adeelahmad · Oct 18, 2026 · Oct 18, 2026
diff --git a/changelog.d/whisper-lang.md b/changelog.d/whisper-lang.md
@@ -0,0 +1,2 @@
+- **Whisper keeps the language it heard.** With no language set in Settings › Transcription, recordings Whisper
+  transcribes now keep the language it detected, as other engines' do, for search filters and the text cleanup.
diff --git a/fastapi_backend/app/domain/ingest.py b/fastapi_backend/app/domain/ingest.py
@@ -351,14 +351,18 @@ def __init__(self, cfg, mlx=False):
             raise EngineMissing(f"{name} isn't installed: uv sync --extra {'mlx' if mlx else 'whisper'}") from e
 
     def transcribe(self, audio, terms=()):
-        """The segments; `terms` (vocabulary.py) go in as the prompt, the names Whisper should expect."""
+        """The segments, in the language set in Settings or, with none set, the one Whisper heard (so the recording's
+        language is kept either way); `terms` (vocabulary.py) go in as the prompt, the names Whisper should expect."""
         hint = {"initial_prompt": vocabulary.prompt(terms)} if terms else {}
         if self.mlx:
-            segs = self.m.transcribe(audio, path_or_hf_repo=self.repo, word_timestamps=True, language=self.lang, **hint)["segments"]
+            got = self.m.transcribe(audio, path_or_hf_repo=self.repo, word_timestamps=True, language=self.lang, **hint)
+            heard, segs = got.get("language"), got["segments"]
             rows = [(s["start"], s["end"], s["text"], [(w["word"], w["start"], w["end"]) for w in s.get("words") or []]) for s in segs]
         else:
-            segs, _ = self.m.transcribe(audio, vad_filter=True, word_timestamps=True, language=self.lang, **hint)
+            segs, info = self.m.transcribe(audio, vad_filter=True, word_timestamps=True, language=self.lang, **hint)
+            heard = getattr(info, "language", None)
             rows = [(s.start, s.end, s.text, [(w.word, w.start, w.end) for w in (s.words or [])]) for s in segs]
+        lang = self.lang or (str(heard).lower() if heard else None)
         return [
             {
                 "t0": int(a * 1000),
@@ -367,7 +371,7 @@ def transcribe(self, audio, terms=()):
                 "raw_text": None,
                 "emotion": None,
                 "event": None,
-                "lang": self.lang,
+                "lang": lang,
                 "words": json.dumps([[w.strip(), int(p * 1000), int(q * 1000)] for w, p, q in ws]),
             }
             for a, b, x, ws in rows

diff --git a/fastapi_backend/tests/domain/test_speech.py b/fastapi_backend/tests/domain/test_speech.py
@@ -617,3 +617,29 @@ def transcribe(self, audio, terms=()):
     monkeypatch.setattr(ingest, "decode", lambda p: clip())
     ingest.transcribe_one(db, eff, rid, log=lambda *a: None, engine=Hearing())
     assert Hearing.heard == words
+
+
+def test_whisper_keeps_the_language_it_heard():
+    """With no language set, the recording's language is the one Whisper detected; a language set in Settings wins."""
+    from types import SimpleNamespace as NS
+
+    from app.domain import ingest
+
+    class Faster:
+        def transcribe(self, audio, **kw):
+            seg = NS(start=0.0, end=1.5, text=" Bonjour à tous.", words=[NS(word=" Bonjour", start=0.0, end=0.6)])
+            return iter([seg]), NS(language="fr", language_probability=0.97)
+
+    class Mlx:
+        def transcribe(self, audio, **kw):
+            return {"language": "de", "segments": [{"start": 0.0, "end": 1.0, "text": " Guten Tag.", "words": []}]}
+
+    def engine(m, mlx, lang):
+        w = object.__new__(ingest.Whisper)
+        w.m, w.mlx, w.repo, w.lang, w.name = m, mlx, "repo", lang, "whisper"
+        return w
+
+    segs = engine(Faster(), False, None).transcribe(None)
+    assert ([s["lang"] for s in segs], ingest._language(segs)) == (["fr"], "fr")
+    assert [s["lang"] for s in engine(Mlx(), True, None).transcribe(None)] == ["de"]
+    assert [s["lang"] for s in engine(Faster(), False, "en").transcribe(None)] == ["en"]  # the setting, as before
Original file line number	Diff line number	Diff line change
		@@ -0,0 +1,2 @@
		- Whisper keeps the language it heard. With no language set in Settings › Transcription, recordings Whisper
		transcribes now keep the language it detected, as other engines' do, for search filters and the text cleanup.