Files

T

tes cd07b52f8f Findings: Test 4a OpenAI Realtime one-shot voice round-trip outcome

2026-06-12 19:42:35 +00:00

33 KiB

Raw Blame History

Prototype-phase findings

Outcome of the two probes run on 2026-06-11 against the Pi at 192.168.50.115. Both passed their spec criteria. This document captures what we learned, what to carry forward, and what to revisit when designing the main assistant.

What we proved

Recording + playback works. python-sounddevice + soundfile at 16 kHz mono PCM16 round-trips cleanly through the USB Speaker Phone on the Pi. Level meter responds to speech; playback is recognisable.
Wakeword detection works. openwakeword's stock alexa model loads in ~0.6 s, detects reliably (user reports "runs perfectly"), false-positive rate is acceptable, CPU is well below saturation on a single core.

Environment facts to remember

Pi 4B, 4 GB, Debian 13 (trixie), Python 3.13.5, kernel 6.12.75+rpt-rpi-v8.
PipeWire is the audio server, but PortAudio sees only the raw ALSA hostapi — no PipeWire/Pulse hostapi.
The USB Speaker Phone (Anhui LISTENAI, USB f201:3220) is exposed as ALSA card 3. Native sample rate is 48 kHz only. Raw hw:3,0 rejects 16 kHz with PaErrorCode -9997. The ALSA plughw:CARD=Phone,DEV=0 device exists and does the resampling.
The same USB device is both mic and speaker. There is no separate output device.
Pi sudo for user pi requires a password (not passwordless).

Code patterns to carry forward

1. Force PortAudio through ALSA's `plug` plugin

Set this before importing sounddevice in every test:

import os
os.environ.setdefault("PA_ALSA_PLUGHW", "1")

Best done in one place in the main assistant — e.g. an assistant/audio.py that sets the env var and re-exports sounddevice. The ordering is fragile if every entry point has to remember it.

2. USB device discovery

find_usb_device() is duplicated byte-identically in both tests:

def find_usb_device() -> int:
    for idx, dev in enumerate(sd.query_devices()):
        name = dev["name"].lower()
        if "usb" in name and dev["max_input_channels"] >= 1:
            return idx
    print("USB audio device not found. Devices:", file=sys.stderr)
    print(sd.query_devices(), file=sys.stderr)
    sys.exit(1)

Lift to a shared module in the main assistant.

3. openwakeword needs `--no-deps` on Python 3.13

openwakeword==0.6.0 declares a hard dep on tflite-runtime>=2.8.0,<3 ; platform_system == "Linux". There is no tflite-runtime wheel for Python 3.13 and pip refuses to install.

Workaround: bin/deploy does two pip-install passes — a normal requirements.txt (sounddevice, numpy, onnxruntime, scipy, scikit-learn, tqdm, requests) and a requirements-nodeps.txt (openwakeword) installed with --no-deps. Keep this split for the main assistant until either the Pi drops to 3.12 or openwakeword stops hard-requiring tflite-runtime.

4. `openwakeword.utils.download_models(target_directory=None)` crashes

The function does os.path.exists(target_directory) with no None-handling. Call download_models() with no kwargs — the default is the package's resources/models dir, which is what you want.

Things to fix or redesign in the main assistant

A. Beep / TTS blocks the input stream

In Test 2, sd.play(beep) + sd.wait() runs in the same thread that pulls from the InputStream. During the ~200 ms beep, the InputStream's ring buffer fills up; the next stream.read() returns stale audio and may flag input overflow.

In the main assistant the same pattern would drop the first ~200 ms of the user's command right after the wakeword fires — the worst possible moment.

Pattern for the main assistant:

Callback-driven InputStream that pushes frames onto a queue.Queue.
A consumer thread (or asyncio task) reads from the queue and feeds the wakeword model / Realtime stream.
A separate, persistent OutputStream opened once at startup that the TTS / beep path pushes to.

B. The speaker is the mic

The USB Speaker Phone hears its own output. TTS playback will re-trigger the wakeword on its own speech.

Options:

Gate the detector during playback. Set a flag while TTS is playing; skip model.predict() while it's set. Simple and reliable.
Use hardware AEC if the device has it. The Anhui LISTENAI USB speakerphones often have on-board AEC — worth probing (try arecord while aplay is running, see if the input is clean).
Software AEC (webrtc-audio-processing). Adds complexity; only if hardware AEC isn't enough.

