From 81a33c906e1d8dfbfff770c23c1cd52787e65257 Mon Sep 17 00:00:00 2001
From: jmj <waysoleward01@naver.com>
Date: Sun, 28 Jun 2026 20:57:23 +0900
Subject: [PATCH] =?UTF-8?q?feat(feedback):=20=EC=A0=84=EB=8B=AC=EB=A0=A5?=
 =?UTF-8?q?=20=ED=94=BC=EB=93=9C=EB=B0=B1=20=EA=B0=95=ED=99=94=20=E2=80=94?=
 =?UTF-8?q?=20=EB=B0=B0=EC=A7=80=C2=B7=EC=BD=94=EC=B9=AD=20=EB=AC=B8?=
 =?UTF-8?q?=EC=9E=A5=C2=B7=EB=B0=9C=EC=9D=8C=20=EC=A0=95=ED=99=95=EB=8F=84?=
 =?UTF-8?q?=C2=B7=EC=9E=90=EA=B8=B0=EC=86=8C=EA=B0=9C=20=EA=B0=9C=EB=B3=84?=
 =?UTF-8?q?=20=EC=9D=8C=EC=84=B1?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

실전 면접 후 "목소리 피드백을 강화하고 싶다"는 요청 반영.

- DeliveryFeedback.assess: 답변 음성 메트릭(어절/분·무음 비율·100어절당 간투어·
  발음 신뢰도)에서 결정론적으로 전달력 배지(GOOD/FAIR/POOR) + 한 줄 코칭 산정(LLM 비호출).
  MessageResult/MessageResponse 에 deliveryRating/deliveryComment + pronunciationAccuracy
  노출(종료 세션·음성 답변만, 기존 게이팅 동일).
- 자기소개 첫인상: 세션 평균이 아니라 '자기소개 답변 단독' 음성 지표로 평가.
  SessionFeedbackRequester 가 generate.feedback.selfIntroVoiceAnalysis 동봉,
  AI feedback_consumer 가 첫인상 평가에 사용, self_intro 프롬프트가 음성 지표 적극 반영.
- 프론트 AnswerCoachingAccordion: 전달력 배지 + 발음 정확도 칩 + 코칭 문장 표시.
- OpenAPI 재생성, 프론트 타입 갱신. 백엔드 DeliveryFeedbackTest 추가, 전체 테스트 통과,
  프론트 빌드 통과(기존 lint 에러 2파일 무관).

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
---
 .../chain/prompts/self_intro_evaluation.py    |  3 +-
 .../messaging/consumers/feedback_consumer.py  |  8 +-
 ai/src/ai_server/model/messages/feedback.py   |  2 +
 backend/CLAUDE.md                             |  5 ++
 backend/openapi.json                          | 10 +++
 .../application/InterviewMessageService.java  |  4 +-
 .../application/SessionFeedbackRequester.java | 41 +++++++--
 .../application/dto/DeliveryFeedback.java     | 84 +++++++++++++++++++
 .../dto/GenerateFeedbackPayload.java          |  4 +-
 .../application/dto/MessageResult.java        | 21 ++++-
 .../presentation/dto/MessageResponse.java     | 11 ++-
 .../application/dto/DeliveryFeedbackTest.java | 54 ++++++++++++
 .../ui/live/AnswerCoachingAccordion.tsx       | 47 +++++++++--
 frontend/src/shared/api/generated.ts          |  4 +
 14 files changed, 271 insertions(+), 27 deletions(-)
 create mode 100644 backend/src/main/java/com/stackup/stackup/session/application/dto/DeliveryFeedback.java
 create mode 100644 backend/src/test/java/com/stackup/stackup/session/application/dto/DeliveryFeedbackTest.java

