Make auto the default speech language and carry what was detected through

New installs start detecting instead of being locked to one language; a stored
value from before this default still wins. The dictation chain asks
transcribe_detected() in auto mode, records the detected code in history as
speech_language, hands it to the cleanup prompt (a detected Turkish recording
gets the Turkish prompt and glossary rule), and reports it on the socket reply.
The stale comment claiming whisper.cpp's -l auto does not detect is corrected.

dikte/app.py

@@ -1058,7 +1058,7 @@ class Dikte:
         if not self._transcripts_pending:
             self._settle(DICTATION, payload)
 
-    def _on_finished(self, _raw, text, warning):
+    def _on_finished(self, _raw, text, warning, speech_language):
         if warning:
             # The text was still pasted, but cleanup did not run. Say so loudly:
             # a rejected key otherwise looks exactly like working dictation.
@@ -1079,9 +1079,10 @@ class Dikte:
                 t("{action}: {preview}", action=action, preview=_preview(text))
             )
         self._transcript_settled({"ok": True, "text": text, "raw": _raw,
-                                  "warning": warning})
+                                  "warning": warning,
+                                  "speech_language": speech_language})
 
-    def _on_ask_finished(self, _raw, text, warning):
+    def _on_ask_finished(self, _raw, text, warning, speech_language):
         agent = assistant.display_name(self.conf)
         if warning:
             # A tool the agent was not allowed to touch otherwise looks exactly
@@ -1102,7 +1103,8 @@ class Dikte:
             )
         self._set_ask_state(IDLE)
         self._settle(ASK, {"ok": True, "answer": text, "question": _raw,
-                           "warning": warning, "agent": agent})
+                           "warning": warning, "agent": agent,
+                           "speech_language": speech_language})
 
     def _on_ask_cancelled(self):
         self.ask_overlay.show_done(t("Stopped."), 2000)

dikte/config.py

@@ -395,7 +395,10 @@ DEFAULTS = {
     "transcribe_model": "gpt-4o-transcribe",           # used when provider is openai
     "groq_transcribe_model": "whisper-large-v3-turbo",
     "openrouter_transcribe_model": "openai/gpt-4o-transcribe",
-    "language": "tr",
+    # Detect on the machine by default, so a new install needs no language to
+    # be told. whisper.cpp detects; the hosted providers detect when handed no
+    # language; a stored value from before this default overrides it.
+    "language": "auto",
     "transcribe_prompt": "",
 
     # --- whisper.cpp, on this machine ---------------------------------------
@@ -702,8 +705,16 @@ class Config:
         return self["cleanup_provider"] == "local"
 
     def cleanup_prompt(self, with_timestamps=False, with_speakers=False,
-                       subtitles=False):
+                       subtitles=False, speech=""):
-        turkish = i18n.language() == "tr"
+        """`speech` is the two-letter code of the language that was heard, when
+        the transcription model reported one. The default prompts and the
+        glossary rule only exist in Turkish and English, so a detected Turkish
+        recording gets the Turkish prompt and any other detected language — or
+        none at all — the English one, which is written not to care what
+        language the transcript is in. Nothing else calls this with it, so the
+        interface language keeps deciding everywhere the speech was not asked
+        about."""
+        turkish = (speech == "tr") if speech else i18n.language() == "tr"
         if subtitles:
             prompt = (self["file_cleanup_prompt"].strip()
                       or default_file_cleanup_prompt())

dikte/ggml.py

@@ -967,12 +967,13 @@ def _whisper_args(settings):
         binary, "-m", str(model),
         "--inference-path", INFERENCE_PATH,
         # Whatever language the request does not name. api.py leaves the field
-        # out when the language is "auto", and the server's own default is
+        # out when the language is "auto", and the server's own language is
-        # English rather than detection.
+        # set here: "auto" makes whisper.cpp detect what it hears.
         "-l", "auto",
         # Stock phrases invented for near-silence come from non-speech tokens,
         # and verbose_json otherwise pays for a language probability sweep
-        # nothing here reads.
+        # nobody asked for. A request that wants the detected language switches
+        # that back on per request.
         "-sns", "-nlp",
     ]
     if int(settings["threads"]) > 0:

dikte/worker.py

@@ -37,7 +37,7 @@ _paste_lock = threading.Lock()
 
 class Pipeline(QObject):
     stage = pyqtSignal(str)          # human-readable progress line
-    finished = pyqtSignal(str, str, str)  # raw transcript, final text, warning
+    finished = pyqtSignal(str, str, str, str)  # raw, final text, warning, language
     failed = pyqtSignal(str)
     cancelled = pyqtSignal()
 