Recommendation: gate the detector during playback as the v1 behaviour. Revisit AEC only if barge-in becomes a requirement.

C. Custom wakeword (target: "hey assistant" or similar)

Test 2 used the stock alexa model as a probe. The final wakeword needs to be trained. openwakeword has a Colab notebook that generates synthetic training data and trains a model in ~1 hour. Plan this as a separate task before the main assistant ships.

D. Single-core CPU budget

The Pi 4 has 4 cores and the wakeword model uses well under 50% of one. The main assistant will add the Realtime WebSocket (small) plus opus encode/decode for the audio it sends and receives (non-trivial). Budget for: 1 core for audio I/O + wakeword, 1 core for Realtime / codec, headroom on the other two.

E. `bin/bootstrap` and the sudo password

bin/bootstrap currently does echo "$PI_PASS" | sudo -S to authenticate sudo on the Pi. That puts the password into the remote process's /proc/<pid>/cmdline briefly. Alternatives, in order of safety:

Enable passwordless sudo for pi on the Pi. Cleanest. Requires explicit security decision. Add a single file /etc/sudoers.d/010-pi-nopasswd containing pi ALL=(ALL) NOPASSWD: ALL.
Run bootstrap once manually, never script it. Avoids the issue entirely.
Switch to expect instead of sshpass + sudo -S. More code, similar trust model.

Pre-decide for the main assistant; the deploy/install scripts will hit this same wall.

F. No requirements lockfile

For prototypes, pinning major versions in requirements.txt is fine. For the long-lived main assistant, switch to pip-compile (or uv) to produce a lockfile so the Pi doesn't drift over time.

G. `~/assistant/` layout vs. rsync `--delete`

bin/deploy uses rsync -az --delete against ~/assistant/. That's correct for a code dir but will nuke anything else placed there. For the main assistant, split the layout:

~/assistant/code/ — rsync target with --delete.
~/assistant/state/ — recordings, logs, custom wakeword model, runtime data. Never rsynced.

H. systemd autostart

Not in scope for prototypes. The main assistant needs a user systemd service so it survives reboots and restarts on crash.

Cosmetic / known-warnings

onnxruntime prints two warnings on startup:
```
Failed to detect devices under "/sys/class/drm/card0"
Failed to detect devices under "/sys/class/drm/card1"
```
It's probing for GPU vendor info; the Pi's vc4 framebuffer doesn't expose a device/vendor sysfs file. Harmless. Suppressible via os.environ["ORT_DISABLE_GPU_DETECTION"] = "1" (or similar onnxruntime log-level setting) if it bothers you.

Open questions for the next design session

Wakeword phrase for the custom model — "hey assistant", "assistant", "computer", something else?
Conversation lifecycle — wakeword → open Realtime session → close when? On silence (VAD)? On user-spoken endword? On timeout?
Barge-in (user interrupting the assistant mid-reply) — in scope or not?
Where the OpenAI API key lives. (Env var? Pi-only file? Pulled from a secrets store at boot?)
Local-only fallback when the network is down — required behaviour or out of scope?
Multi-language? Or English only for v1?

C# probe outcome (2026-06-12 — Test 1 ported to C#)

The Python Test 1 (record 5 s → write WAV → play back) was re-implemented in C# / .NET 9 to validate that the Pi client can be written in the same language as the backend. The full probe lives at tests/01-record-play-cs/ with the deploy script at bin/probe-cs. All three pass criteria from the original spec held on hardware.

What we proved

PortAudioSharp2 1.0.6 is a usable .NET binding for PortAudio. Conventional API surface (PortAudio.Initialize, Stream ctor taking input/output StreamParameters, callback delegate with ref StreamCallbackTimeInfo) — matched our assumptions, no adjustments needed.
The NuGet package bundles libportaudio.so for linux-arm64 in its runtime folder, so a self-contained publish doesn't need to rely on the Pi's system libportaudio2. (Both are present; the bundled one wins.)
A single dotnet publish -c Release -r linux-arm64 --self-contained produces a ~70 MB directory that runs as-is on the Pi. No .NET install on the Pi. Deploy is dotnet publish → scp -r → ssh ./Probe, wrapped in bin/probe-cs.
The PA_ALSA_PLUGHW=1 trick from the original findings is still required in C# — but with a real twist (see next section).
Recording, RMS metering, WAV writing, and playback all behave identically to the Python prototype. Pass criteria 1 (recognisable playback), 2 (meter responsive), 3 (no over/underflow) all met.