diff --git a/ai/src/ai_server/chain/prompts/self_intro_evaluation.py b/ai/src/ai_server/chain/prompts/self_intro_evaluation.py
index 870028be..0b2c697c 100644
--- a/ai/src/ai_server/chain/prompts/self_intro_evaluation.py
+++ b/ai/src/ai_server/chain/prompts/self_intro_evaluation.py
@@ -23,6 +23,7 @@
     "지원 직군: {job_category} / 면접 모드: {mode}\n\n"
     "=== 면접관 질문(자기소개) ===\n{self_intro_question}\n\n"
     "=== 지원자 자기소개 답변 ===\n{self_intro_answer}\n\n"
-    "=== (참고) 세션 전체 음성 지표 — 자기소개 단독이 아님 ===\n{voice_analysis_summary}\n\n"
+    "=== 자기소개 답변의 음성 지표(없으면 세션 평균) ===\n{voice_analysis_summary}\n\n"
+    "전달력 평가 시 위 음성 지표(발화 속도·무음·간투어)를 적극 반영하세요.\n\n"
     "{format_instructions}"
 )
diff --git a/ai/src/ai_server/messaging/consumers/feedback_consumer.py b/ai/src/ai_server/messaging/consumers/feedback_consumer.py
index b7464306..40a2d206 100644
--- a/ai/src/ai_server/messaging/consumers/feedback_consumer.py
+++ b/ai/src/ai_server/messaging/consumers/feedback_consumer.py
@@ -183,13 +183,19 @@ async def _evaluate_self_intro(
         if pair is None:
             return None  # 레거시 세션(자기소개 없음) 또는 빈 답변 — 건너뜀
         question, answer = pair
+        # 자기소개 답변 단독 음성 지표가 있으면 그걸, 없으면 세션 평균을 폴백으로 사용.
+        intro_voice_summary = (
+            _build_voice_analysis_summary(req.self_intro_voice_analysis)
+            if req.self_intro_voice_analysis is not None
+            else voice_analysis_summary
+        )
         try:
             ev = await self._self_intro_evaluator.evaluate(
                 job_category=req.job_category,
                 mode=req.mode,
                 self_intro_question=question.content,
                 self_intro_answer=answer.content,
-                voice_analysis_summary=voice_analysis_summary,
+                voice_analysis_summary=intro_voice_summary,
             )
         except Exception as exc:  # noqa: BLE001
             log.warning(
diff --git a/ai/src/ai_server/model/messages/feedback.py b/ai/src/ai_server/model/messages/feedback.py
index 0b9ca553..7b2cdee1 100644
--- a/ai/src/ai_server/model/messages/feedback.py
+++ b/ai/src/ai_server/model/messages/feedback.py
@@ -62,6 +62,8 @@ class GenerateFeedbackRequest(BaseModel):
     messages: list[FeedbackMessageItem] = Field(default_factory=list)
     context_document_ids: list[int] = Field(default_factory=list)
     voice_analysis_summary: VoiceAnalysisSummary | None = None
+    # 자기소개 답변 단독의 음성 메트릭. 첫인상 평가에서 세션 평균 대신 사용. 없으면 None.
+    self_intro_voice_analysis: VoiceAnalysisSummary | None = None
     # 다직군 패널 가중: 사용된 일반질문의 직군별 개수. 비면 단일 직군 평가.
     domain_question_counts: dict[str, int] = Field(default_factory=dict)
     # 직무 맞춤(JOB_TAILORED) 모드 전용. 회사명 + 채용공고(JD). '직무 적합도' 평가의 근거.
diff --git a/backend/CLAUDE.md b/backend/CLAUDE.md
index f42ab8db..7fe58134 100644
--- a/backend/CLAUDE.md
+++ b/backend/CLAUDE.md
@@ -376,6 +376,11 @@ docker compose up -d
   `FeedbackCallbackService` 가 각 메시지에 `recordCoaching` 기록. `MessageResult`/`MessageResponse` 가 답변
   평가 점수 + 복기 + **답변 전달력 메트릭(WPM·무음·간투어, `MessageVoiceAnalysis` 에서 파싱)** 을 노출하되
   **종료 세션 조회에서만**(`expectedSignal` 과 동일 게이팅). 프론트는 답변 버블 아래 '복기' 아코디언으로 표시.
+- **전달력 피드백 강화 본 구현**: `MessageResult`/`MessageResponse` 가 `pronunciationAccuracy`(STT 신뢰도 근사)
+  와 **결정론적 전달력 평가**(`DeliveryFeedback.assess` — 어절/분·무음 비율·100어절당 간투어·발음 임계치 →
+  `deliveryRating` GOOD/FAIR/POOR + `deliveryComment` 한 줄 코칭, LLM 비호출)를 추가 노출(종료 세션만, 음성 답변만).
+  자기소개 첫인상은 세션 평균이 아니라 **자기소개 답변 단독 음성 지표**로 평가 — `SessionFeedbackRequester` 가
+  `generate.feedback.selfIntroVoiceAnalysis`(자기소개 답변의 WPM/무음/간투어)를 동봉하고 AI 첫인상 평가가 이를 사용.
 - **직무 맞춤 면접 모드(JOB_TAILORED) 본 구현**: `SessionMode.JOB_TAILORED` 추가(V18 — mode CHECK 갱신 +
   `target_company_name`/`target_job_description` 컬럼). 이 모드는 회사명+채용공고(JD)를 받아(JD 필수,
   `SessionService` 검증 `SESSION_JD_REQUIRED`) `InterviewSession.assignTargetRole` 로 보관. JD 는
diff --git a/backend/openapi.json b/backend/openapi.json
index e5962af8..353cb425 100644
--- a/backend/openapi.json
+++ b/backend/openapi.json
@@ -2843,6 +2843,16 @@
               "type" : "integer",
               "format" : "int32"
             }
+          },
+          "pronunciationAccuracy" : {
+            "type" : "number",
+            "format" : "double"
+          },
+          "deliveryRating" : {
+            "type" : "string"
+          },
+          "deliveryComment" : {
+            "type" : "string"
           }
         }
       },
diff --git a/backend/src/main/java/com/stackup/stackup/session/application/InterviewMessageService.java b/backend/src/main/java/com/stackup/stackup/session/application/InterviewMessageService.java
index 799da424..29bbdb56 100644
--- a/backend/src/main/java/com/stackup/stackup/session/application/InterviewMessageService.java
+++ b/backend/src/main/java/com/stackup/stackup/session/application/InterviewMessageService.java
@@ -74,7 +74,9 @@ private MessageResult toResultWithAudioUrls(
         Double wpm = voice == null ? null : voice.getSpeakingRateWpm();
         Double silence = voice == null ? null : voice.getSilenceDurationSec();
         Map<String, Integer> fillers = voice == null ? null : parseFillers(voice.getFillerWordCounts());
-        return MessageResult.of(m, ttsUrl, audioUrl, revealInsights, wpm, silence, fillers);
+        Double pronunciation = voice == null ? null : voice.getPronunciationAccuracy();
+        return MessageResult.of(
+            m, ttsUrl, audioUrl, revealInsights, wpm, silence, fillers, pronunciation);
     }
 
