Strip mistral-specific fields in SpeechRequest.to_openai

[jaideeppyne]

What

SpeechRequest.to_openai() leaks mistral-internal fields (id, max_tokens) into the OpenAI-compatible request payload, while its sibling TranscriptionRequest.to_openai() deliberately strips them.

SpeechRequest(id="req-123", input="hi", voice="female", model="tts-1").to_openai()
# {'id': 'req-123', 'max_tokens': None, 'temperature': 0.7, 'top_p': 1.0, 'model': ..., 'input': ..., 'voice': ..., 'seed': ...}
#  ^^^^ leaked                ^^^^^^^^^^ leaked

Why

SpeechRequest.to_openai does self.model_dump(...) and returns it directly, so every internal field ends up in the OpenAI request. TranscriptionRequest.to_openai already handles exactly this:

# transcription/request.py
# remove mistral-specific
# TODO: revisit which fields to expose in the OpenAI format
default_exclude = ("id", "max_tokens", "strict_audio_validation", "streaming")
default_exclude += exclude
for k in default_exclude:
    openai_request.pop(k, None)

So the two to_openai conversions are inconsistent, and a SpeechRequest's internal id / max_tokens are sent to the OpenAI endpoint.

How

Apply the same exclusion in SpeechRequest.to_openai, and add an exclude parameter for parity with TranscriptionRequest.to_openai:

-    def to_openai(self, **kwargs: Any) -> dict[str, Any]:
+    def to_openai(self, exclude: tuple = (), **kwargs: Any) -> dict[str, Any]:
         ...
         openai_request["seed"] = openai_request.pop("random_seed")
         openai_request.update(kwargs)
+
+        # remove mistral-specific fields, mirroring TranscriptionRequest.to_openai
+        # TODO: revisit which fields to expose in the OpenAI format
+        default_exclude = ("id", "max_tokens", "strict_audio_validation", "streaming")
+        default_exclude += exclude
+        for k in default_exclude:
+            openai_request.pop(k, None)

(I kept the exclusion list identical to TranscriptionRequest so the two stay in sync; strict_audio_validation/streaming simply aren't present on SpeechRequest and pop(..., None) is a no-op for them.)

Tests

Added test_speech_to_openai_excludes_mistral_specific_fields: asserts id/max_tokens are stripped, the OpenAI fields (model/input/voice) are preserved, and the exclude argument is honored. It fails on main (id is present) and passes with this change; the speech to_openai tests stay green (10 passed). ruff check is clean.

---

Disclosure: this change was prepared with AI assistance and reviewed/verified by me before submission.