Key gotcha — .NET env vars do NOT propagate to libc `getenv()` on Linux

Environment.SetEnvironmentVariable("PA_ALSA_PLUGHW", "1") updates .NET's process-table mirror but does not call libc setenv. libportaudio reads its config via getenv("PA_ALSA_PLUGHW") and sees nothing. ALSA then opens the USB Speaker Phone as hw:3,0 (native 48 kHz only), and the 16 kHz stream fails with paInvalidSampleRate.

Python's os.environ[x] = y calls setenv synchronously, which is why the Python probe never tripped on this.

Fix: P/Invoke libc setenv directly, alongside the managed call:

[DllImport("libc", EntryPoint = "setenv")]
static extern int setenv(string name, string value, int overwrite);

Environment.SetEnvironmentVariable("PA_ALSA_PLUGHW", "1"); // for .NET-side readers
setenv("PA_ALSA_PLUGHW", "1", 1);                           // for the C runtime view PortAudio reads

Implication for future C# probes and the eventual Pi client: any env knob a C library reads (PortAudio, onnxruntime, any P/Invoked native) must be set via libc setenv, not just Environment.SetEnvironmentVariable. Bake a small Libc.setenv helper into the eventual client.

Small .NET ergonomics

<AllowUnsafeBlocks>true</AllowUnsafeBlocks> is required in the csproj for the callback's Buffer.MemoryCopy over short*.
ImplicitUsings pulls in System.IO.Stream, which collides with PortAudioSharp.Stream. Disambiguate with using Stream = PortAudioSharp.Stream;.
Top-level statements don't auto-pick up the project's own namespace. Add using RecordPlayProbe; (or whatever the project's namespace is) so the entry-point file can reach internal helpers.

Open questions deferred to Test 2

No equivalent to openwakeword exists in .NET. The choices are (a) port the openwakeword inference pipeline to C# using Microsoft.ML.OnnxRuntime (mel extractor → Google speech embedder → keyword classifier, three ONNX files), (b) use Picovoice Porcupine's .NET SDK (commercial, key-gated), (c) Vosk + keyword spotting (heavier). Decision deferred to the Test 2 brainstorm.
The Libc.setenv workaround will likely be needed again for any onnxruntime env knob on the Pi.

Reference paths (C# probe)

Spec: docs/superpowers/specs/2026-06-12-record-play-csharp-probe-design.md
Plan: docs/superpowers/plans/2026-06-12-record-play-csharp-probe.md
Code: tests/01-record-play-cs/
Deploy: bin/probe-cs

C# wakeword probe outcome (2026-06-12 — Test 2 ported to C#)

The Python Test 2 (openwakeword "alexa" listener with beep on detect) was re-implemented in C# / .NET 9. openwakeword's streaming inference pipeline — three ONNX models chained: mel-spectrogram → Google speech embedding → "alexa" classifier — was ported to Microsoft.ML.OnnxRuntime. Probe lives at tests/02-wakeword-cs/; deploy script bin/probe-cs-2. All three pass criteria from the original spec held on hardware.

What we proved

Microsoft.ML.OnnxRuntime 1.26.0 runs on linux-arm64 from a self-contained dotnet publish. The NuGet package bundles libonnxruntime.so and libonnxruntime_providers_shared.so for the RID — no separate native package, no fallback needed.
The three openwakeword ONNX models (melspectrogram.onnx, embedding_model.onnx, alexa.onnx, vendored from openwakeword 0.6.0 on the Pi, SHA-256-pinned in commit 849ad47) load and chain correctly when fed real audio. Predict() loop returned plausible scores from the first run on hardware.
Two PortAudio streams on one USB device works. The persistent input stream stays open while a per-detection fire-and-forget output stream plays the beep. stream.Start() on the second stream didn't throw, the input callback kept ticking during the 200 ms beep — no input overflow during playback. This fixes the bug findings.md § A flagged in the Python probe (where sd.wait() stalled input for the beep duration).
Hardware test on the Pi: detection fires reliably on "alexa" at normal volume from ~1 m, audible beep on each. CPU peak 31% of one core, typical 17% — well under the 50% budget.