     private Map<String, Integer> parseFillers(String json) {
diff --git a/backend/src/main/java/com/stackup/stackup/session/application/SessionFeedbackRequester.java b/backend/src/main/java/com/stackup/stackup/session/application/SessionFeedbackRequester.java
index f022995e..5a4ca929 100644
--- a/backend/src/main/java/com/stackup/stackup/session/application/SessionFeedbackRequester.java
+++ b/backend/src/main/java/com/stackup/stackup/session/application/SessionFeedbackRequester.java
@@ -61,10 +61,9 @@ public void onSessionEnded(SessionEndedEvent event) {
             return;
         }
 
-        List<MessageItem> messages = messageRepository
-            .findBySession_IdOrderBySequenceNumberAsc(event.sessionId()).stream()
-            .map(this::toItem)
-            .toList();
+        List<InterviewMessage> msgEntities =
+            messageRepository.findBySession_IdOrderBySequenceNumberAsc(event.sessionId());
+        List<MessageItem> messages = msgEntities.stream().map(this::toItem).toList();
         List<Long> contextDocumentIds = contextRepository.findBySession_Id(event.sessionId()).stream()
             .map(c -> c.getDocument().getId())
             .toList();
@@ -76,6 +75,8 @@ public void onSessionEnded(SessionEndedEvent event) {
             domainQuestionCounts.merge(jc, 1, Integer::sum);
         }
 
