 3, 3)    # wrist rotation

NLF — Full-Body SMPL-X

Produces 55 SMPL-X joints (body + hands + face landmarks) per frame.

from sltk.extraction.nlf import NLFExtractor, NLFConfig

with NLFExtractor(NLFConfig(device="cuda:0")) as ext:
    ext.load_model()
    result = ext.extract_from_video("video.mp4", "video_nlf.h5")

TEASER — FLAME Face Parameters

Produces FLAME 3D face parameters: jaw pose, expression coefficients, shape, eyelid state, and head pose per frame. This is what drives NMS detection.

from sltk.extraction.teaser import TeaserExtractor, TeaserConfig

with TeaserExtractor(TeaserConfig(device="cuda:0")) as ext:
    ext.load_model()
    result = ext.extract_from_video("video.mp4", "video_teaser.h5")

Batch Processing

All extractors support batch processing over a directory:

from pathlib import Path
from sltk.extraction.wilor import WiLoRExtractor

with WiLoRExtractor() as ext:
    ext.load_model()
    results = ext.process_batch(
        video_paths=list(Path("videos/").glob("*.mp4")),
        output_dir=Path("poses/"),
        skip_existing=True,
    )
    for path, result in results.items():
        print(f"{path}: {result.num_frames} frames, {result.num_detections} detections")

Model Weights

All weights are auto-downloaded from HuggingFace Hub and cached at ~/.cache/sltk/weights/. Override with environment variables:

Variable	Model
SLTK_WILOR_MODEL_PATH	WiLoR hand model
SLTK_NLF_MODEL_PATH	NLF body model
SLTK_TEASER_CHECKPOINT	TEASER face model
SLTK_SIGNREP_CHECKPOINT	SignRep embedding model
SLTK_SEGMENTOR_PATH	Segmenter model

---

Segmentation: Finding Sign Boundaries

The segmenter is a 4-layer Transformer that reads WiLoR hand features and predicts per-frame BIO labels (0=OUT, 1=IN_SIGN, 2=BEGIN).

From WiLoR H5

from sltk.segmentation.runner import get_runner, segment_h5
from sltk.segmentation.output import OutputFormat
from sltk.segmentation.h5_loader import h5_to_features
from sltk.segmentation.postprocess import extract_segments

# High-level: segment a file → ELAN or JSON
segment_h5(
    "video_wilor.h5",
    output_path="video_segments.eaf",
    output_format=OutputFormat.ELAN,
    fps=25.0,
    media_path="video.mp4",
)

# Low-level: get raw predictions
features = h5_to_features("video_wilor.h5")  # (T, 192)
runner = get_runner()
labels = runner.predict(features)             # (T,) values 0/1/2
segments = extract_segments(labels)           # [(start_frame, end_frame), ...]

---

Gloss Spotting: Matching Signs to a Dictionary

SignRep extracts 768-dim visual features from 16-frame sliding windows, then matches each detected segment against a dictionary of known sign features using cosine similarity.

from sltk.embedding.pipeline import SignRepPipeline

pipeline = SignRepPipeline()

# Extract dense features from the full video
continuous = pipeline.extract_continuous("video.mp4", stride=4)

# Load your sign dictionary (folder of .npz files, one per sign)
dictionary = pipeline.load_dictionary(["/data/dictionaries/bsldict/signrep/"])

# Match segments to dictionary entries
result = pipeline.spot(
    features=continuous,
    segments=[{"segment_id": 0, "start_frame": 12, "end_frame": 45}],
    dictionary=dictionary,
    top_k=10,
)

for seg in result.segments:
    print(f"Segment {seg.start_ms}ms-{seg.end_ms}ms:")
    for gl in seg.top_glosses:
        print(f"  {gl['gloss']} ({gl['similarity']:.3f})")

# Save as ELAN (creates Rank-1..N and Score-1..N tiers)
result.save_eaf("video_spotted.eaf", media_path="video.mp4")

Building a Dictionary

Before spotting, build a dictionary from isolated sign videos:

result = pipeline.extract_dictionary("isolated_sign.mp4", method="middle")
result.save_npz("dictionary/HELLO.npz")

---

Non-Manual Signals: Face and Head Analysis

Detect blinks, head nods/shakes/tilts, mouth movements, eyebrow raises, gaze direction, and eye squints from TEASER face-tracking data.

from sltk.nms.runner import detect_nms, export_results

# Detect all NMS events from a TEASER H5 file
blinks, nms_events, quality = detect_nms(
    "video_teaser.h5",
    detectors={"all"},           # or specific: {"blink", "nod", "mouth"}
    smpl_path="video_nlf.h5",   # optional: enables gaze detection
)

# Export to ELAN (one tier per signal type)
export_results(
    blinks, nms_events, "video_teaser.h5",
    output_dir="output/",
    formats=["elan", "json", "csv"],
    media_path="video.mp4",
)

Available Detectors

Detector	ELAN Tier	Signal	Input
blink	BLINK	Eye closures	TEASER eyelid params
nod	HEAD-NOD	Vertical head oscillation	TEASER head pitch
shake	HEAD-SHAKE	Horizontal head oscillation	TEASER head yaw
tilt	HEAD-TILT	Side-to-side head tilt	TEASER head roll
mouth	MOUTH-MOVEMENT	Lip/mouth movement	FLAME expression
eyebrow	EYEBROW-RAISE	Eyebrow raise/furrow	FLAME expression
gaze	EYE-GAZE	Gaze direction	NLF eye pose
squint	EYE-SQUINT	Partial eye closure	TEASER eyelid

---

Feature Representations

SLTK computes several feature representations from the extracted H5 pose data.

