Fix seed field not bridged in SpeechRequest OpenAI conversion

[winklemad]

No linked issue — found by reading the code.

The bug

SpeechRequest uses the field name random_seed, while the OpenAI request format uses seed. Neither converter bridges the two:

• to_openai() returns the raw model_dump, so it emits random_seed (a mistral-common-internal name) instead of seed.
• from_openai() keeps only keys that are model fields, so an OpenAI request's seed is never mapped to random_seed and is silently dropped:

from mistral_common.protocol.speech.request import SpeechRequest

SpeechRequest.from_openai({"input": "hi", "voice": "female", "seed": 123}).random_seed
# None   -> the seed provided in the OpenAI request is lost

SpeechRequest(input="hi", voice="female", random_seed=123).to_openai()["seed"]
# KeyError: 'seed'   -> the key is emitted as "random_seed" instead

Smoking gun

The sibling TranscriptionRequest already bridges the two names in both directions:

# TranscriptionRequest.to_openai
openai_request["seed"] = openai_request.pop("random_seed")
# TranscriptionRequest.from_openai
seed = openai_request.get("seed")
...
converted_dict["random_seed"] = seed

SpeechRequest does neither, so it is inconsistent with its sibling.

Fix

Translate random_seed ⇄ seed in both SpeechRequest.to_openai and SpeechRequest.from_openai, mirroring TranscriptionRequest. A seed now survives the round trip and an OpenAI-style request with seed is honored.

Testing

Added test_convert_speech_request_seed, asserting to_openai() emits seed (not random_seed) and from_openai({"seed": ...}) sets random_seed. It fails before the change (KeyError: 'seed') and passes after; the full tests/test_converters.py suite passes (104 passed). ruff check / ruff format --check are clean.

[juliendenize]

Thanks for the contribution, made a minor update and will merge :)