Anvil Voice latency baseline for OpenClaw Talk (2026-07-07)¶

Publication redaction: The operator-specific network hostname was replaced with a public-safe placeholder. The measurements and event ordering are unchanged.

STATUS: BASELINE CAPTURED for voice-latency-model-ab:T001. The active Mini-local audio profile routes STT and TTS through loopback MLX Audio serves on Companion Node, and routes the LLM turn to the Primary Node router as fast-local / qwen36-27b.

Topology addendum (2026-07-08). This remains historical evidence for the optional Mini-local audio mode. It is no longer the reference OpenClaw Talk or candidate benchmark topology. Companion Node's 16 GB RAM is reserved for OpenClaw Gateway, Anvil Voice Realtime/proxy, Claude Code, and Codex during normal validation; use Dark-host or Mini-proxied audio for future candidate A/B.

This note records the current baseline before candidate LLM A/B testing. It is not a promotion recommendation and does not change production routing.

Environment¶

Field	Value
Capture date	2026-07-07
Gateway / voice host	Companion Node (`mini-host.example`)
Benchmark checkout	`/Users/operator/anvil-serving-t007`
Benchmark checkout revision	`6d024c2480b9f935b06ae26213ca7cfb69568485`
Checkout status	detached `HEAD`; unrelated dirty file `docs/findings/2026-07-voice-16gb-mini.md`
Voice manifest	`examples/voice/openclaw-anvil-voice.toml`
Voice profile	`mini-audio`
Command path	`/Users/operator/anvil-serving-t007/.venv/bin/anvil-serving`
Realtime endpoint	`ws://127.0.0.1:8765/v1/realtime`
LLM endpoint	`http://100.64.0.10:8000/v1`
LLM request model	`fast-local`
Router auth env	`ANVIL_ROUTER_TOKEN`
STT endpoint	`http://127.0.0.1:30010/v1`
TTS endpoint	`http://127.0.0.1:30011/v1`

The non-interactive SSH shell did not load ANVIL_ROUTER_TOKEN; the already running Realtime process did have it. The benchmark command used that token in-memory as an environment variable and did not print or write the token.

Active Services¶

Mini processes at capture time:

PID	Role	Command
14711	STT	`.venv/bin/python -m mlx_audio.server --host 127.0.0.1 --port 30010 --log-dir /tmp/anvil-voice-mini`
14712	TTS	`.venv/bin/python -m mlx_audio.server --host 127.0.0.1 --port 30011 --log-dir /tmp/anvil-voice-mini`
61505	Realtime	`.venv/bin/anvil-serving voice run --config examples/voice/openclaw-anvil-voice.toml --profile mini-audio`

Port checks:

Port	Status
`127.0.0.1:8765`	open
`127.0.0.1:30010`	open
`127.0.0.1:30011`	open

Endpoint model identity:

Stage	Model
STT	`mlx-community/parakeet-tdt-0.6b-v3`
TTS	`mlx-community/Kokoro-82M-bf16`

Route Proof¶

Decision-only /v1/route probe:

Field	Value
URL	`http://100.64.0.10:8000/v1/route`
HTTP status	`200`
Request model	`fast-local`
Provider	`fast-local`
Served model	`qwen36-27b`
Tier	`local`
Work class	`chat-fast`
Validation errors	none

Router response summary:

{
  "confidence": 0.9,
  "model": "qwen36-27b",
  "provider": "fast-local",
  "reason": "pinned; quality gate: on",
  "tier": "local",
  "work_class": "chat-fast"
}

Benchmark Command¶

The baseline was captured on Companion Node:

cd ~/anvil-serving-t007
.venv/bin/anvil-serving voice benchmark \
  --config examples/voice/openclaw-anvil-voice.toml \
  --profile mini-audio

The same command run from the Windows checkout was a negative control: it could not reach 127.0.0.1:30010 because that loopback address is Mini-local, not the operator workstation.

Three-Run Baseline¶

The benchmark uses a generated STT/TTS round trip. In this capture the STT hypothesis was empty for all three runs (stt_wer = 1.0), so the STT quality metric is not useful here. The stage timings still exercise the configured STT, LLM, and TTS HTTP paths and are valid for baseline latency attribution.

Run	TTFA ms	Turn latency ms	STT ms	LLM ms	TTS ms	TTS RTF	TTS requests	Output bytes
1	3041.72	3299.18	329.72	421.54	2547.93	0.8025	2	152400
2	1014.89	1230.51	500.21	364.88	365.42	0.1151	2	152400
3	465.04	627.60	99.15	229.42	299.02	0.0942	2	152400

Median across all three runs:

Metric	Median
TTFA	1014.89 ms
Turn latency	1230.51 ms
STT	329.72 ms
LLM	364.88 ms
TTS	365.42 ms

Steady-state view using runs 2-3 only:

Metric	Average
TTFA	739.97 ms
Turn latency	929.06 ms
STT	299.68 ms
LLM	297.15 ms
TTS	332.22 ms

Run 1 is a warm-up outlier dominated by TTS. Runs 2-3 are more representative of the warmed Mini-local audio path. Compared with prior live Realtime Talk timings, the benchmark path is a useful smoke measurement but not a substitute for a full iOS/OpenClaw conversational turn.

Interpretation¶

For this CLI benchmark, first-audio latency is the sum of:

STT HTTP round trip and decode,
LLM response generation through the Dark router,
the first TTS audio response.

The warmed benchmark split is roughly balanced between STT, LLM, and TTS, with LLM around 229-365 ms in runs 2-3. The earlier live Realtime Talk traces were different: STT was much lower, first LLM output was the largest perceived latency contributor, and TTS controlled the full-turn tail. That difference is why the model A/B should include both CLI benchmark evidence and live Talk validation.

Verification¶

Local unit tests passed from the Windows checkout on the task branch:

python -m pytest tests/voice/test_voice_benchmark.py tests/voice/test_voice_cli.py -q
67 passed in 3.98s

The Mini benchmark command passed three times after providing the router token through ANVIL_ROUTER_TOKEN.

Evidence Rerun¶

The T001 evidence gate was rerun on 2026-07-08 after reopening the task because the earlier verification commands were not captured as typed Anvil evidence. The rerun used the same Mini checkout and command, with the router token copied in-memory from the already running Realtime process and not printed or written.

Run	TTFA ms	Turn latency ms	STT ms	LLM ms	TTS ms	TTS RTF	TTS requests	Output bytes
1	774.13	965.70	107.45	447.99	410.26	0.1292	2	152400
2	611.29	789.06	106.28	356.82	325.95	0.1027	2	152400
3	586.71	762.12	94.38	353.88	313.87	0.0989	2	152400

Rerun median:

Metric	Median
TTFA	611.29 ms
Turn latency	789.06 ms
STT	106.28 ms
LLM	356.82 ms
TTS	325.95 ms

The rerun confirms the warmed path remains under one second TTFA in this CLI benchmark. The first-audio latency and total turn latency are still separate numbers: TTFA is the user-perceived first audio boundary, while turn latency also includes complete TTS output generation.

Follow-Up¶

Use this as optional Mini-local audio context only, not as the reference qwen36-27b baseline for candidate LLM A/B.
Run future candidate A/B with Mini model-free and audio fixed to dark-audio or mini-dark-audio-proxy.
For model promotion, require live Talk validation with memory, tool calls, session transcript delivery, and no duplicate-message spam.