v3.37.1

Fixed

• Local STT with a blank API key works again on current Home Assistant. Home Assistant 2026.8 and 2026.9 ship a newer OpenAI SDK whose client refuses an empty API key with Missing credentials, so every utterance against a keyless local server (Speaches, for example) failed before a request was ever sent. The entity now hands the SDK a placeholder that never reaches the wire and strips the Authorization header from each request, so a keyless server still receives no bearer token at all — the behavior v3.37.0 intended. Entries that worked around this by entering a dummy key can leave the key blank again.
• The recommended local model no longer points at a repository that has gone private. Systran/faster-whisper-large-v3-turbo disappeared from Hugging Face, so a Speaches server could not download the v3.37.0 default. The recommended (and pre-filled) model is now deepdml/faster-whisper-large-v3-turbo-ct2, the public CTranslate2 conversion of the same weights. Existing entries keep whatever model name they have; a server that already holds the Systran files keeps serving them.
• Docs: the Speaches docker-compose example now pre-downloads the model — Speaches only serves models already on disk — keeps it resident between utterances, and notes the compute type needed on Pascal-generation GPUs.