feat(openai): add GPT Transcribe audio support

src/celeste/modalities/audio/providers/openai/models.py

@@ -82,6 +82,14 @@
             AudioParameter.OUTPUT_FORMAT: Choice(options=_RESPONSE_FORMAT_OPTIONS),
         },
     ),
+    Model(
+        id="gpt-transcribe",
+        provider=Provider.OPENAI,
+        display_name="GPT Transcribe",
+        streaming=False,
+        operations={Modality.AUDIO: {Operation.TRANSCRIBE}},
+        parameter_constraints=_TRANSCRIBE_CONSTRAINTS,
+    ),
     Model(
         id="gpt-4o-mini-transcribe",
         provider=Provider.OPENAI,

src/celeste/providers/openai/audio/parameters.py

@@ -91,7 +91,8 @@ def map(
         validated_value = self._validate_value(value, model)
         if validated_value is None:
             return request
-        request["language"] = validated_value
+        field = "languages[]" if model.id == "gpt-transcribe" else "language"
+        request[field] = validated_value
         return request
 
 

tests/unit_tests/test_openai_audio_transcribe.py

@@ -8,15 +8,16 @@
 from celeste.mime_types import AudioMimeType
 from celeste.modalities.audio.io import AudioInput
 from celeste.modalities.audio.providers.openai.client import OpenAIAudioClient
+from celeste.modalities.audio.providers.openai.models import MODELS
 from celeste.models import Model
 
 
-def _client() -> OpenAIAudioClient:
+def _client(model_id: str = "gpt-4o-mini-transcribe") -> OpenAIAudioClient:
     return OpenAIAudioClient(
         model=Model(
-            id="gpt-4o-mini-transcribe",
+            id=model_id,
             provider=Provider.OPENAI,
-            display_name="GPT-4o Mini Transcribe",
+            display_name=model_id,
             operations={Modality.AUDIO: {Operation.TRANSCRIBE}},
         ),
         auth=AuthHeader(secret=SecretStr("test")),
@@ -40,6 +41,21 @@ def test_build_request_adds_model_and_default_response_format() -> None:
     assert request["language"] == "en"
 
 
+def test_gpt_transcribe_uses_plural_language_field() -> None:
+    audio = AudioArtifact(data=b"fake-audio", mime_type=AudioMimeType.WAV)
+    request = _client("gpt-transcribe")._build_request(
+        AudioInput(audio=audio), language="en"
+    )
+    assert request["languages[]"] == "en"
+    assert "language" not in request
+
+
+def test_gpt_transcribe_is_registered_for_file_transcription() -> None:
+    model = next(model for model in MODELS if model.id == "gpt-transcribe")
+    assert model.operations == {Modality.AUDIO: {Operation.TRANSCRIBE}}
+    assert model.streaming is False
+
+
 def test_parse_content_returns_transcript_text() -> None:
     assert _client()._parse_content({"text": "hello world"}) == "hello world"