diff --git a/stackvox/text.py b/stackvox/text.py
index 8daffb7..3675c91 100644
--- a/stackvox/text.py
+++ b/stackvox/text.py
@@ -18,6 +18,7 @@
     "markdown_to_paragraphs",
     "strip_emoji",
     "strip_thousands_separators",
+    "versions_to_words",
     "decimals_to_words",
     "expand_units",
     "apply_pronunciations",
@@ -84,6 +85,21 @@ def strip_thousands_separators(text: str) -> str:
     return re.sub(r"(?<=\d),(?=\d)", "", text)
 
 
+_VERSION = re.compile(r"(?<!\d)\d+(?:\.\d+){2,}(?!\d)")
+
+
+def versions_to_words(text: str) -> str:
+    """1.2.3 -> "1 point 2 point 3"; 0.7.0 -> "0 point 7 point 0".
+
+    A dotted version string has more than one decimal point, which
+    `decimals_to_words` can't handle — it splits only the first, leaving a stray
+    full stop mid-number ("0.7.0" -> "0 point 7.0"). Run this before it. Two-part
+    numbers are left untouched so ordinary decimals still read digit-by-digit,
+    and a trailing sentence stop (``upgrade to 1.2.3.``) is preserved.
+    """
+    return _VERSION.sub(lambda m: m.group(0).replace(".", " point "), text)
+
+
 def decimals_to_words(text: str) -> str:
     """1198.9 -> "1198 point 9"; 770.72 -> "770 point 7 2". Removes the bare
     "." between digits, which TTS can otherwise read as a full stop."""
@@ -287,6 +303,7 @@ def _shape_paragraph(
     if expand_units_flag:  # units BEFORE decimals (see note above)
         text = expand_units(text, locale)
     if expand_numbers_flag:
+        text = versions_to_words(text)  # multi-dot versions BEFORE the decimal split
         text = decimals_to_words(text)
     return re.sub(r"[ \t]{2,}", " ", text).strip()
 
diff --git a/tests/test_text.py b/tests/test_text.py
index 5538445..6206c8f 100644
--- a/tests/test_text.py
+++ b/tests/test_text.py
@@ -10,6 +10,7 @@
     shape_pauses,
     strip_emoji,
     strip_thousands_separators,
+    versions_to_words,
 )
 
 # --- numbers ---------------------------------------------------------------
@@ -28,6 +29,31 @@ def test_year_is_not_a_decimal():
     assert decimals_to_words("built in 2023") == "built in 2023"
 
 
+def test_semver_all_parts_spoken():
+    assert versions_to_words("0.7.0") == "0 point 7 point 0"
+    assert versions_to_words("1.2.3") == "1 point 2 point 3"
+
+
+def test_semver_multi_digit_component():
+    assert versions_to_words("1.20.3") == "1 point 20 point 3"
+
+
+def test_version_preserves_trailing_sentence_stop():
+    assert versions_to_words("upgrade to 1.2.3.") == "upgrade to 1 point 2 point 3."
+
+
+def test_two_part_number_is_left_for_the_decimal_pass():
+    # Only two components — an ordinary decimal, not a version.
+    assert versions_to_words("3.14") == "3.14"
+    assert decimals_to_words(versions_to_words("3.14")) == "3 point 1 4"
+
+
+def test_semver_normalizes_end_to_end():
+    out = normalize_for_speech("Upgraded to 0.7.0 today.", markdown=False)
+    assert "0 point 7 point 0" in out
+    assert "0 point 7.0" not in out
+
+
 # --- units -----------------------------------------------------------------