diff --git a/pyproject.toml b/pyproject.toml
index 98987a5..544cc1a 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -10,7 +10,7 @@ readme = "README.md"
 requires-python = ">=3.11"
 dependencies = [
     "pose-format>=0.8.1",
-    "numpy<2",
+    "numpy",
     "pympi-ling",
     "torch",
     "pose-anonymization",
@@ -20,12 +20,7 @@ dependencies = [
 
 [project.optional-dependencies]
 dgs = [
-    "sign_language_datasets",
-    "tensorflow<2.19",
-    "tensorflow-datasets==4.9.2",
-    "pyarrow",  # transitive: tensorflow-datasets needs it at import time
     "lxml",
-    "webvtt-py",  # transitive: sign_language_datasets eagerly imports mediapi_skel
 ]
 platform = [
     "httpx",
diff --git a/sign_language_segmentation/datasets/dgs/dataset.py b/sign_language_segmentation/datasets/dgs/dataset.py
index b576efb..74cf9ee 100644
--- a/sign_language_segmentation/datasets/dgs/dataset.py
+++ b/sign_language_segmentation/datasets/dgs/dataset.py
@@ -7,25 +7,15 @@
 from pose_format import Pose
 from pose_format.pose_body import EmptyPoseBody
 
-# shim: sign_language_datasets imports get_dl_dirname which was renamed in tfds >=4.9
-import tensorflow_datasets.core.download.resource as _tfds_resource
-if not hasattr(_tfds_resource, "get_dl_dirname"):
-    _tfds_resource.get_dl_dirname = _tfds_resource.get_dl_fname
-
-from sign_language_datasets.datasets.dgs_corpus.dgs_utils import get_elan_sentences
-
+from sign_language_segmentation.datasets.dgs.utils import (
+    get_elan_sentences,
+    is_joke,
+)
 from sign_language_segmentation.datasets.common import CACHE_DIR, BaseSegmentationDataset, Split, md5sum
 
 EXCLUDED_IDS: set[str] = {"1289910", "1245887", "1289868", "1246064", "1584617"}
 
 
-def is_joke(corpus_dir: Path, doc_id: str) -> bool:
-    cmdi_path = corpus_dir / "videos" / doc_id / "data.cmdi"
-    if not cmdi_path.exists():
-        return False
-    return "<cmdp:Task>Joke</cmdp:Task>" in cmdi_path.read_text()
-
-
 class DGSSegmentationDataset(BaseSegmentationDataset):
     """reads DGS corpus poses on the fly with efficient partial reading."""
 
diff --git a/sign_language_segmentation/datasets/dgs/utils.py b/sign_language_segmentation/datasets/dgs/utils.py
new file mode 100644
index 0000000..3041281
--- /dev/null
+++ b/sign_language_segmentation/datasets/dgs/utils.py
@@ -0,0 +1,78 @@
+import pympi
+from pathlib import Path
+
+def is_joke(corpus_dir: Path, doc_id: str) -> bool:
+    cmdi_path = corpus_dir / "videos" / doc_id / "data.cmdi"
+    if not cmdi_path.exists():
+        return False
+    return "<cmdp:Task>Joke</cmdp:Task>" in cmdi_path.read_text()
+
+def get_elan_sentences(elan_path: str):
+    eaf = pympi.Elan.Eaf(elan_path)  # TODO add "suppress_version_warning=True" when pympi 1.7 is released
+
+    timeslots = eaf.timeslots
+
+    for participant in ["A", "B"]:
+        german_tier_name = "Deutsche_Übersetzung_" + participant
+        if german_tier_name not in eaf.tiers:
+            continue
+
+        german_text = eaf.tiers[german_tier_name][0]
+
+        english_tier_name = "Translation_into_English_" + participant
+        english_text = list(eaf.tiers[english_tier_name][0].values()) if english_tier_name in eaf.tiers else []
+
+        all_glosses = []
+        for hand in ["r", "l"]:
+            hand_tier = "Lexem_Gebärde_" + hand + "_" + participant
+            if hand_tier not in eaf.tiers:
+                continue
+
+            gloss = {
+                _id: {"start": timeslots[s], "end": timeslots[e], "gloss": val, "hand": hand}
+                for _id, (s, e, val, _) in eaf.tiers[hand_tier][0].items()
+            }
+            for tier in ["Lexeme_Sign", "Gebärde", "Sign"]:
+                items = eaf.tiers[tier + "_" + hand + "_" + participant][1]
+                for ref, val, _1, _2 in items.values():
+                    if ref in gloss:  # 2 files have a missing reference
+                        gloss[ref][tier] = val
+
+            all_glosses += list(gloss.values())
+
+        all_mouthings = []
+
+        tier_name = "Mundbild_Mundgestik_" + participant
+        items = eaf.tiers[tier_name][0]
+
+        # structure of entries:
+        # {'a2768296': ('ts42', 'ts43', 'tochter', None), ... }
+
+        for s, e, val, _ in items.values():
+            mouthing_entry = {"start": timeslots[s], "end": timeslots[e], "mouthing": val}
+            all_mouthings.append(mouthing_entry)
+
+        for _id, (s, e, val, _) in german_text.items():
+            sentence = {"id": _id, "participant": participant, "start": timeslots[s], "end": timeslots[e], "german": val}
+
+            # Add English sentence
+            english_sentence = [val2 for (s2, e2, val2, _) in english_text if s == s2 and e == e2]
+            sentence["english"] = english_sentence[0] if len(english_sentence) > 0 else None
+
+            # Add glosses
+            sentence["glosses"] = list(
+                sorted(
+                    [item for item in all_glosses if item["start"] >= sentence["start"] and item["end"] <= sentence["end"]],
+                    key=lambda d: d["start"],
+                )
+            )
+
+            # add mouthings
+            sentence["mouthings"] = list(
+                sorted(
+                    [item for item in all_mouthings if item["start"] >= sentence["start"] and item["end"] <= sentence["end"]],
+                    key=lambda d: d["start"],
+                )
+            )
+
+            yield sentence
\ No newline at end of file