+        List<MessageVoiceAnalysis> voiceAnalyses =
+            voiceAnalysisRepository.findByMessage_Session_Id(event.sessionId());
         GenerateFeedbackPayload payload = new GenerateFeedbackPayload(
             session.getId(),
             session.getMode().name(),
@@ -84,10 +85,11 @@ public void onSessionEnded(SessionEndedEvent event) {
             event.reason(),
             messages,
             contextDocumentIds,
-            summarizeVoiceAnalysis(event.sessionId()),
+            summarize(voiceAnalyses),
             domainQuestionCounts,
             session.getTargetCompanyName(),
-            session.getTargetJobDescription()
+            session.getTargetJobDescription(),
+            selfIntroVoiceAnalysis(msgEntities, voiceAnalyses)
         );
 
         publisher.publishToAi(
@@ -124,8 +126,31 @@ private MessageItem toItem(InterviewMessage m) {
         );
     }
 
-    private VoiceAnalysisSummary summarizeVoiceAnalysis(Long sessionId) {
-        List<MessageVoiceAnalysis> analyses = voiceAnalysisRepository.findByMessage_Session_Id(sessionId);
+    // 자기소개 답변(첫 질문에 대한 INTERVIEWEE 답변)의 음성 메트릭만 따로 요약. 첫인상 평가에서
+    // 세션 평균 대신 자기소개 단독 전달력을 보게 한다. 자기소개 답변·음성 분석이 없으면 null.
+    private VoiceAnalysisSummary selfIntroVoiceAnalysis(
+        List<InterviewMessage> msgEntities, List<MessageVoiceAnalysis> analyses) {
+        InterviewMessage selfIntroQuestion = msgEntities.stream()
+            .filter(InterviewMessage::isSelfIntroduction)
+            .findFirst().orElse(null);
+        if (selfIntroQuestion == null) {
+            return null;
+        }
+        Long answerId = msgEntities.stream()
+            .filter(m -> m.getParentMessage() != null
+                && selfIntroQuestion.getId().equals(m.getParentMessage().getId()))
+            .map(InterviewMessage::getId)
+            .findFirst().orElse(null);
+        if (answerId == null) {
+            return null;
+        }
+        List<MessageVoiceAnalysis> own = analyses.stream()
+            .filter(a -> answerId.equals(a.getMessage().getId()))
+            .toList();
+        return own.isEmpty() ? null : summarize(own);
+    }
+
+    private VoiceAnalysisSummary summarize(List<MessageVoiceAnalysis> analyses) {
         if (analyses.isEmpty()) {
             return new VoiceAnalysisSummary(0, null, 0.0, Map.of());
         }
diff --git a/backend/src/main/java/com/stackup/stackup/session/application/dto/DeliveryFeedback.java b/backend/src/main/java/com/stackup/stackup/session/application/dto/DeliveryFeedback.java
new file mode 100644
index 00000000..4a5e761e
--- /dev/null
+++ b/backend/src/main/java/com/stackup/stackup/session/application/dto/DeliveryFeedback.java
@@ -0,0 +1,84 @@
+package com.stackup.stackup.session.application.dto;
+
+import java.util.ArrayList;
+import java.util.List;
+import java.util.Map;
+
+// 답변 음성 메트릭(WPM·무음·간투어·발음 신뢰도) → 결정론적 전달력 평가. LLM 비호출.
+// rating: GOOD/FAIR/POOR (이슈 0/1/2+개), comment: 답변별 한 줄 전달력 코칭.
+// 임계치는 어절(공백 기준) 단위 기준. content 로 어절 수를 세어 무음 비율·간투어 밀도를 정규화.
+public record DeliveryFeedback(String rating, String comment) {
+
+    private static final double WPM_SLOW = 90;        // 어절/분 미만이면 느림
+    private static final double WPM_FAST = 150;        // 어절/분 초과면 빠름
+    private static final double SILENCE_RATIO_HIGH = 0.35;  // 전체 발화시간 중 무음 비율
+    private static final double FILLERS_PER_100_HIGH = 3.0;  // 100어절당 간투어 개수
+    private static final double PRONUNCIATION_LOW = 0.85;    // STT 신뢰도 근사
+
