@gobing-ai/knowledge-kit 0.0.8 → 0.0.10

package/plugins/generations/dailynews-gen/src/script-builder.ts

@@ -0,0 +1,430 @@
+import type { Doc } from '@gobing-ai/kk-core';
+import { z } from 'zod';
+
+export const EffectConfigSchema = z.object({
+    type: z.string().min(1),
+    enabled: z.boolean().optional(),
+    params: z.record(z.string(), z.number()).optional(),
+});
+
+export type EffectConfig = z.infer<typeof EffectConfigSchema>;
+
+export const VoiceSpeakerSchema = z.object({
+    profile: z.string().min(1).optional(),
+    profile_id: z.string().min(1).optional(),
+    name: z.string().min(1).optional(),
+    role: z.string().min(1).optional(),
+    engine: z.string().min(1).optional(),
+    language: z.string().min(1).optional(),
+    emotion: z.string().min(1).optional(),
+    instruct: z.string().min(1).optional(),
+    effects_chain: z.array(EffectConfigSchema).optional(),
+    effect_preset: z.string().min(1).optional(),
+    personality: z.boolean().optional(),
+});
+
+export type VoiceSpeaker = z.infer<typeof VoiceSpeakerSchema>;
+
+export const VoiceScriptSegmentSchema = z.object({
+    speaker: z.string().min(1).optional(),
+    profile: z.string().min(1).optional(),
+    profile_id: z.string().min(1).optional(),
+    text: z.string().min(1),
+    emotion: z.string().min(1).optional(),
+    instruct: z.string().min(1).optional(),
+    engine: z.string().min(1).optional(),
+    language: z.string().min(1).optional(),
+    gap_ms: z.number().int().nonnegative().optional(),
+    max_chunk_chars: z.number().int().optional(),
+    crossfade_ms: z.number().int().optional(),
+    personality: z.boolean().optional(),
+    seed: z.number().int().nonnegative().optional(),
+    effects_chain: z.array(EffectConfigSchema).optional(),
+    effect_preset: z.string().min(1).optional(),
+});
+
+export type VoiceScriptSegment = z.infer<typeof VoiceScriptSegmentSchema>;
+
+export const VoiceScriptSchema = z.object({
+    title: z.string().min(1),
+    language: z.enum(['zh', 'en']),
+    default_profile: z.string().min(1).optional(),
+    default_profile_id: z.string().min(1).optional(),
+    default_engine: z.string().min(1).optional(),
+    default_emotion: z.string().min(1).optional(),
+    default_effects_chain: z.array(EffectConfigSchema).optional(),
+    default_effect_preset: z.string().min(1).optional(),
+    max_chunk_chars: z.number().int().optional(),
+    crossfade_ms: z.number().int().optional(),
+    speakers: z.record(z.string(), VoiceSpeakerSchema).optional(),
+    segments: z.array(VoiceScriptSegmentSchema).min(1),
+});
+
+export type VoiceScript = z.infer<typeof VoiceScriptSchema>;
+
+export interface ScriptBuilderOptions {
+    language?: string;
+    voiceProfile?: string;
+    voiceProfileId?: string;
+    title?: string;
+    date?: Date;
+    emotion?: string;
+    effectPreset?: string;
+    speakers?: Record<string, VoiceSpeaker>;
+}
+
+const ZH_DIGITS = ['零', '一', '二', '三', '四', '五', '六', '七', '八', '九', '十'];
+
+export function formatChineseCount(n: number): string {
+    if (n === 2) return '两';
+    if (n >= 0 && n <= 10) return ZH_DIGITS[n] ?? String(n);
+    return String(n);
+}
+
+export function formatBroadcastDate(date: Date, language: 'zh' | 'en'): string {
+    if (language === 'zh') {
+        const year = date.getFullYear();
+        const month = date.getMonth() + 1;
+        const day = date.getDate();
+        const weekdays = ['日', '一', '二', '三', '四', '五', '六'];
+        const weekday = weekdays[date.getDay()] ?? '日';
+        return `${year}年${month}月${day}日，星期${weekday}`;
+    }
+    return new Intl.DateTimeFormat('en-US', {
+        weekday: 'long',
+        month: 'long',
+        day: 'numeric',
+        year: 'numeric',
