LoveLogicAI commited on
Commit
6ce54c8
Β·
verified Β·
1 Parent(s): 71c3187

Latency budget + benchmark plan

Browse files
Files changed (1) hide show
  1. docs/LATENCY.md +64 -0
docs/LATENCY.md ADDED
@@ -0,0 +1,64 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # LATENCY β€” benchmarking methods & budget tables
2
+
3
+ ## Budget (from the concept package, to be measured β€” not yet achieved)
4
+
5
+ These are the *targets* the MVP slice is designed against. Nothing in this
6
+ repo has been benchmarked yet; the harness below is how every number gets
7
+ produced and re-produced.
8
+
9
+ | Stage | Local CPU-only | Local + GPU/ANE | Pika (published) |
10
+ |---|---|---|---|
11
+ | ASR finalize | 400–700ms | 150–250ms | β€” |
12
+ | Agent first token | 300–900ms | 150–400ms | β€” |
13
+ | TTS first chunk | 200–400ms | 80–150ms | β€” |
14
+ | Avatar render | 500–900ms | 200–350ms | β€” |
15
+ | **Perceived turn** | **~2.5–4.5s** | **~1.2–2.0s** | **~1.5s** |
16
+
17
+ ## The mitigations that make 2.5s feel like 1.5s
18
+
19
+ 1. **Speculative dispatch at ~600ms of silence** β€” partial ASR transcripts
20
+ go to the agent before the utterance finalizes; resumed speech cancels
21
+ the generation. Saves 200–600ms.
22
+ 2. **Sentence-level pre-roll** β€” the agent's first sentence starts TTS
23
+ while later sentences are still generating. Saves one TTS round-trip.
24
+ 3. **Lead-out motion** β€” avatar mouth motion starts ~150ms before audio
25
+ playout, perceptually erasing the render handshake.
26
+ 4. **Cached identity** β€” avatar/voice assets are loaded and warmed at
27
+ session start, never per turn.
28
+
29
+ ## Fidelity ladder rungs & FPS targets
30
+
31
+ | Rung | Name | Target FPS | Host |
32
+ |---|---|---|---|
33
+ | 4 | FULL_PORTRAIT | 20–24 | GPU / Apple ANE |
34
+ | 3 | VISEME_STILL | β‰₯15 (MVP) | CPU |
35
+ | 2 | AUDIO_REACTIVE_WAVE | β‰₯10 | CPU (weak) |
36
+ | 1 | STATIC_CARD | n/a | anything |
37
+
38
+ Rung selection is `select_rung()` at join time; runtime degradation
39
+ triggers when measured FPS < target for 5 consecutive seconds (event:
40
+ `persona.rung_change`).
41
+
42
+ ## Benchmark harness (MVP acceptance)
43
+
44
+ `python benchmarks/turn_latency.py` (lands with the MVP slice) measures,
45
+ per turn over β‰₯50 synthetic Q&A turns:
46
+
47
+ - `t_asr_final`: speech-end β†’ final transcript
48
+ - `t_first_token`: transcript β†’ agent first sentence
49
+ - `t_tts_chunk`: sentence β†’ first audio chunk
50
+ - `t_avatar_frame`: audio chunk β†’ first frame on the virtual camera
51
+ - `t_perceived`: speech-end β†’ first audible+visible output
52
+
53
+ Win condition from the concept: **agent joins, answers within 4s with a
54
+ moving mouth, leaves, and produces a valid replay log** β€” i.e.
55
+ `t_perceived` p50 < 4.0s on CPU at rung 3, and
56
+ `EventLog(...).verify() == (True, None)`.
57
+
58
+ ## Reporting rules
59
+
60
+ - Report p50/p95, machine, and library versions with every number.
61
+ - Faster-whisper int8, Piper medium voices; note thread counts.
62
+ - If FULL_PORTRAIT can't hold 20 FPS on Apple Silicon, video becomes a
63
+ cloud-burst feature and rung 3 becomes the local ceiling β€” that is a
64
+ finding to report, not a failure to hide.