+    // 음성 메트릭이 하나라도 있으면 평가, 전부 없으면(텍스트 답변 등) null.
+    public static DeliveryFeedback assess(
+        String content, Double wpm, Double silenceSec,
+        Map<String, Integer> fillers, Double pronunciation) {
+        int fillerTotal = fillers == null ? 0
+            : fillers.values().stream().mapToInt(Integer::intValue).sum();
+        boolean noMetrics = wpm == null && silenceSec == null
+            && fillerTotal == 0 && pronunciation == null;
+        if (noMetrics) {
+            return null;
+        }
+
+        int words = wordCount(content);
+        List<String> issues = new ArrayList<>();
+
+        if (wpm != null && words > 0) {
+            if (wpm > WPM_FAST) {
+                issues.add("말이 빠른 편입니다(약 %d 어절/분). 핵심 문장 앞에서 한 박자 쉬어 강조하세요."
+                    .formatted(Math.round(wpm)));
+            } else if (wpm < WPM_SLOW) {
+                issues.add("말이 다소 느립니다(약 %d 어절/분). 군더더기를 줄이고 핵심을 먼저 말해보세요."
+                    .formatted(Math.round(wpm)));
+            }
+        }
+
+        // 무음 비율: duration ≈ words / wpm * 60 (WPM 의 분모가 전체 오디오 길이).
+        if (silenceSec != null && silenceSec > 0 && wpm != null && wpm > 0 && words > 0) {
+            double durationSec = words / wpm * 60.0;
+            double ratio = durationSec > 0 ? silenceSec / durationSec : 0;
+            if (ratio > SILENCE_RATIO_HIGH) {
+                issues.add("말 사이 침묵이 깁니다(총 %d초). 답변을 미리 구조화해 끊김을 줄여보세요."
+                    .formatted(Math.round(silenceSec)));
+            }
+        }
+
+        if (fillerTotal > 0 && words > 0) {
+            double per100 = fillerTotal * 100.0 / words;
+            if (per100 > FILLERS_PER_100_HIGH) {
+                String top = fillers.entrySet().stream()
+                    .max(Map.Entry.comparingByValue()).map(Map.Entry::getKey).orElse("어");
+                issues.add("간투어가 잦습니다('%s' 등 %d회). 한 문장을 끝맺고 잠깐 멈추는 연습이 도움이 됩니다."
+                    .formatted(top, fillerTotal));
+            }
+        }
+
+        if (pronunciation != null && pronunciation < PRONUNCIATION_LOW) {
+            issues.add("발음이 다소 뭉개져 인식 정확도가 낮습니다. 또박또박 발음해보세요.");
+        }
+
+        String rating = switch (issues.size()) {
+            case 0 -> "GOOD";
+            case 1 -> "FAIR";
+            default -> "POOR";
+        };
+        String comment = issues.isEmpty()
+            ? "전달이 안정적입니다. 속도·간투어·침묵 모두 적정 범위입니다."
+            : String.join(" ", issues);
+        return new DeliveryFeedback(rating, comment);
+    }
+
+    private static int wordCount(String content) {
+        if (content == null || content.isBlank()) {
+            return 0;
+        }
+        return content.trim().split("\\s+").length;
+    }
+}
diff --git a/backend/src/main/java/com/stackup/stackup/session/application/dto/GenerateFeedbackPayload.java b/backend/src/main/java/com/stackup/stackup/session/application/dto/GenerateFeedbackPayload.java
index d684d698..74d1d798 100644
--- a/backend/src/main/java/com/stackup/stackup/session/application/dto/GenerateFeedbackPayload.java
+++ b/backend/src/main/java/com/stackup/stackup/session/application/dto/GenerateFeedbackPayload.java
@@ -18,7 +18,9 @@ public record GenerateFeedbackPayload(
     Map<String, Integer> domainQuestionCounts,
     // 직무 맞춤 모드 전용. 회사명 + 채용공고(JD). '직무 적합도' 평가의 근거. 다른 모드는 null.
     String targetCompanyName,
-    String targetJobDescription
+    String targetJobDescription,
+    // 자기소개 답변 단독의 음성 메트릭. 첫인상 평가에서 세션 평균 대신 사용. 없으면 null.
+    VoiceAnalysisSummary selfIntroVoiceAnalysis
 ) {
 
