Blame for docs/LATENCY.md

# LATENCY — benchmarking methods & budget tables

## MEASURED — MVP slice, 2026-10-04, cpu-upgrade (8 vCPU cgroup quota, CPU-only)

End-to-end run `ml-intern-freestream-e2e-speculative` (final: job `6ac2e756404719ba37651973`,
two turns, speculative dispatch ON, ASR tiny.en @1 thread, clause-level chunking):
livekit-server v1.13.7 dev mode in-container, bot joined via the real CLI,
test client asked spoken questions, timings measured at the receiving client.

| Metric | Measured | Target | Verdict |
|---|---|---|---|
| **t_perceived, turn 1** (speculative + cached disclosure) | **2.94s** | ≤ 4.0s (win condition) | **PASS** 🔥 |
| t_perceived, turn 2 (speculative, dynamic reply) | 5.79s | — | both turns via speculative path |
| t_perceived, roleplay mode (LLM agent in the loop) | 7.1–7.4s | relaxed (12s) | HF router generation dominates |
| t_ready (CLI join → bot ready, incl. warmup) | 12.5s | — | — |
| ASR finalize (tiny.en int8, 1 thread) | 1.25–1.46s | 400–700ms | over; standalone RTF 0.25 |
| First-chunk synth, turn 2 | 4.09s | 200–400ms | clause-capped (was 7.47s) |
| bot→client transport gap | 0.04s (t1) / 0.45s (t2) | — | negligible |
| cgroup CPU throttling during turns | **0 periods, 0ms** | — | contention myth settled |
| Mouth movement | 13/13 checks incl. both turns | >0 | PASS |
| Replay logs | both hash chains verify | valid | PASS |

Findings from the profiling campaign:

1. **cgroup oversubscription is the #1 CPU pitfall**: the container reports
   `os.cpu_count()=64` under an 8-core quota. Whisper tiny at 8 threads
   measured **39.5s** vs **1.16s** at 1 thread (same transcript). Thread
   counts are pinned low (FREESTREAM_WHISPER_THREADS).
2. **The fixed consent disclosure is pre-synthesizable** (CONSENT.md #2):
   caching it at session start removes Piper entirely from the first turn's
   audio path.
3. **bot→client transport is nearly free (0.04–0.5s)** — measured with
   absolute cross-process timestamps.
4. **Speculative dispatch works** (concept mitigation #1): dispatching ASR +
   agent + first-chunk synth at ~300ms of silence (with cancellation on
   speech resume and dedup against the final event) took turn 1 from 3.12s
   to **2.94s**. Both measured turns completed via the speculative path; the
   cancellation path is exercised by the segmenter's "resumed" event.
5. **The contention question is settled: zero cgroup throttling**
   (cpu.stat deltas across turns: 0 periods, 0ms). The in-context vs
   standalone RTF gap is not quota starvation — it is cold-start effects
   (page cache, allocator warmup) and benign thread scheduling. The ORT
   thread-cap countermeasure was not needed (knob FREESTREAM_ORT_THREADS
   remains for other hosts).
6. **First-chunk synth length was the real turn-2+ bottleneck**: a 99-char
   first sentence cost 7.47s of synthesis. Clause-level chunking (cap ~60
   chars at comma boundaries) cut it to ~4s. Later chunks pipeline behind
   playback.
7. **Quantization verdict — asymmetric**: int8-dynamic-quantizing the Piper
   voice (63.2 → 18.7 MB) made TTS **3–4× slower** (RTF 2.50–2.98 vs 0.79–0.86
   fp32): dequant overhead dominates the conv-heavy VITS graph on this CPU
   (no VNNI). Rejected here; artifact + verdict archived at
   [LoveLogicAI/piper-en_US-amy-medium-int8](https://huggingface.co/LoveLogicAI/piper-en_US-amy-medium-int8).
   ASR was already int8 (CTranslate2) and got faster via English-only tiny.en
   @1 thread.
8. **whisper tiny.en vs base accuracy**: base transcribed the bot name
   "Streamy" correctly; tiny.en hears "Streami". The latency profile uses
   tiny.en; base (+~2s ASR) is the accuracy upgrade path.

Remaining levers, in order of expected yield: (a) the ~0.9s of unexplained
utterance-finalization lag beyond the 600ms hangover; (b) in-context ASR/synth
contention (standalone RTF 0.25/0.79 vs in-context ~0.5/~1.2) — the render
loop is exonerated, so the suspect is WebRTC encode + thread interference;
(c) the concept's speculative dispatch at ~600ms silence (overlaps agent
response with utterance tail) — the largest structural win for turn-2+,
but a bigger build.

The tables below are the original design targets, kept for reference.

## Budget (from the concept package; measured results above)

These are the *targets* the MVP slice was designed against; the MEASURED
section above is the current scorecard against them.

| Stage | Local CPU-only | Local + GPU/ANE | Pika (published) |
|---|---|---|---|
| ASR finalize | 400–700ms | 150–250ms | — |
| Agent first token | 300–900ms | 150–400ms | — |
| TTS first chunk | 200–400ms | 80–150ms | — |
| Avatar render | 500–900ms | 200–350ms | — |
| **Perceived turn** | **~2.5–4.5s** | **~1.2–2.0s** | **~1.5s** |

## The mitigations that make 2.5s feel like 1.5s

1. **Speculative dispatch at ~600ms of silence** — partial ASR transcripts
   go to the agent before the utterance finalizes; resumed speech cancels
   the generation. Saves 200–600ms.
2. **Sentence-level pre-roll** — the agent's first sentence starts TTS
   while later sentences are still generating. Saves one TTS round-trip.
3. **Lead-out motion** — avatar mouth motion starts ~150ms before audio
   playout, perceptually erasing the render handshake.
4. **Cached identity** — avatar/voice assets are loaded and warmed at
   session start, never per turn.

## Fidelity ladder rungs & FPS targets

| Rung | Name | Target FPS | Host |
|---|---|---|---|
| 4 | FULL_PORTRAIT | 20–24 | GPU / Apple ANE |
| 3 | VISEME_STILL | ≥15 (MVP) | CPU |
| 2 | AUDIO_REACTIVE_WAVE | ≥10 | CPU (weak) |
| 1 | STATIC_CARD | n/a | anything |

Rung selection is `select_rung()` at join time; runtime degradation
triggers when measured FPS < target for 5 consecutive seconds (event:
`persona.rung_change`).

## Benchmark harness (MVP acceptance)

`python benchmarks/turn_latency.py` (lands with the MVP slice) measures,
per turn over ≥50 synthetic Q&A turns:

- `t_asr_final`: speech-end → final transcript
- `t_first_token`: transcript → agent first sentence
- `t_tts_chunk`: sentence → first audio chunk
- `t_avatar_frame`: audio chunk → first frame on the virtual camera
- `t_perceived`: speech-end → first audible+visible output

Win condition from the concept: **agent joins, answers within 4s with a
moving mouth, leaves, and produces a valid replay log** — i.e.
`t_perceived` p50 < 4.0s on CPU at rung 3, and
`EventLog(...).verify() == (True, None)`.

## Reporting rules

- Report p50/p95, machine, and library versions with every number.
- Faster-whisper int8, Piper medium voices; note thread counts.
- If FULL_PORTRAIT can't hold 20 FPS on Apple Silicon, video becomes a
  cloud-burst feature and rung 3 becomes the local ceiling — that is a
  finding to report, not a failure to hide.