@@ -120,12 +120,24 @@ class Pipeline(QObject):
         try:
             self.stage.emit(t("Transcribing…"))
             target = conf.transcribe_target()
-            raw = api.transcribe(
+            # The spoken language is only knowable after the fact, and only the
-                target,
+            # local server says what it heard: auto mode asks it there, and
-                wav_path,
+            # every other run (a fixed language, or a hosted provider that
-                language=conf["language"],
+            # detects but stays silent) transcribes as before.
-                prompt=conf["transcribe_prompt"],
+            auto = conf["language"] == "auto"
-            )
+            if auto:
+                raw, detected = api.transcribe_detected(
+                    target, wav_path, language=conf["language"],
+                    prompt=conf["transcribe_prompt"],
+                )
+            else:
+                raw = api.transcribe(
+                    target,
+                    wav_path,
+                    language=conf["language"],
+                    prompt=conf["transcribe_prompt"],
+                )
+                detected = ""
 
             if conf["filter_hallucinations"] and vad.looks_like_hallucination(raw, duration):
                 self._discard(wav_path)
@@ -145,7 +157,7 @@ class Pipeline(QObject):
                 self.stage.emit(t("Cleaning up…"))
                 cleaned = True
                 try:
-                    text = cleanup.run(raw, conf, conf.cleanup_prompt())
+                    text = cleanup.run(raw, conf, conf.cleanup_prompt(speech=detected))
                 except api.ApiError as exc:
                     # Keep the transcript, but never let the failure pass unseen:
                     # a rejected key would otherwise look like working dictation.
@@ -183,6 +195,9 @@ class Pipeline(QObject):
                 "question": question,
                 "assistant": assistant.provider(conf) if ask else "",
                 "assistant_model": assistant.model(conf) if ask else "",
+                # The language the run actually spoke: the detected code, or the
+                # configured one when nothing was detected to replace it.
+                "speech_language": detected or conf["language"],
                 "raw": raw,
                 "text": text,
             }
@@ -221,7 +236,7 @@ class Pipeline(QObject):
                     time.sleep(0.35)
                     paste.copy_bytes(previous)
 
-            self.finished.emit(raw, text, warning)
+            self.finished.emit(raw, text, warning, detected)
 
         except assistant.Cancelled:
             self.cancelled.emit()

tests/test_config.py

@@ -259,6 +259,16 @@ class CleanupPrompt(DikteTest):
     def test_no_glossary_means_no_rule_about_one(self):
         self.assertEqual(cfg.Config().cleanup_prompt(), cfg.CLEANUP_PROMPT_EN)
 
+    def test_a_detected_turkish_recording_gets_the_turkish_prompt(self):
+        """Auto mode learns what was heard, and that decides the prompt rather
+        than the interface language."""
+        self.write_config({"ui_language": "en", "transcribe_prompt": "Paraşüt"})
+        conf = cfg.Config()
+        self.assertIn("KONUŞMACININ KULLANDIĞI İSİM VE TERİMLER",
+                      conf.cleanup_prompt(speech="tr"))
+        self.assertIn("NAMES AND TERMS THE SPEAKER USES",
+                      conf.cleanup_prompt(speech="de"))
+
     def test_subtitles_use_their_own_prompt(self):
         conf = cfg.Config()
         self.assertNotEqual(conf.cleanup_prompt(subtitles=True), conf.cleanup_prompt())

