Quantize-to-fit results: tiny.en ASR 3.11s pass; piper int8 rejected (3-4x slower), artifact archived

docs/LATENCY.md

@@ -2,15 +2,16 @@
 
 ## MEASURED — MVP slice, 2026-10-04, cpu-upgrade (8 vCPU cgroup quota, CPU-only)
 
-End-to-end run `ml-intern-freestream-e2e-wincondition` (final: job `6ac2e0f3404719ba376516e5`):
+End-to-end run `ml-intern-freestream-e2e-wincondition` (final: job `6ac2e372404719ba3765176e`,
+ASR profile tiny.en @1 thread):
 livekit-server v1.13.7 dev mode in-container, bot joined via the real CLI,
 test client asked a spoken question, timings measured at the receiving client.
 
 | Metric | Measured | Target | Verdict |
 |---|---|---|---|
-| **t_perceived** (speech-end → first bot audio at client) | **3.75s** | ≤ 4.0s (win condition) | **PASS** |
+| **t_perceived** (speech-end → first bot audio at client) | **3.11s** | ≤ 4.0s (win condition) | **PASS** |
 | t_ready (CLI join → bot ready, incl. model + disclosure warmup) | 14.6s | — | — |
-| ASR finalize (whisper tiny int8, 2 threads) | 2.04s | 400–700ms | over budget |
+| ASR finalize (whisper tiny.en int8, 1 thread) | 1.33s | 400–700ms | over, halved from 2.04s |
 | TTS first-audio path (cached consent disclosure) | ~0.2s (pre-roll + transport) | — | PASS |
 | Full turn (speech-end → answer finished) | 18.9s (14.9s of audio) | — | pipelined |
 | Piper in-context RTF (later sentences) | ~0.8–1.2 | <1.0 | borderline |
@@ -33,9 +34,18 @@ Findings from the measurement campaign:
    while sentence 1 plays, mitigation #2) keeps the turn flowing: 14.9s of
    audio completed in 18.9s wall clock. Fixing the contention is the next
    performance item.
-4. **whisper tiny vs base accuracy**: base transcribed the bot name
-   "Streamy" correctly; tiny heard "Streami"/"streaming". The latency profile
-   uses tiny; base (+~2s ASR) is the accuracy upgrade path.
+4. **whisper tiny.en vs base accuracy**: base transcribed the bot name
+   "Streamy" correctly; tiny/tiny.en heard "Streami"/"streaming". The latency
+   profile uses tiny.en (measured best: standalone RTF 0.25 @1 thread vs 0.31
+   for tiny @2); base (+~2s ASR) is the accuracy upgrade path.
+5. **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 for this hardware; artifact + verdict archived at
+   [LoveLogicAI/piper-en_US-amy-medium-int8](https://huggingface.co/LoveLogicAI/piper-en_US-amy-medium-int8)
+   for re-testing on VNNI-capable hosts. ASR, in contrast, was **already**
+   int8 (CTranslate2 compute_type="int8") and got faster by switching to the
+   English-only tiny.en at 1 thread.
 
 The tables below are the original design targets, kept for reference.