pulse-engine 0.2.35__tar.gz → 0.2.36__tar.gz

{pulse_engine-0.2.35 → pulse_engine-0.2.36}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: pulse-engine
-Version: 0.2.35
+Version: 0.2.36
 Summary: Pulse Engine — Hybrid framework for building Pulse products
 Author: Pulse Team
 Requires-Python: >=3.11,<3.13

{pulse_engine-0.2.35 → pulse_engine-0.2.36}/pyproject.toml

@@ -1,6 +1,6 @@
 [tool.poetry]
 name = "pulse-engine"
-version = "0.2.35"
+version = "0.2.36"
 description = "Pulse Engine — Hybrid framework for building Pulse products"
 authors = ["Pulse Team"]
 readme = "README.md"

{pulse_engine-0.2.35 → pulse_engine-0.2.36}/src/pulse_engine/adapters/transcription/providers/deepgram.py

@@ -45,9 +45,18 @@ class DeepgramProvider(TranscriptionProvider):
 
         results = response.get("results", {})
         channels = results.get("channels", [{}])
-        alt = (channels[0].get("alternatives", [{}]) or [{}])[0]
+        channel = channels[0] if channels else {}
+        alt = (channel.get("alternatives", [{}]) or [{}])[0]
         transcript = str(alt.get("transcript", ""))
-        detected = str(alt.get("language") or language)
+        # With detect_language the code is reported on the channel, not the
+        # alternative. Falling back to the requested value would echo "auto"
+        # back at the caller and discard the one useful output of detection —
+        # which is what lets a caller assert the source language was plausible.
+        detected = str(
+            alt.get("language")
+            or channel.get("detected_language")
+            or language
+        )
 
         utterances = results.get("utterances") or []
         if utterances:
@@ -75,12 +84,26 @@ class DeepgramProvider(TranscriptionProvider):
         client = DeepgramClient(self._api_key)
         with open(audio_path, "rb") as fobj:
             payload = {"buffer": fobj.read()}
-        options = PrerecordedOptions(
-            model=model or _DEEPGRAM_MODEL,
-            smart_format=True,
-            utterances=True,
-            language=None if language in ("", "auto") else language,
-        )
+
+        # "auto" must ask Deepgram to DETECT the language, not merely omit it.
+        # Omitting `language` makes Deepgram default to English, so a Hindi
+        # broadcast comes back as English phonetic gibberish rather than failing:
+        # a 47-minute programme returned 390 characters instead of 39,042, which
+        # downstream consumers cannot distinguish from a genuinely quiet source.
+        # `detect_language=True` returns the right transcript and reports the
+        # detected code on the alternative.
+        detect = language in ("", "auto") or language is None
+        option_kwargs: dict[str, Any] = {
+            "model": model or _DEEPGRAM_MODEL,
+            "smart_format": True,
+            "utterances": True,
+        }
+        if detect:
+            option_kwargs["detect_language"] = True
+        else:
+            option_kwargs["language"] = language
+
+        options = PrerecordedOptions(**option_kwargs)
         resp = client.listen.rest.v("1").transcribe_file(payload, options)
         result: dict[str, Any] = resp.to_dict()
         return result