tests/test_worker.py

@@ -35,7 +35,7 @@ class Chain(DikteTest):
                   cleaned="Book it for Thursday.",
                   cleanup_error=None, answer=("Booked.", ""), rms=None,
                   clipboard=b"what was there before", paste_error=None,
-                  focus=None):
+                  detected="en", focus=None):
         pipeline = worker.Pipeline(self.conf)
         done, failures, stages, cancels = [], [], [], []
         pipeline.finished.connect(lambda *args: done.append(args))
@@ -45,14 +45,19 @@ class Chain(DikteTest):
 
         cleanup = (mock.Mock(side_effect=cleanup_error) if cleanup_error
                    else mock.Mock(return_value=cleaned))
+        # Auto mode takes the detection path; a fixed language the plain one.
+        # Both are mocked so the chain runs either way without a server.
+        behavior = {"side_effect": transcribe_error} if transcribe_error \
+            else {"return_value": transcript}
+        detect_behavior = {"side_effect": transcribe_error} if transcribe_error \
+            else {"return_value": (transcript, detected)}
         calls = {}
         # The chain reports its own failures on stderr, which a test run has no
         # use for.
         with contextlib.redirect_stderr(io.StringIO()), \
-                mock.patch.object(
+                mock.patch.object(api, "transcribe", **behavior) as tr, \
-                    api, "transcribe",
+                mock.patch.object(api, "transcribe_detected",
-                    **({"side_effect": transcribe_error} if transcribe_error
+                                  **detect_behavior) as tdet, \
-                       else {"return_value": transcript})) as tr, \
                 mock.patch.object(api, "cleanup", cleanup), \
                 mock.patch.object(assistant, "ask", return_value=answer) as ask_call, \
                 mock.patch.object(paste, "copy") as copy, \
@@ -62,7 +67,8 @@ class Chain(DikteTest):
                                   return_value=clipboard) as read_clipboard, \
                 mock.patch.object(worker.time, "sleep", lambda seconds: None):
             press.side_effect = paste_error
-            calls = {"transcribe": tr, "cleanup": cleanup, "ask": ask_call,
+            calls = {"transcribe": tr, "transcribe_detected": tdet,
+                     "cleanup": cleanup, "ask": ask_call,
                      "copy": copy, "copy_bytes": copy_bytes, "press": press,
                      "read_clipboard": read_clipboard}
             pipeline._work(self.wav, duration,
@@ -77,7 +83,8 @@ class Chain(DikteTest):
         run = self.run_chain()
         self.assertEqual(run["failures"], [])
         self.assertEqual(run["done"][0],
-                         ("uh, book it for Thursday", "Book it for Thursday.", ""))
+                         ("uh, book it for Thursday", "Book it for Thursday.",
+                          "", "en"))
         run["copy"].assert_called_once_with("Book it for Thursday.")
         run["press"].assert_called_once_with(self.conf["paste_shortcut"],
                                              focus=None)
@@ -129,7 +136,7 @@ class Chain(DikteTest):
         self.conf["restore_clipboard"] = True
         run = self.run_chain(paste_error=paste.PasteError("not trusted"))
         self.assertEqual(run["failures"], [])
-        raw, text, warning = run["done"][0]
+        raw, text, warning, _lang = run["done"][0]
         self.assertIn("not trusted", warning)
         run["copy_bytes"].assert_not_called()
 
@@ -182,17 +189,41 @@ class Chain(DikteTest):
         self.assertEqual(run["transcribe"].call_args.kwargs["language"], "tr")
         self.assertEqual(run["transcribe"].call_args.kwargs["prompt"], "Paraşüt")
 