+    }).format(date);
+}
+
+export function normalizeBroadcastText(raw: string, isZh: boolean): string {
+    let text = raw.trim();
+
+    // 1. Strip markdown links, bold, headers, code, blockquotes
+    // Preserves paralinguistic tags like [laugh] or [sigh] because they have no (url)
+    text = text.replace(/\[([^\]]+)\]\([^)]+\)/g, '$1');
+    text = text.replace(/[*#_`>~]/g, '');
+    text = text.replace(/\\"/g, '');
+    text = text.replace(/["“”«»「」『』]/g, '');
+
+    // 2. Normalize whitespace
+    text = text.replace(/\s+/g, ' ');
+
+    if (isZh) {
+        text = text.replace(/[''‘’]/g, '');
+        // Replace English punctuation with Chinese pauses
+        text = text.replace(/[:：]/g, '，');
+        text = text.replace(/[;；]/g, '。');
+        text = text.replace(/--+/g, '，');
+        text = text.replace(/\.{2,}/g, '。');
+        text = text.replace(/([0-9]+)\s*条/g, (_, num) => `${formatChineseCount(parseInt(num, 10))}条`);
+        text = text.replace(/[!！]/g, '。');
+        text = text.replace(/\s*，\s*/g, '，');
+        text = text.replace(/\s*。\s*/g, '。');
+
+        // Ensure clean final period
+        text = text.replace(/[，、：；\s]+$/, '');
+        if (!text.endsWith('。') && !text.endsWith('？')) {
+            text += '。';
+        }
+    } else {
+        text = text.replace(/[:;]/g, ',');
+        text = text.replace(/--+/g, ',');
+        text = text.replace(/\.{2,}/g, '.');
+        text = text.replace(/[,\s]+$/, '');
+        if (!text.endsWith('.') && !text.endsWith('?')) {
+            text += '.';
+        }
+    }
+
+    return text;
+}
+
+/**
+ * Synthesizes a natural, concise spoken perspective / takeaway based on the story facts.
+ * Gives the monologue a demonstrative tech podcast talk feel rather than mechanical news reading.
+ */
+export function synthesizePersonalUnderstanding(title: string, body: string, isZh: boolean): string {
+    const combined = `${title} ${body}`.toLowerCase();
+
+    if (isZh) {
+        if (/jakub|szymon|首席科学家|创始团队|技术领袖|黄金搭档/i.test(combined)) {
+            return '在各种技术光环背后，正是这样专注而不知疲倦的底层攻坚者，为今天整个大模型生态奠定了最坚实的技术基石。';
+        }
+        if (/算力|吉瓦|工厂|电力|芯片|基础设施|基建|硬件|能源/i.test(combined)) {
+            return '这其实释放了一个非常明确的信号：AI竞争的下半场，已经不再只是单纯拼模型参数，更是电力、能源和全栈工业基础设施的硬核较量。';
+        }
+        if (/sora|视频|生成|cameo|chatgpt/i.test(combined)) {
+            return '我个人觉得，这次最大的看点是内容创作门槛的再次降低。不过正如团队提到的，如何平衡创作自由与防止低质内容泛滥，会是整个行业接下来最值得观察的关键考题。';
+        }
+        if (/融资|投资|市场|商业|估值|合作/i.test(combined)) {
+            return '这项进展反映出资本与产业界正在加速形成合力，后续实际业务场景的落地效果非常值得我们持续跟进。';
+        }
+        return '这项新动态反映出技术演进正在进一步提速，对相关领域的实际影响也非常值得我们持续关注。';
+    }
+
+    // English spoken insights
+    if (/jakub|szymon|chief scientist|founding team|technical leader|partnership/i.test(combined)) {
+        return 'Behind the headlines, it is relentless technical leadership and deep engineering rigor that continue to drive these foundational milestones.';
+    }
+    if (/compute|gigawatt|factory|power|chip|infrastructure|datacenter/i.test(combined)) {
+        return 'This signals clearly that the AI frontier is no longer just about algorithms, but a massive contest of power, energy, and full-stack industrial engineering.';
+    }
+    if (/sora|video|generation|cameo|chatgpt/i.test(combined)) {
+        return 'In my view, the biggest takeaway here is the dramatic lowering of creation barriers, while content quality governance will be the key test ahead.';
+    }
+    return 'This development highlights how rapidly the ecosystem is moving from experimental proofs of concept into large-scale execution.';
+}
+
+export function buildNewsVoiceScript(docs: Doc[], options?: ScriptBuilderOptions): VoiceScript {