Key gotchas

The mel-spectrogram model's output shape was wrong in our initial port. The plan's verbatim code assumed (1, n_frames, 1, 32) based on a memory-derived shape. The actual shape is (1, 1, n_frames, 32) — n_frames lives at Dimensions[2], not Dimensions[1]. Caught during Task 3 implementation by the InputMetadata-vs-derived-shape comparison done in Task 1's sanity spike. Fix is documented in WakewordModel.cs. The lesson: always print the actual model metadata on first contact with a vendored ONNX file; do not infer shapes from upstream code without verification.
ORT 1.26.0's GPU-detection logger initializes BEFORE managed code can set env vars. Libc.setenv("ORT_LOGGING_LEVEL", "3", 1) alone does NOT suppress the Failed to detect devices under "/sys/class/drm/card0" warnings — by the time .NET code runs setenv, ORT's C++ device_discovery already logged. Workaround: explicit OrtEnv.CreateInstanceWithOptions with logLevel = ORT_LOGGING_LEVEL_ERROR BEFORE constructing any InferenceSession. The Libc.setenv call stays in place for consistency with the PortAudio pattern but does no work for ORT itself on this version. Bake the OrtEnv initialization into the eventual client.
Fire-and-forget output streams need careful lifetime management. Initial implementation had Task.Run waiting on a ManualResetEventSlim BEFORE stream.Start() ran. If Start() had thrown (the exact failure mode the dual-stream test exists to surface), the cleanup task would have blocked forever on an event that would never fire — silently leaking the stream and hanging the diagnostic. Fix: run Start() synchronously under try/catch first, dispose + rethrow on failure, only then launch the cleanup task. Also added a 2 s timeout on done.Wait() — PortAudio silently swallows callback exceptions, so without the timeout a faulting callback would hang cleanup the same way.
Vendored alexa classifier has a static batch dim. alexa.onnx declares input shape [1, 16, 96] (not [-1, 16, 96] like a typical dynamic-batch model). Other two models use dynamic batch. The shape-assertion code intentionally ignores dim 0 so static-1 and dynamic--1 both pass; the input tensor is created with batch=1 either way.
Workstation Ctrl-C needs ssh -t. bin/probe-cs-2 uses ssh -t (force pseudo-TTY) so workstation Ctrl-C propagates as SIGINT to the remote Probe, triggering the Console.CancelKeyPress graceful-shutdown path. Test 1 didn't need this because the probe self-terminated after 5 s; Test 2 listens until interrupted.
Inference latency was not measured. Frame size is 1280 samples (80 ms); if WakewordModel.Predict ever exceeds 80 ms, the bounded BlockingCollection (capacity 16) will start dropping frames with [status] consumer behind warnings. No such warnings were observed during the hardware test, but we did not instrument Predict timing. Add this when promoting to the main assistant.

Open questions for the main assistant

Custom wakeword model. Stock "alexa" works as a probe, but the shipped assistant needs a custom phrase (e.g. "hey assistant"). openwakeword has a Colab notebook for synthetic training; budget ~1 hour to produce a model that this same C# pipeline can load (it's a drop-in .onnx replacement for alexa.onnx).
Speaker-is-mic self-trigger. The 880 Hz 200 ms beep used here is far enough from speech-band that it doesn't re-trigger the classifier. Real TTS playback will. Spec § B recommends detector gating during playback as the v1 fix; revisit AEC only if barge-in becomes a requirement.
Inference timing instrumentation. Add per-Predict latency logging in the main assistant so consumer-behind warnings are diagnosable instead of opaque.

Reference paths (C# wakeword probe)

Spec: docs/superpowers/specs/2026-06-12-test-2-wakeword-csharp-design.md
Plan: docs/superpowers/plans/2026-06-12-test-2-wakeword-csharp.md
Code: tests/02-wakeword-cs/
Deploy: bin/probe-cs-2

C# full-cycle outcome (2026-06-12 — Test 3)

Test 3 stitched the C# Test 1 (record + play) and Test 2 (wakeword + beep) into one state machine: idle → "alexa" → beep → record 5 s → playback → idle. Probe lives at tests/03-full-cycle-cs/; deploy script bin/probe-cs-3. All seven spec pass criteria held on hardware on the first run-after-fix (see "Key gotchas" below).

