Voice Realtime proof: official openai SDK client <-> anvil Realtime server

STATUS: LIVE CAPTURED. The proof harness is scripts/voice/realtime_sdk_client_demo.py. A passing --capture run appends a session row below and writes the artifact bundle under the temp anvil-voice-captures/ directory unless an explicit prefix is supplied.

Related: docs/findings/2026-07-04-hf-speech-to-speech-review.md s5 (the Realtime server, "verified with the official OpenAI Python SDK as client") · anvil_serving/voice/realtime/{app,ws,pool,service,events}.py · scripts/voice/realtime_sdk_client_demo.py

What This Proves

The T014 acceptance command:

python scripts/voice/realtime_sdk_client_demo.py --capture

loads the primary-node manifest by default, starts the same package-owned Realtime server wiring used by anvil-serving voice run, connects with the official openai Python SDK Realtime client, sends synthesized PCM speech through input_audio_buffer.append/commit, renders live input transcript events, cancels the first response with response.cancel, sends a second spoken turn, and requires a completed assistant audio response after the interruption. Capture validation fails if any output for the cancelled response arrives after the client sends response.cancel. Text-enabled runs additionally require assistant transcript deltas whose concatenation equals one correlated terminal transcript, with that terminal arriving before response.done.

Known Caveats

1. The default proof uses configured TTS to synthesize the user's two spoken inputs. That is intentional: it keeps the acceptance command automated and still exercises audio input, STT, LLM, TTS, Realtime protocol events, and the official SDK WebSocket path end to end.
2. The server remains a deliberately partial Realtime implementation: server-VAD path, no item delete/truncate, no granular content-part streaming, and loopback-only unauthenticated default unless realtime_token_env is configured for non-loopback binds.
3. The proof logs automated audio/transcript/latency evidence. It is not a subjective speech-quality review.

Session Log

timestamp (UTC)	turn kind	barge-in tested?	transcript(s)	events captured	audio bytes	completed TTFA / latency ms	proof bundle
2026-07-06T06:09:48Z	audio/audio	yes	Please count slowly from one to twenty so I can interrupt you.; Interrupting you now, please answer briefly how many countries are in Africa.	50	75800	303.35 / 588.02	C:\Users\operator\AppData\Local\Temp\anvil-voice-captures\realtime-sdk-20260706T060944Z.session.json
2026-07-22T18:14:43Z	audio/audio	yes	Please count slowly from 1 to 20, so I can interrupt you.; Interrupting you now, please answer briefly, how many countries are in Africa?	40	41146	253.48 / 414.21	C:\Users\operator\AppData\Local\Temp\anvil-voice-captures\realtime-sdk-20260722T181440Z.session.json

Decision

T014 is satisfied by the 2026-07-06T06:09:48Z run: the official SDK connected, audio input produced live transcripts, response.cancel interrupted resp_1 with no post-cancel output events, and resp_2 completed with assistant audio. The 2026-07-22T18:14:43Z run additionally proves issue #281's assistant-text contract against the live router/STT/TTS topology: resp_2 streamed the assistant transcript Fifty-five., the terminal matched its deltas and preceded response.done, 41,146 output-audio bytes were captured, and the harness reported no acceptance errors. T017 should independently verify the artifact bundle and delivery branch.