+    const language = z.enum(['zh', 'en']).parse((options?.language || 'zh').toLowerCase());
+    const isZh = language === 'zh';
+    const voiceProfile = options?.voiceProfile || process.env.VOICEBOX_DEFAULT_PROFILE || 'Robin';
+    const voiceProfileId = options?.voiceProfileId;
+    const defaultEmotion = options?.emotion || 'calm, professional';
+    const defaultEffectPreset = options?.effectPreset;
+    const broadcastDate = options?.date ?? new Date();
+    const dateStr = formatBroadcastDate(broadcastDate, language);
+
+    if (docs.length === 0) {
+        const title = options?.title || (isZh ? '今日AI科技漫谈（暂无更新）' : 'AI Daily Briefing (No Updates)');
+        const segments: VoiceScriptSegment[] = isZh
+            ? [
+                  {
+                      text: `大家好，今天是${dateStr}。欢迎收听由${voiceProfile}为您带来的今日AI科技漫谈。今天科技圈暂无最新的精选动态。`,
+                      gap_ms: 400,
+                      emotion: 'calm',
+                  },
+                  {
+                      text: `感谢您的收听，我是${voiceProfile}，我们下期节目再见。`,
+                      gap_ms: 200,
+                      emotion: 'warm',
+                  },
+              ]
+            : [
+                  {
+                      text: `Hello and welcome. Today is ${dateStr}. This is ${voiceProfile} with your AI Daily Briefing. There are no new featured stories today.`,
+                      gap_ms: 400,
+                      emotion: 'calm',
+                  },
+                  {
+                      text: `Thank you for listening. This is ${voiceProfile}, and have a great day.`,
+                      gap_ms: 200,
+                      emotion: 'warm',
+                  },
+              ];
+
+        return VoiceScriptSchema.parse({
+            title,
+            language,
+            default_profile: voiceProfile,
+            default_profile_id: voiceProfileId,
+            default_emotion: defaultEmotion,
+            default_effect_preset: defaultEffectPreset,
+            speakers: options?.speakers,
+            segments,
+        });
+    }
+
+    const title =
+        options?.title ||
+        (isZh ? `今日AI科技漫谈（共${docs.length}条）` : `AI Daily Briefing (${docs.length} stories)`);
+    const segments: VoiceScriptSegment[] = [];
+
+    let hostName =
+        (options?.speakers?.host?.name as string) ||
+        voiceProfile.replace(/[-_]?(news|voice|bot|profile)$/i, '') ||
+        voiceProfile;
+    if (hostName.length > 0) {
+        hostName = hostName.charAt(0).toUpperCase() + hostName.slice(1);
+    }
+
+    // 1. Opening segment: Conversational podcast hook + overview
+    const countText = isZh ? `${formatChineseCount(docs.length)}条` : `${docs.length}`;
+    const titlesSummary = docs
+        .slice(0, 3)
+        .map((d) => (d.title ?? '').trim())
+        .filter(Boolean)
+        .join('、');
+
+    const introText = isZh
+        ? `大家好，今天是${dateStr}，欢迎收听由${hostName}为您带来的今日AI科技漫谈。今天科技圈有几条非常有看点的大消息，我精选了${countText}最值得关注的重点动态来和大家聊一聊${titlesSummary ? `，包括${titlesSummary}` : ''}。话不多说，我们马上进入今天的正题。`
+        : `Hello and welcome. Today is ${dateStr}. This is ${hostName} with your AI Daily Briefing. In today's episode, we have ${countText} fascinating stories from the AI world. Let's dive straight in.`;
+
+    segments.push({
+        text: normalizeBroadcastText(introText, isZh),
+        gap_ms: 700, // Breathing pause before the first story
+        emotion: 'confident, warm',
+    });
+
+    // 2. Story segments: demonstrative talk with facts + conversational insight
+    docs.forEach((doc, idx) => {
+        const isLast = idx === docs.length - 1 && docs.length > 1;
+        const cleanTitle = (doc.title ?? '').trim();
+        const cleanBody = (doc.body ?? '').trim();
+