     public record MessageItem(
diff --git a/backend/src/main/java/com/stackup/stackup/session/application/dto/MessageResult.java b/backend/src/main/java/com/stackup/stackup/session/application/dto/MessageResult.java
index bdd205cd..c700ed96 100644
--- a/backend/src/main/java/com/stackup/stackup/session/application/dto/MessageResult.java
+++ b/backend/src/main/java/com/stackup/stackup/session/application/dto/MessageResult.java
@@ -39,7 +39,11 @@ public record MessageResult(
     // 답변 전달력 메트릭(음성 답변에만, 종료 세션 조회에서만). 별도 엔티티(MessageVoiceAnalysis)에서 온다.
     Double speakingRateWpm,
     Double silenceDurationSec,
-    Map<String, Integer> fillerWordCounts
+    Map<String, Integer> fillerWordCounts,
+    Double pronunciationAccuracy,
+    // 위 메트릭에서 결정론적으로 산정한 전달력 평가(배지 GOOD/FAIR/POOR + 한 줄 코칭). 음성 답변에만.
+    String deliveryRating,
+    String deliveryComment
 ) {
     public static MessageResult of(InterviewMessage m) {
         return of(m, null, null, false);
@@ -51,13 +55,19 @@ public static MessageResult of(InterviewMessage m, String ttsAudioUrl, String au
 
     public static MessageResult of(
         InterviewMessage m, String ttsAudioUrl, String audioFileUrl, boolean revealInsights) {
-        return of(m, ttsAudioUrl, audioFileUrl, revealInsights, null, null, null);
+        return of(m, ttsAudioUrl, audioFileUrl, revealInsights, null, null, null, null);
     }
 