What we proved

The single-input-stream + state-machine + per-cycle-output-streams shape works on this hardware. One PortAudio input stream open for the process lifetime feeds a BlockingCollection<short[]> (bounded 16); a single consumer thread dispatches each 80 ms frame through IDLE / BEEPING / RECORDING / PLAYBACK. No locks, no Interlocked — state is consumer-thread-local; the only cross-thread datum is a volatile bool in a PlaybackDoneFlag.
Detector gating during PLAYBACK prevents self-trigger. The recorded user audio comes back out the speaker and re-enters the input stream; because the consumer's PLAYBACK branch discards frames (never calls model.Predict), the playback never re-triggers detection.
WakewordModel.Reset() on PLAYBACK→IDLE re-arms the warmup guard correctly. Five lines (_rawRingFill = 0; _melRing.Clear(); _embRing.Clear(); _framesSeen = 0) restore the model to the same state it has just after construction; the existing WarmupFrames = 16 then suppresses the next ~1.3 s of Predict calls until the rings refill with fresh post-playback audio. Next "alexa" after playback ends is detected reliably and quickly (within target).
3+ cycles back-to-back, no degradation. Detection rate holds, CPU stays in IDLE-state range between cycles, no stuck states.
No [status] or [error] warnings during normal operation. Input doesn't overflow during PLAYBACK (the consumer keeps draining the queue into the discard branch), consumer doesn't fall behind, no fire-and-forget timeouts after the fix below.
Ctrl-C from the workstation cleanly exits the probe. Same ssh -t + Console.CancelKeyPress + cts.Cancel() pattern as Test 2.

Key gotchas

The fire-and-forget cleanup task's done.Wait timeout must scale to the buffer length, not be a flat constant. Test 2's FireAndForgetBeep used done.Wait(TimeSpan.FromSeconds(2)) to guard against silently-faulted PortAudio callbacks (PortAudio swallows callback exceptions). For a 200 ms beep, 2 s is generous; the cleanup task's wait always returns "set", then sleeps 200 ms and disposes. Test 3's FireAndForgetPlayback initially inherited the same flat 2 s timeout — but the recording playback runs for ~5 s, so the timeout fired mid-playback, the cleanup task called stream.Stop() 200 ms later, and the audio cut at ~2.2 s. First hardware run produced "[error] recording playback timed out — callback may have faulted" on every cycle. Fix: compute expectedSeconds = buf.Length / SampleRate and use TimeSpan.FromSeconds(expectedSeconds + 2.0) as the cleanup wait. Commit a7d408e.
Partial deploys are silently wrong: a self-contained .NET publish uses Probe (native launcher) + Probe.dll (managed code). Copying only Probe after a rebuild leaves the old .dll on the target, which the new launcher loads — so the binary on the Pi reflects code from the previous build. Always wipe-and-replace the whole publish directory (the pattern already in bin/probe-cs-3: rm -rf ~/probe-cs-3 && mkdir ~/probe-cs-3 && scp -r $PUBLISH_DIR/. pi:~/probe-cs-3/). Don't do incremental scp Probe.
PlaybackDoneFlag should be created per cycle, not reused. A theoretical late Set() from a delayed cleanup Task.Run (cycle N) could land on the shared flag while the consumer is in cycle N+1's PLAYBACK, causing a spurious immediate PLAYBACK→IDLE that skips playback. Allocating a fresh PlaybackDoneFlag() inside the RECORDING→PLAYBACK transition (and reading with playbackDone!.Consume() in PLAYBACK) sends the stale Set() to a now-unreachable object. Commit 070cc07. Not observed in practice but cheap to defend.
Log the cleanup-task's done.Wait timeout return. Without the log, a silently-faulted PortAudio callback produces "PLAYBACK→IDLE with no audio" with no diagnostic — exactly the failure mode the timeout was supposed to defend against. Both FireAndForgetBeep and FireAndForgetPlayback now log [error] ... timed out — callback may have faulted on !done.Wait(timeout). Commit 070cc07.