+        // Conversational lead
+        const lead = isZh
+            ? idx === 0
+                ? `首先来聊聊大家非常关注的【${cleanTitle}】。`
+                : isLast
+                  ? `最后，来跟大家分享一条很有温度的幕后故事，【${cleanTitle}】。`
+                  : `接着我们把目光转向另一条重要进展，【${cleanTitle}】。`
+            : idx === 0
+              ? `First up today, let's look at ${cleanTitle}.`
+              : isLast
+                ? `And finally today, let's wrap up with ${cleanTitle}.`
+                : `Next, turning our attention to ${cleanTitle}.`;
+
+        const normalizedLead = normalizeBroadcastText(lead, isZh);
+        const normalizedBody = normalizeBroadcastText(cleanBody, isZh);
+        const insight = synthesizePersonalUnderstanding(cleanTitle, cleanBody, isZh);
+        const normalizedInsight = normalizeBroadcastText(insight, isZh);
+
+        const factText = isZh ? `${normalizedLead}${normalizedBody}` : `${normalizedLead} ${normalizedBody}`;
+
+        // Emit facts followed by personal insight commentary
+        segments.push({
+            text: factText,
+            gap_ms: 500, // Natural breathing pause before giving personal perspective
+            emotion: 'professional, clear',
+        });
+        segments.push({
+            text: normalizedInsight,
+            gap_ms: 800, // Distinct 800ms pause before the next story
+            emotion: 'thoughtful, personal',
+        });
+    });
+
+    // 3. Outro segment: Warm, conversational wrap-up
+    const outroText = isZh
+        ? `好了，以上就是今天AI科技漫谈的全部内容。不知道今天这几条消息有没有给您带来一些新的启发？感谢您的收听，我是${hostName}，祝您拥有充实高效的一天，我们下期节目不见不散！`
+        : `That wraps up today's AI Daily Briefing. I hope today's highlights gave you some useful takeaways. Thank you for listening. This is ${hostName}, wishing you a productive and wonderful day. Until next time!`;
+
+    segments.push({
+        text: normalizeBroadcastText(outroText, isZh),
+        gap_ms: 300,
+        emotion: 'warm, friendly',
+    });
+
+    return VoiceScriptSchema.parse({
+        title,
+        language,
+        default_profile: voiceProfile,
+        default_profile_id: voiceProfileId,
+        default_emotion: defaultEmotion,
+        default_effect_preset: defaultEffectPreset,
+        speakers: options?.speakers,
+        segments,
+    });
+}
+
+export function formatVoiceScriptToYaml(script: VoiceScript): string {
+    const lines: string[] = [`title: ${JSON.stringify(script.title)}`, `language: ${JSON.stringify(script.language)}`];
+
+    if (script.default_profile) {
+        lines.push(`default_profile: ${JSON.stringify(script.default_profile)}`);
+    }
+    if (script.default_profile_id) {
+        lines.push(`default_profile_id: ${JSON.stringify(script.default_profile_id)}`);
+    }
+    if (script.default_engine) {
+        lines.push(`default_engine: ${JSON.stringify(script.default_engine)}`);
+    }
+    if (script.default_emotion) {
+        lines.push(`default_emotion: ${JSON.stringify(script.default_emotion)}`);
+    }
+    if (script.default_effect_preset) {
+        lines.push(`default_effect_preset: ${JSON.stringify(script.default_effect_preset)}`);
+    }
+    if (script.default_effects_chain && script.default_effects_chain.length > 0) {
+        lines.push('default_effects_chain:');
+        for (const eff of script.default_effects_chain) {
+            lines.push(`  - type: ${JSON.stringify(eff.type)}`);
+            if (eff.enabled !== undefined) lines.push(`    enabled: ${eff.enabled}`);
+            if (eff.params) {
+                lines.push('    params:');
+                for (const [k, v] of Object.entries(eff.params)) {
+                    lines.push(`      ${k}: ${v}`);
+                }
+            }
+        }
+    }
+