     // revealInsights=true 면 답변 평가·복기·전달력 메트릭·expectedSignal 노출(종료 세션). 라이브/대기 중엔 false.
     public static MessageResult of(
         InterviewMessage m, String ttsAudioUrl, String audioFileUrl, boolean revealInsights,
-        Double speakingRateWpm, Double silenceDurationSec, Map<String, Integer> fillerWordCounts) {
+        Double speakingRateWpm, Double silenceDurationSec, Map<String, Integer> fillerWordCounts,
+        Double pronunciationAccuracy) {
+        DeliveryFeedback delivery = revealInsights
+            ? DeliveryFeedback.assess(
+                m.getContent(), speakingRateWpm, silenceDurationSec,
+                fillerWordCounts, pronunciationAccuracy)
+            : null;
         return new MessageResult(
             m.getId(),
             m.getSession().getId(),
@@ -85,7 +95,10 @@ public static MessageResult of(
             revealInsights ? m.getCoachingComment() : null,
             revealInsights ? speakingRateWpm : null,
             revealInsights ? silenceDurationSec : null,
-            revealInsights ? fillerWordCounts : null
+            revealInsights ? fillerWordCounts : null,
+            revealInsights ? pronunciationAccuracy : null,
+            delivery == null ? null : delivery.rating(),
+            delivery == null ? null : delivery.comment()
         );
     }
 }
diff --git a/backend/src/main/java/com/stackup/stackup/session/presentation/dto/MessageResponse.java b/backend/src/main/java/com/stackup/stackup/session/presentation/dto/MessageResponse.java
index 4b698fa3..890f35de 100644
--- a/backend/src/main/java/com/stackup/stackup/session/presentation/dto/MessageResponse.java
+++ b/backend/src/main/java/com/stackup/stackup/session/presentation/dto/MessageResponse.java
@@ -37,7 +37,11 @@ public record MessageResponse(
     // 답변 전달력 메트릭(음성 답변·종료 세션에서만).
     Double speakingRateWpm,
     Double silenceDurationSec,
-    java.util.Map<String, Integer> fillerWordCounts
+    java.util.Map<String, Integer> fillerWordCounts,
+    Double pronunciationAccuracy,
+    // 전달력 메트릭에서 산정한 배지(GOOD/FAIR/POOR)와 한 줄 코칭.
+    String deliveryRating,
+    String deliveryComment
 ) {
     public static MessageResponse from(MessageResult r) {
         return new MessageResponse(
@@ -67,7 +71,10 @@ public static MessageResponse from(MessageResult r) {
             r.coachingComment(),
             r.speakingRateWpm(),
             r.silenceDurationSec(),
-            r.fillerWordCounts()
+            r.fillerWordCounts(),
+            r.pronunciationAccuracy(),
+            r.deliveryRating(),
+            r.deliveryComment()
         );
     }
 }
diff --git a/backend/src/test/java/com/stackup/stackup/session/application/dto/DeliveryFeedbackTest.java b/backend/src/test/java/com/stackup/stackup/session/application/dto/DeliveryFeedbackTest.java
new file mode 100644
index 00000000..f3d4f17a
--- /dev/null
+++ b/backend/src/test/java/com/stackup/stackup/session/application/dto/DeliveryFeedbackTest.java
@@ -0,0 +1,54 @@
+package com.stackup.stackup.session.application.dto;
+
+import static org.assertj.core.api.Assertions.assertThat;
+
+import java.util.Map;
+import org.junit.jupiter.api.Test;
+
+class DeliveryFeedbackTest {
+
+    // 약 60어절 분량의 답변(무음 비율·간투어 밀도 정규화의 분모).
+    private static final String ANSWER = "저는 백엔드 개발자로서 "
+        + "대용량 트래픽을 처리하는 결제 시스템을 설계하고 운영한 경험이 있습니다 "
+        + "특히 데이터베이스 인덱스를 튜닝하고 캐시 계층을 도입해 응답 속도를 크게 "
+        + "개선했으며 장애 상황에서도 안정적으로 동작하도록 모니터링과 알림 체계를 "
+        + "구축했습니다 또한 동료들과 코드 리뷰 문화를 정착시켜 품질을 높였습니다 "
+        + "앞으로도 견고한 시스템을 만드는 개발자가 되고 싶습니다 감사합니다";
+
+    @Test
+    void noMetrics_returnsNull() {
+        assertThat(DeliveryFeedback.assess(ANSWER, null, null, null, null)).isNull();
+        assertThat(DeliveryFeedback.assess(ANSWER, null, null, Map.of(), null)).isNull();
+    }
+
+    @Test
+    void cleanDelivery_isGood() {
+        DeliveryFeedback fb = DeliveryFeedback.assess(ANSWER, 120.0, 4.0, Map.of(), 0.95);
+        assertThat(fb).isNotNull();
+        assertThat(fb.rating()).isEqualTo("GOOD");
+        assertThat(fb.comment()).contains("안정");
+    }
+
+    @Test
+    void fastPace_isFlagged() {
+        DeliveryFeedback fb = DeliveryFeedback.assess(ANSWER, 180.0, 4.0, Map.of(), 0.95);
+        assertThat(fb.rating()).isEqualTo("FAIR");
+        assertThat(fb.comment()).contains("빠른");
+    }
+
+    @Test
+    void multipleIssues_arePoor() {
+        // 빠른 속도 + 잦은 간투어 + 낮은 발음 신뢰도 → 2건 이상 → POOR.
+        DeliveryFeedback fb = DeliveryFeedback.assess(
+            ANSWER, 185.0, 4.0, Map.of("어", 6, "그", 5), 0.70);
+        assertThat(fb.rating()).isEqualTo("POOR");
+        assertThat(fb.comment()).contains("간투어");
+    }
+
+    @Test
+    void normalFillerDensity_notFlagged() {
+        // ~60어절에 간투어 1회 → 100어절당 ~1.7개로 임계(3) 미만 → 미지적.
+        DeliveryFeedback fb = DeliveryFeedback.assess(ANSWER, 120.0, 4.0, Map.of("어", 1), 0.95);
+        assertThat(fb.rating()).isEqualTo("GOOD");
+    }
+}
diff --git a/frontend/src/features/interview/ui/live/AnswerCoachingAccordion.tsx b/frontend/src/features/interview/ui/live/AnswerCoachingAccordion.tsx
index 06c97362..8225e897 100644
--- a/frontend/src/features/interview/ui/live/AnswerCoachingAccordion.tsx
+++ b/frontend/src/features/interview/ui/live/AnswerCoachingAccordion.tsx
@@ -12,6 +12,13 @@ function score5(v?: number | null): string | null {
   return typeof v === 'number' ? `${v}/5` : null
 }
 