WiLoR Segmenter Features (192-dim)

MANO rotation matrices converted to axis-angle, both hands concatenated. Used by the Transformer segmenter.

from sltk.segmentation.h5_loader import h5_to_features
features = h5_to_features("video_wilor.h5")  # (T, 192)

Angle Features (104-dim)

Body joint angles + hand Euler angles from MANO rotations.

from sltk.processing.features import compute_angle_features
angles = compute_angle_features(body_poses, right_hand, left_hand)  # (T, 104)

HaMeR Features (288-dim)

Flattened MANO rotation matrices for both hands.

from sltk.processing.features import compute_hamer_features
hamer = compute_hamer_features(
    mano_global_orient_right, mano_hand_pose_right,
    mano_global_orient_left, mano_hand_pose_left,
)  # (T, 288)

SignRep Embeddings (768-dim)

Dense visual features from the SignRep ViT model. 16-frame windows, L2-normalized.

from sltk.embedding.pipeline import SignRepPipeline
pipeline = SignRepPipeline()
continuous = pipeline.extract_continuous("video.mp4", stride=4)
# continuous.features: (num_windows, 768)

Combined Features from H5

from sltk.processing.features import load_features_from_nlf_wilor
angles, hamer = load_features_from_nlf_wilor("video_nlf.h5", "video_wilor.h5")

---

ELAN File I/O

Writing a New ELAN File

from sltk.io.elan_roundtrip import ElanDocument

doc = ElanDocument.new(video_path="video.mp4")
doc.add_tier("Gloss")
doc.add_tier("NMS")
doc.add_segment("Gloss", 0.0, 1.5, "HELLO")
doc.add_segment("Gloss", 1.5, 3.0, "WORLD")
doc.add_segment("NMS", 0.2, 0.8, "nod")
doc.save("output.eaf")

Reading and Modifying Existing ELAN Files

doc = ElanDocument.open("annotations.eaf")
tiers = doc.get_tiers()         # list of TierInfo
segments = doc.get_segments()   # list of SegmentInfo (all tiers)

# Add new annotations from pipeline results
doc.add_tier("AutoSegmentation")
doc.add_segment("AutoSegmentation", 1.0, 2.5, "SIGN")
doc.save()  # preserves all original XML structure

Simple Read/Write

from sltk.io import read_eaf, write_eaf
from sltk.data import Segment, SegmentList

# Read
segments = read_eaf("annotations.eaf", tiers=["Gloss"])

# Write
new_segments = SegmentList([
    Segment(start=0.0, end=1.5, label="HELLO", tier="Gloss"),
    Segment(start=1.5, end=3.0, label="WORLD", tier="Gloss"),
])
write_eaf(new_segments, "output.eaf", video_path="video.mp4")

---

Web Interface

SLTK includes a React frontend for browsing workspaces, running processing jobs, and exploring corpus data.

# Development: FastAPI (port 8000) + Vite (port 5173)
bash scripts/run_dev.sh

# Production
cd frontend && npm ci && npm run build && cd ..
sltk serve --host 0.0.0.0 --port 8000

Route	Page	Purpose
/	Workspaces	Create/switch workspaces, scan directories
/process	Process	Submit segmentation and spotting jobs
/explore	Explore	Search glosses, view video clips, corpus statistics
/viewer	Viewer	Video playback with annotation overlay
/analysis/*	Analysis	Vocabulary, concordance, n-grams, collocations, durations

---

CLI Reference

sltk convert input.npy output.h5 --from mediapipe --to wilor --fps 25
sltk evaluate predictions.txt references.txt --task translation
sltk to-elan segments.json --video source.mp4 --output annotations.eaf
sltk from-elan annotations.eaf --output segments.json --tier Gloss
sltk info video_wilor.h5
sltk serve --host 0.0.0.0 --port 8000 --reload

Configuration

Variable	Description	Default
SLTK_CORS_ORIGINS	Allowed CORS origins	http://localhost:5173,http://localhost:3000
SLTK_ALLOWED_PATHS	Filesystem whitelist for API	/vol/research,/home
SLTK_WEIGHTS_DIR	Override weight cache location	~/.cache/sltk/weights/

Supported Pose Formats

Format	Extractor	Keypoints	Description
WiLoR	WiLoRExtractor	21 per hand	MANO 3D hand mesh with rotation matrices
NLF/SMPL-X	NLFExtractor	55 joints	Full body with axis-angle rotations
TEASER/FLAME	TeaserExtractor	FLAME params	Face parameters: jaw, expression, shape, eyelid
MediaPipe	MediaPipeExtractor	33+42+468	Fast 2D/3D holistic landmarks
RTMPose	RTMPoseExtractor	133	COCO-WholeBody, multi-person

All stored as HDF5 (.h5) files.

Testing

pytest                       # Full suite (1500+ tests)
pytest -m "not slow"         # Skip slow tests
pytest -m api                # API tests only

License

CC-BY-NC-4.0