+    if (script.speakers && Object.keys(script.speakers).length > 0) {
+        lines.push('speakers:');
+        for (const [spkKey, spk] of Object.entries(script.speakers)) {
+            lines.push(`  ${spkKey}:`);
+            if (spk.profile) lines.push(`    profile: ${JSON.stringify(spk.profile)}`);
+            if (spk.profile_id) lines.push(`    profile_id: ${JSON.stringify(spk.profile_id)}`);
+            if (spk.name) lines.push(`    name: ${JSON.stringify(spk.name)}`);
+            if (spk.role) lines.push(`    role: ${JSON.stringify(spk.role)}`);
+            if (spk.language) lines.push(`    language: ${JSON.stringify(spk.language)}`);
+            if (spk.engine) lines.push(`    engine: ${JSON.stringify(spk.engine)}`);
+            if (spk.emotion) lines.push(`    emotion: ${JSON.stringify(spk.emotion)}`);
+            if (spk.instruct) lines.push(`    instruct: ${JSON.stringify(spk.instruct)}`);
+            if (spk.effect_preset) lines.push(`    effect_preset: ${JSON.stringify(spk.effect_preset)}`);
+            if (spk.personality !== undefined) lines.push(`    personality: ${spk.personality}`);
+            if (spk.effects_chain && spk.effects_chain.length > 0) {
+                lines.push('    effects_chain:');
+                for (const eff of spk.effects_chain) {
+                    lines.push(`      - type: ${JSON.stringify(eff.type)}`);
+                    if (eff.enabled !== undefined) lines.push(`        enabled: ${eff.enabled}`);
+                    if (eff.params) {
+                        lines.push('        params:');
+                        for (const [k, v] of Object.entries(eff.params)) {
+                            lines.push(`          ${k}: ${v}`);
+                        }
+                    }
+                }
+            }
+        }
+    }
+
+    lines.push('segments:');
+    for (const segment of script.segments) {
+        lines.push(`  - text: ${JSON.stringify(segment.text)}`);
+        if (segment.speaker) {
+            lines.push(`    speaker: ${JSON.stringify(segment.speaker)}`);
+        }
+        if (segment.profile) {
+            lines.push(`    profile: ${JSON.stringify(segment.profile)}`);
+        }
+        if (segment.profile_id) {
+            lines.push(`    profile_id: ${JSON.stringify(segment.profile_id)}`);
+        }
+        if (segment.emotion) {
+            lines.push(`    emotion: ${JSON.stringify(segment.emotion)}`);
+        }
+        if (segment.instruct) {
+            lines.push(`    instruct: ${JSON.stringify(segment.instruct)}`);
+        }
+        if (segment.gap_ms !== undefined) {
+            lines.push(`    gap_ms: ${segment.gap_ms}`);
+        }
+        if (segment.seed !== undefined) {
+            lines.push(`    seed: ${segment.seed}`);
+        }
+        if (segment.personality !== undefined) {
+            lines.push(`    personality: ${segment.personality}`);
+        }
+        if (segment.effect_preset) {
+            lines.push(`    effect_preset: ${JSON.stringify(segment.effect_preset)}`);
+        }
+        if (segment.effects_chain && segment.effects_chain.length > 0) {
+            lines.push('    effects_chain:');
+            for (const eff of segment.effects_chain) {
+                lines.push(`      - type: ${JSON.stringify(eff.type)}`);
+                if (eff.enabled !== undefined) lines.push(`        enabled: ${eff.enabled}`);
+                if (eff.params) {
+                    lines.push('        params:');
+                    for (const [k, v] of Object.entries(eff.params)) {
+                        lines.push(`          ${k}: ${v}`);
+                    }
+                }
+            }
+        }
+    }
+    return `${lines.join('\n')}\n`;
+}