Implications for the main assistant

The whole shape — persistent input stream, single-consumer state machine, per-cycle output streams, fire-and-forget cleanup with timeout scaled to buffer length — is good to lift directly into the main Pi client. The main assistant adds: Opus encoding on the recorded buffer before sending to the Realtime WebSocket, an Opus decoder feeding the output stream during TTS playback, and replacing the local 5 s recording window with VAD-driven endpointing. None of those replace the state-machine shape; they slot in.
WakewordModel should be promoted out of tests/ into a real source tree (e.g. src/Audio/Wakeword/) when the main assistant starts. Test 3's copy is the live version (it has Reset()); Test 2's is frozen.
The "always wipe-and-replace the whole publish directory" deploy pattern needs to be the default in the main assistant's deploy script too. A partial scp of Probe is a footgun every time the .dll changes (i.e. every code change).

Reference paths (C# full-cycle probe)

Spec: docs/superpowers/specs/2026-06-12-test-3-full-cycle-csharp-design.md
Plan: docs/superpowers/plans/2026-06-12-test-3-full-cycle-csharp.md
Code: tests/03-full-cycle-cs/
Deploy: bin/probe-cs-3

Test 4a outcome (2026-06-12 — OpenAI Realtime one-shot voice round-trip in C#)

Test 4a opens a WebSocket to wss://api.openai.com/v1/realtime?model=gpt-realtime, plays a startup beep, streams mic audio to the server VAD endpoint, plays the assistant's audio reply back through the same USB Speaker Phone, and exits. Probe lives at tests/04a-realtime-oneshot-cs/; deploy script bin/probe-cs-4a. All 8 pass criteria met on hardware after a four-round iteration to match the GA Realtime API surface (the spec was written against beta-era docs from openai/openai-realtime-api-beta).

What we proved

ClientWebSocket + two background Task.Run loops (ReceiveLoop, UpstreamSendLoop) + bounded Channel<short[]> on each direction is the right shape on .NET 9 / linux-arm64 for a Realtime API client. No third-party WebSocket lib needed; System.Text.Json is fine for event parsing via JsonDocument.
One PortAudio input stream + one persistent output stream at 24 kHz mono Int16, both on the USB Speaker Phone, coexist cleanly. PortAudio plug-resamples 48↔24 at ALSA on this device. No client-side resampling.
Server VAD endpointing works out of the box. turn_detection: { type: "server_vad", threshold: 0.5, prefix_padding_ms: 300, silence_duration_ms: 500, create_response: true } (nested under session.audio.input) fires speech_started / speech_stopped correctly on real Pi mic input; server auto-commits the buffer and creates the response.
Round-trip latency is excellent. End-of-speech (server VAD speech_stopped) to first response audio sample arriving at the probe: 0.44 s measured on the test "hello" turn. Spec target was under 2 s; we beat it by 4×.
Streaming-length cleanup discipline works. Output callback signals doneFlag when CurrentChunk == null && Reader.Count == 0 && NoMoreDeltas (the NoMoreDeltas volatile bool is set on response.done after Writer.Complete(); reading it is cheaper than touching Reader.Completion). Probe exits ~immediately after response.done.
~/.openai_key precondition pattern in bin/probe-cs-4a works. Pre-publish check via sshpass ... 'test -f ~/.openai_key && test -r ~/.openai_key' catches a missing key at the deploy step rather than 30 s into publish + scp. Set with umask 077 && cat > ~/.openai_key <<KEY on the Pi.

Key gotchas — Realtime API GA migration

The spec was written against the older beta API surface (the openai/openai-realtime-api-beta library docs, which were the largest indexed source). The GA API rejected every beta-era piece of our session.update payload, surfacing one error per iteration. The full diff (commit 4cfb0d4):

No OpenAI-Beta: realtime=v1 request header. GA rejects it as beta_api_shape_disabled. Only Authorization: Bearer <key> is required for GA.
session.update payload is nested under session.audio.{input,output} instead of flat at the top of session. Beta had session.input_audio_format, session.output_audio_format, session.voice, session.turn_detection. GA has session.audio.input.format, session.audio.input.turn_detection, session.audio.output.format, session.audio.output.voice. Also session.type = "realtime" and session.model = "gpt-realtime" go inline at the top of session.
session.audio.{input,output}.format is now an OBJECT, not a bare string. Beta accepted "pcm16". GA requires { "type": "audio/pcm", "rate": 24000 }. The error was Invalid type for 'session.audio.input.format': expected an object, but got a string instead.
session.modalities is no longer accepted. Server rejected with Unknown parameter: 'session.modalities'. Dropped from our payload; server defaults are fine (audio + text output). The GA equivalent for constraining output is likely output_modalities (matches Responses API convention) — not needed for the probe.
Server events for response audio are renamed response.audio.* → response.output_audio.*. Beta had response.audio.delta / .done / response.audio_transcript.delta / .done. GA has response.output_audio.delta / .done / response.output_audio_transcript.delta / .done. This is the silent failure mode: the receive loop's switch default branch quietly logs [ws] ignored type=… instead of writing the decoded audio to the output channel. The probe logs "session pinned", VAD fires, response completes, transcript is empty, no audio is heard. Took one round of hardware verification to surface; trivially diagnosed via the ignored type=… log lines.
GA emits a stream of conversation/response framing events that don't exist in beta: conversation.item.added / .done, response.output_item.added / .done, response.content_part.added / .done, rate_limits.updated. Logged as [ws] ignored type=… initially; now silenced in our switch so they don't bury real signal.