+    def test_auto_mode_asks_for_the_detected_language_and_records_it(self):
+        run = self.run_chain(detected="tr")
+        told = run["transcribe_detected"].call_args.kwargs
+        self.assertEqual(told["language"], "auto")
+        self.assertEqual(cfg.read_history()[0]["speech_language"], "tr")
+        self.assertEqual(run["done"][0][3], "tr")
+        run["transcribe"].assert_not_called()
+
+    def test_the_detected_language_is_told_to_the_cleanup_prompt(self):
+        # The mock stands in for api.cleanup, which the cleanup module calls
+        # with (text, key, model, system_prompt, …); the prompt is the fourth.
+        self.conf["transcribe_prompt"] = "Paraşüt"
+        run = self.run_chain(detected="tr")
+        prompt = run["cleanup"].call_args.args[3]
+        # Turkish was detected, so the Turkish glossary rule is appended.
+        self.assertIn("KONUŞMACININ KULLANDIĞI İSİM VE TERİMLER", prompt)
+
+    def test_a_fixed_language_needs_no_detection(self):
+        self.conf["language"] = "en"
+        run = self.run_chain()
+        run["transcribe"].assert_called_once()
+        run["transcribe_detected"].assert_not_called()
+        self.assertEqual(cfg.read_history()[0]["speech_language"], "en")
+
     # ---- silence and stock phrases ----------------------------------------
 
     def test_room_tone_costs_no_api_call(self):
         run = self.run_chain(rms=[0.00001] * 60)
-        run["transcribe"].assert_not_called()
+        run["transcribe_detected"].assert_not_called()
         self.assertIn("No speech", run["failures"][0])
 
     def test_the_silence_check_can_be_switched_off(self):
         self.conf["skip_silent"] = False
         run = self.run_chain(rms=[0.00001] * 60)
-        run["transcribe"].assert_called_once()
+        run["transcribe_detected"].assert_called_once()
 
     def test_a_stock_phrase_from_a_short_clip_is_thrown_away(self):
         run = self.run_chain(duration=2.0, transcript="Altyazı M.K.")
@@ -208,7 +239,7 @@ class Chain(DikteTest):
 
     def test_a_failed_cleanup_still_pastes_the_transcript(self):
         run = self.run_chain(cleanup_error=api.ApiError("rate limited"))
-        _raw, text, warning = run["done"][0]
+        _raw, text, warning, _lang = run["done"][0]
         self.assertEqual(text, "uh, book it for Thursday")
         self.assertIn("rate limited", warning)
         run["copy"].assert_called_once_with("uh, book it for Thursday")
@@ -220,6 +251,9 @@ class Chain(DikteTest):
         self.assertEqual(cfg.read_history()[0]["cleanup_error"], "bad key")
 
     def test_a_failed_transcription_ends_the_run(self):
+        # This path mocks api.transcribe, so it wants
+        # the plain (fixed-language) transcription.
+        self.conf["language"] = "tr"
         pipeline = worker.Pipeline(self.conf)
         failures = []
         pipeline.failed.connect(failures.append)
@@ -231,6 +265,9 @@ class Chain(DikteTest):
         copy.assert_not_called()
 
     def test_a_clipboard_that_will_not_take_it(self):
+        # This path mocks api.transcribe, so it wants
+        # the plain (fixed-language) transcription.
+        self.conf["language"] = "tr"
         pipeline = worker.Pipeline(self.conf)
         failures = []
         pipeline.failed.connect(failures.append)
@@ -243,6 +280,9 @@ class Chain(DikteTest):
         self.assertIn("wl-copy", failures[0])
 
     def test_an_unexpected_error_is_reported_rather_than_swallowed(self):
+        # This path mocks api.transcribe, so it wants
+        # the plain (fixed-language) transcription.
+        self.conf["language"] = "tr"
         pipeline = worker.Pipeline(self.conf)
         failures = []
         pipeline.failed.connect(failures.append)
@@ -280,6 +320,9 @@ class Chain(DikteTest):
         run["press"].assert_not_called()
 
     def test_a_command_that_was_cancelled(self):
+        # This path mocks api.transcribe, so it wants
+        # the plain (fixed-language) transcription.
+        self.conf["language"] = "tr"
         pipeline = worker.Pipeline(self.conf)
         cancels = []
         pipeline.cancelled.connect(lambda: cancels.append(True))
@@ -289,6 +332,9 @@ class Chain(DikteTest):
         self.assertEqual(cancels, [True])
 
     def test_an_agent_that_is_not_installed(self):
+        # This path mocks api.transcribe, so it wants
+        # the plain (fixed-language) transcription.
+        self.conf["language"] = "tr"
         pipeline = worker.Pipeline(self.conf)
         failures = []
         pipeline.failed.connect(failures.append)