+// 전달력 배지(결정론적 평가) → 라벨 + 색상 토큰.
+const DELIVERY_RATING: Record<string, { label: string; cls: string }> = {
+  GOOD: { label: '전달력 우수', cls: 'bg-success-50 text-success-700' },
+  FAIR: { label: '전달력 보통', cls: 'bg-warning-50 text-warning-700' },
+  POOR: { label: '전달력 미흡', cls: 'bg-danger-50 text-danger-700' },
+}
+
 // 답변 음성 전달력 메트릭 → 표시용 칩 문자열들.
 function deliveryChips(message: Message): string[] {
   const chips: string[] = []
@@ -25,6 +32,9 @@ function deliveryChips(message: Message): string[] {
   if (fillers.length > 0) {
     chips.push('간투어 ' + fillers.map(([w, c]) => `${w} ${c}`).join(' · '))
   }
+  if (typeof message.pronunciationAccuracy === 'number') {
+    chips.push(`발음 정확도 ${Math.round(message.pronunciationAccuracy * 100)}%`)
+  }
   return chips
 }
 
@@ -33,8 +43,13 @@ function deliveryChips(message: Message): string[] {
 export function AnswerCoachingAccordion({ message }: { message: Message }) {
   const [open, setOpen] = useState(false)
   const delivery = deliveryChips(message)
+  const rating = message.deliveryRating ? DELIVERY_RATING[message.deliveryRating] : undefined
   const hasContent = Boolean(
-    message.modelAnswer || message.answerRewrite || message.coachingComment || delivery.length > 0,
+    message.modelAnswer ||
+      message.answerRewrite ||
+      message.coachingComment ||
+      message.deliveryComment ||
+      delivery.length > 0,
   )
   if (!hasContent) return null
 
@@ -61,19 +76,33 @@ export function AnswerCoachingAccordion({ message }: { message: Message }) {
       </button>
       {open && (
         <div className="mt-1.5 flex flex-col gap-3 rounded-lg border border-border bg-surface px-4 py-3 text-left">
-          {delivery.length > 0 && (
+          {(delivery.length > 0 || rating || message.deliveryComment) && (
             <div className="flex flex-col gap-1">
-              <span className="text-caption text-fg-subtle">전달력</span>
-              <div className="flex flex-wrap gap-1.5">
-                {delivery.map((d) => (
+              <div className="flex items-center gap-2">
+                <span className="text-caption text-fg-subtle">전달력</span>
+                {rating && (
                   <span
-                    key={d}
-                    className="rounded-pill bg-info-50 px-2 py-0.5 text-caption text-info-700"
+                    className={`rounded-pill px-2 py-0.5 text-caption font-medium ${rating.cls}`}
                   >
-                    {d}
+                    {rating.label}
                   </span>
-                ))}
+                )}
               </div>
+              {delivery.length > 0 && (
+                <div className="flex flex-wrap gap-1.5">
+                  {delivery.map((d) => (
+                    <span
+                      key={d}
+                      className="rounded-pill bg-info-50 px-2 py-0.5 text-caption text-info-700"
+                    >
+                      {d}
+                    </span>
+                  ))}
+                </div>
+              )}
+              {message.deliveryComment && (
+                <p className="text-caption text-fg-muted">{message.deliveryComment}</p>
+              )}
             </div>
           )}
           {evals.length > 0 && (
diff --git a/frontend/src/shared/api/generated.ts b/frontend/src/shared/api/generated.ts
index ad0ec384..5ef70d07 100644
--- a/frontend/src/shared/api/generated.ts
+++ b/frontend/src/shared/api/generated.ts
@@ -917,6 +917,10 @@ export interface components {
             fillerWordCounts?: {
                 [key: string]: number;
             };
+            /** Format: double */
+            pronunciationAccuracy?: number;
+            deliveryRating?: string;
+            deliveryComment?: string;
         };
         VoiceStreamBeginResponse: {
             /** Format: int64 */