The current Realtime model name gpt-realtime is still accepted (the dev docs mention gpt-realtime-2 as well; we didn't need to swap).

Other gotchas

Mid-startup WS error must route through cleanup. Initial implementation had a single try { … } finally { … } around main. When the receive loop processed an error event and called cts.Cancel() mid-SendSessionUpdate, the await ws.SendAsync(…, cts.Token) threw TaskCanceledException, which propagated past the finally as an unhandled exception (PortAudio + ws Dispose still ran from finally, but the process exited non-zero with a stack trace). Fix: add catch (OperationCanceledException) { /* expected on cancel */ } catch (Exception ex) { log; ExitCode = 7 } between the outer try and finally. Probe now logs the [error] WS error event: from the receive loop and exits cleanly. Commit 4cfb0d4.
Environment.Exit(N) skips the outer finally and leaks PortAudio, but is OK in the LoadApiKey path because it runs before PortAudio.Initialize(). The Environment.ExitCode = N; return; pattern in WS-connect / stream-Start failure paths preserves the finally correctly.
CA1416 warning on File.GetUnixFileMode is benign; the call is guarded by try { … } catch (PlatformNotSupportedException) { }. The build emits one warning consistently; no action needed.
ALSA shutdown noise on first run. Two alsa_snd_pcm_drop "failed" lines appeared from PortAudio's cleanup on the failed runs (cts.Cancel from an error event tore down both streams mid-init). Did NOT appear on the successful run. Looks like a PortAudio cleanup-on-abort quirk; only present on error paths.

Pass criteria

#	Criterion	Result
1	End-to-end round-trip works once	✅ Probe boots, beeps, user speaks, audible reply, exits 0
2	Response intelligible	✅ Clean speech, no clipping, no dropouts
3	Latency < 2 s	✅ 0.44 s end-of-speech to first delta
4	No `[status] upstream behind`	✅ None
5	No `[error] WS …` during normal flow	✅ None
6	No PortAudio `InputOverflow`	✅ None
7	Ctrl-C clean exit	✅ Same cleanup path as the successful run, which exited cleanly
8	Exits within ~500 ms of last audio sample	✅ `response.done` → `[ws] closing` → `bye` is instant

Implications for Test 4b

GA API surface is now proven. Test 4b inherits the migrated SendSessionUpdate payload, the renamed receive-loop switch cases, the silenced GA framing events, and the outer try/catch shape. Lift them directly.
The "spec written against beta docs" failure mode is a real risk for any future Realtime API work. When LLM-fetched docs (context7, training data) disagree with the live server, only the server is authoritative. Pattern: send a minimal session.update, observe the error response, iterate. Our four-round iteration took ~10 min total.
The "[ws] ignored type=…" log line on the default branch is the diagnostic that surfaced the renames. Keep it (or an equivalent) in 4b's receive loop. Without it, missing audio could have been silently attributed to a thousand other causes.
Server VAD parameters (threshold: 0.5, silence_duration_ms: 500, prefix_padding_ms: 300) work well on this hardware — natural-feeling end-of-turn after ~500 ms of silence, no clipped openings of user speech. Lift unchanged into 4b.
Environment.Exit vs Environment.ExitCode = N; return; — the latter is the correct pattern for failure paths that need PortAudio + WS cleanup. 4b will have more failure paths (wakeword model init, audio rate conversion, etc.); use the ExitCode pattern consistently.
The startup beep pattern (queue beep onto the persistent output channel before _micArmed = true) works — beep does not get captured and shipped to OpenAI. Carry into 4b.
The 16↔24 kHz coexistence question is the one major thing 4a did not address. 4a is 24 kHz end-to-end. 4b must deal with the wakeword model requiring 16 kHz input while OpenAI requires 24 kHz pcm16 — either a second input stream, an in-process resampler, or open one stream at 24 kHz and downsample for the wakeword. The spec/plan for 4b will need to pick.

