Speech-to-Text streaming is gRPC only, 25 KB per message, config first

Use the gRPC client for streaming; do not look for a REST streaming endpoint. Chunk audio so no single stream message exceeds 25 KB. Send the config in the first request and audio after. If you are just transcribing a file rather than live input, use synchronous or async recognize instead of streaming.

Context: Official docs (Cloud Speech-to-Text, streaming): documents the streaming gotchas agents hit. Streaming recognition is available through gRPC only; there is no REST streaming endpoint. Each message in the stream, including the first, is limited to 25 KB of audio, and exceeding it throws an error. The first StreamingRecognize request must carry the config with the recognizer and streaming settings; audio goes in the subsequent requests. For streaming a local file the docs note synchronous recognition is usually simpler.