In-process speech-to-text via Whisper (Bumblebee + Nx) — a local AgentSea.Voice.STT with no transcription API.

Build a serving once (it loads the model), then pass it per call:

serving = AgentSea.Voice.STT.Whisper.serving("openai/whisper-tiny")
{:ok, %{text: text}} = AgentSea.Voice.STT.Whisper.transcribe(audio, serving: serving)

The serving accepts what Bumblebee's Whisper serving accepts (raw PCM samples as an Nx.Tensor, or a file path when ffmpeg is available). The serving call is injectable (:run) so the transcription plumbing is testable without a model. Add :exla for real throughput (see AgentSea.Embedder.Bumblebee).