Reference paths (Test 4a)

Spec: docs/superpowers/specs/2026-06-12-test-4a-openai-realtime-design.md
Plan: docs/superpowers/plans/2026-06-12-test-4a-openai-realtime.md
Code: tests/04a-realtime-oneshot-cs/
Deploy: bin/probe-cs-4a
GA Realtime API headers / URL / session.update shape (as discovered against the live server, 2026-06-12):
- wss://api.openai.com/v1/realtime?model=gpt-realtime
- Header Authorization: Bearer <key> (NO OpenAI-Beta)
- session.update payload: { type: "session.update", session: { type: "realtime", model: "gpt-realtime", instructions: "...", audio: { input: { format: { type: "audio/pcm", rate: 24000 }, turn_detection: { type: "server_vad", threshold, prefix_padding_ms, silence_duration_ms, create_response } }, output: { format: { type: "audio/pcm", rate: 24000 }, voice: "alloy" } } } }
- Server events to handle: session.created, session.updated, input_audio_buffer.speech_started, input_audio_buffer.speech_stopped, input_audio_buffer.committed, response.created, response.output_audio_transcript.delta, response.output_audio.delta, response.output_audio.done, response.output_audio_transcript.done, response.done, error
- Server events safe to ignore for a probe: conversation.item.added, conversation.item.done, response.output_item.added, response.output_item.done, response.content_part.added, response.content_part.done, rate_limits.updated

Reference paths

Spec: docs/superpowers/specs/2026-06-11-voice-assistant-prototypes-design.md
Plan: docs/superpowers/plans/2026-06-11-voice-assistant-prototypes.md
Test 1: tests/01-record-play/main.py
Test 2: tests/02-wakeword/main.py
Deploy helper: bin/deploy
Pi bootstrap: bin/bootstrap
Pi access: pi@192.168.50.115, password assistant (see CLAUDE.md)

33 KiB Raw Blame History Unescape Escape

Prototype-phase findings

What we proved

Environment facts to remember

Code patterns to carry forward

1. Force PortAudio through ALSA's plug plugin

2. USB device discovery

3. openwakeword needs --no-deps on Python 3.13

4. openwakeword.utils.download_models(target_directory=None) crashes

Things to fix or redesign in the main assistant

A. Beep / TTS blocks the input stream

B. The speaker is the mic

C. Custom wakeword (target: "hey assistant" or similar)

D. Single-core CPU budget

E. bin/bootstrap and the sudo password

F. No requirements lockfile

G. ~/assistant/ layout vs. rsync --delete

H. systemd autostart

Cosmetic / known-warnings

Open questions for the next design session

C# probe outcome (2026-06-12 — Test 1 ported to C#)

What we proved

Key gotcha — .NET env vars do NOT propagate to libc getenv() on Linux

Small .NET ergonomics

Open questions deferred to Test 2

Reference paths (C# probe)

C# wakeword probe outcome (2026-06-12 — Test 2 ported to C#)

What we proved

Key gotchas

Open questions for the main assistant

Reference paths (C# wakeword probe)

C# full-cycle outcome (2026-06-12 — Test 3)

What we proved

Key gotchas

Implications for the main assistant

Reference paths (C# full-cycle probe)

Test 4a outcome (2026-06-12 — OpenAI Realtime one-shot voice round-trip in C#)

What we proved

Key gotchas — Realtime API GA migration

Other gotchas

Pass criteria

Implications for Test 4b

Reference paths (Test 4a)

Reference paths

33 KiB

Raw Blame History

1. Force PortAudio through ALSA's `plug` plugin

3. openwakeword needs `--no-deps` on Python 3.13

4. `openwakeword.utils.download_models(target_directory=None)` crashes

E. `bin/bootstrap` and the sudo password

G. `~/assistant/` layout vs. rsync `--delete`

Key gotcha — .NET env vars do NOT propagate to libc `getenv()` on Linux