package/plugins/generations/dailynews-gen/tsconfig.json

@@ -0,0 +1,4 @@
+{
+    "extends": "../../../tooling/typescript/base.json",
+    "include": ["src", "tests"]
+}

package/plugins/generations/voice-gen/src/concat.ts

@@ -157,13 +157,15 @@ export function concatWavs(wavBuffers: Uint8Array[], gapsMs: number[] = []): Uin
         if (!wav) {
             continue;
         }
-        const gap = gapsMs[i] ?? 0;
-        if (gap > 0) {
-            const numGapBytes = Math.floor(first.sampleRate * (gap / 1000)) * first.blockAlign;
-            if (numGapBytes > 0) {
-                const silence = new Uint8Array(numGapBytes);
-                pieces.push(silence);
-                totalDataLength += numGapBytes;
+        if (i > 0) {
+            const gap = gapsMs[0] === 0 ? (gapsMs[i] ?? 0) : (gapsMs[i - 1] ?? 0);
+            if (gap > 0) {
+                const numGapBytes = Math.floor(first.sampleRate * (gap / 1000)) * first.blockAlign;
+                if (numGapBytes > 0) {
+                    const silence = new Uint8Array(numGapBytes);
+                    pieces.push(silence);
+                    totalDataLength += numGapBytes;
+                }
             }
         }
         pieces.push(wav.dataBytes);

package/plugins/generations/voice-gen/src/index.ts

@@ -4,10 +4,14 @@ import { type Content, ContentSchema, type Doc, DocListSchema } from '@gobing-ai
 import { createNodeFileSystem } from '@gobing-ai/ts-runtime';
 import { echoError } from '@gobing-ai/ts-utils';
 import { concatWavs } from './concat';
+import { isMp3Requested, type Mp3Transcoder, transcodeWavToMp3 } from './mp3';
+import { auditVoiceSegments } from './qc';
 import { createVoiceboxClient, type VoiceboxClient, type VoiceboxGenerateBody } from './voicebox-client';
 import {
     mergeVoiceScripts,
     parseDocToVoiceScript,
+    resolveEffectsChain,
+    resolveInstruct,
     VOICEBOX_CHUNK_CHARS_DEFAULT,
     VOICEBOX_CHUNK_CHARS_MAX,
     VOICEBOX_CHUNK_CHARS_MIN,
@@ -18,6 +22,8 @@ import {
 } from './voicescript';
 
 export * from './concat';
+export * from './mp3';
+export * from './qc';
 export * from './voicebox-client';
 export * from './voicescript';
 
@@ -35,15 +41,19 @@ export async function processGeneratorIO(
     inputPath: string,
     outputPath: string,
     clientOverride?: VoiceboxClient,
+    transcodeMp3?: Mp3Transcoder,
 ): Promise<void> {
     const fs = createNodeFileSystem();
     const outDir = dirname(outputPath);
     const outStem = basename(outputPath, extname(outputPath));
     const audioPath = resolve(join(outDir, `${outStem}.wav`));
+    const mp3Path = resolve(join(outDir, `${outStem}.mp3`));
+    const wantMp3 = isMp3Requested();
 
-    // Delete existing output and sibling WAV at start
+    // Delete existing output and sibling audio at start
     await fs.deleteFile(outputPath);
     await fs.deleteFile(audioPath);
+    await fs.deleteFile(mp3Path);
 
     let docs: Doc[];
     try {
@@ -76,21 +86,30 @@ export async function processGeneratorIO(
         // 4. Generate each segment
         const segmentWavs: Uint8Array[] = [];
         const segmentGaps: number[] = [];
-        const segmentMetadata: Array<{ generationId: string; profile: string; duration: number }> = [];
+        const segmentMetadata: Array<{
+            generationId: string;
+            profile: string;
+            speaker?: string;
+            emotion?: string;
+            duration: number;
+        }> = [];
         let totalDuration = 0;
 
         for (const segment of combinedScript.segments) {
             const speakerConfig = segment.speaker ? combinedScript.speakers?.[segment.speaker] : undefined;
             const profileTarget =
+                segment.profile_id ??
                 segment.profile ??
+                speakerConfig?.profile_id ??
                 speakerConfig?.profile ??
+                combinedScript.default_profile_id ??
                 combinedScript.default_profile ??
                 envDefaultProfile ??
                 (typeof docs[0]?.metadata?.voiceProfile === 'string' ? docs[0].metadata.voiceProfile : undefined);
 
             if (!profileTarget) {
                 throw new Error(
-                    'No voice profile specified for segment (set profile, speaker profile, default_profile, or VOICEBOX_DEFAULT_PROFILE)',
+                    'No voice profile specified for segment (set profile, profile_id, speaker profile, default_profile, default_profile_id, or VOICEBOX_DEFAULT_PROFILE)',
                 );
             }
 
