Eterna TeknolojiAkıllı Çağrı Yönetimi

04Targets

Five quantified targets — literature-grounded and locked

Every target is measured with bootstrap confidence intervals on a frozen evaluation set locked from M4 onward.

fig. — quantified targets

5 targets · M4 frozen set

T1[K9]
target40%

%40+

Türkçe özel ad WER azalması

T2[K19]
target30%

≤ %20

Telefon bandı Türkçe WER

T3[K24]
target60%

p95 ≤ 2 s

Tur başına uçtan uca gecikme

T4[K5]
target95%

≥ %95

Alan doğruluğu / asılsız veri ≤ %1

T5[K9]
target90%

≥ %90

Serbest konum → idari birim

Detailed target cards

Each target is locked with its measurement method, success criterion, and linked work package.

T1WP3 · WP5

Turkish proper-name WER reduction (correction layer on)

≥ 40% relative drop

Literature anchor: [K9]51% relative WER reduction (K9); confidence threshold 0.5 (K11)

T2WP2 · WP3

Phone-band Turkish WER

WER ≤ 20%; fine-tune contributes ≥ 30% relative drop vs baseline

Literature anchor: [K1]Whisper Turkish base WER 4.3%–14.2%; LoRA fine-tune up to 52.38% reduction

T3WP3 · WP4 · WP8 · WP10

End-to-end per-turn latency

RTF < 1 and p95 ≤ 2 s

Literature anchor: [K24]Whisper-Streaming 3.3 s (K24); p95 reporting (K25)

T4WP7 · WP10

Record field accuracy & hallucinated-record rate

Field accuracy ≥ 95%; hallucinated-record rate ≤ 1%

Literature anchor: [K5]LLM correction precision / F1 (K5)

T5WP6 · WP10

Free-form location → administrative-unit accuracy

≥ 90% on golden set