@@ -98,7 +117,21 @@ export async function processGeneratorIO(
 
             const engine = segment.engine ?? speakerConfig?.engine ?? combinedScript.default_engine;
             const language = segment.language ?? speakerConfig?.language ?? combinedScript.language ?? 'en';
-            const instruct = segment.instruct ?? speakerConfig?.instruct;
+            const rawInstruct = segment.instruct ?? speakerConfig?.instruct;
+            const rawEmotion = segment.emotion ?? speakerConfig?.emotion ?? combinedScript.default_emotion;
+            const instruct = resolveInstruct(rawInstruct, rawEmotion);
+
+            const personality = segment.personality ?? speakerConfig?.personality ?? false;
+            const seed = segment.seed ?? null;
+
+            const effects_chain =
+                segment.effects_chain ??
+                resolveEffectsChain(undefined, segment.effect_preset) ??
+                speakerConfig?.effects_chain ??
+                resolveEffectsChain(undefined, speakerConfig?.effect_preset) ??
+                combinedScript.default_effects_chain ??
+                resolveEffectsChain(undefined, combinedScript.default_effect_preset) ??
+                null;
 
             const envChunk = process.env.VOICEBOX_MAX_CHUNK_CHARS
                 ? parseInt(process.env.VOICEBOX_MAX_CHUNK_CHARS, 10)
@@ -131,8 +164,10 @@ export async function processGeneratorIO(
                 instruct,
                 max_chunk_chars,
                 crossfade_ms,
-                personality: false,
+                personality,
                 normalize: true,
+                seed,
+                effects_chain,
             };
 
             const { id } = await client.generate(generateBody);
@@ -154,6 +189,8 @@ export async function processGeneratorIO(
             segmentMetadata.push({
                 generationId: id,
                 profile: resolvedProfile.name,
+                speaker: segment.speaker,
+                emotion: rawEmotion,
                 duration: segDuration,
             });
         }
@@ -163,18 +200,34 @@ export async function processGeneratorIO(
         await fs.ensureDir(outDir);
         await Bun.write(audioPath, concatenatedWav);
 
+        const metadata: Record<string, unknown> = {
+            generator: 'kk:voice-gen',
+            audioPath,
+            duration: totalDuration,
+            segments: segmentMetadata,
+        };
+
+        // 6. Quality Control Audit
+        const qc = await auditVoiceSegments(combinedScript, segmentWavs, segmentMetadata, client);
+        metadata.qc = qc;
+
+        if (process.env.VOICE_GEN_FAIL_ON_QC === 'true' && !qc.passed) {
+            throw new Error(`Voice Quality Control failed (${qc.overallScore}/100):\n${qc.criticalIssues.join('\n')}`);
+        }
+
+        if (wantMp3) {
+            const transcode = transcodeMp3 ?? transcodeWavToMp3;
+            await transcode(audioPath, mp3Path);
+            metadata.mp3Path = mp3Path;
+        }
+
         // 6. Build Content
         const content: Content = {
             title: combinedScript.title || docs[0]?.title || 'Generated voice',
             body: Bun.YAML.stringify(combinedScript),
             format: 'audio',
             references: [],
-            metadata: {
-                generator: 'kk:voice-gen',
-                audioPath,
-                duration: totalDuration,
-                segments: segmentMetadata,
-            },
+            metadata,
         };
 
         const validatedContent = ContentSchema.parse(content);
@@ -182,6 +235,7 @@ export async function processGeneratorIO(
     } catch (err: unknown) {
         await fs.deleteFile(outputPath);
         await fs.deleteFile(audioPath);
+        await fs.deleteFile(mp3Path);
         throw err;
     }
 }