@kmchai/cortico-world-voice 0.2.1 → 0.3.2

package/CHANGELOG.md

@@ -1,3 +1,20 @@
+## 0.3.2
+
+- IndexTTS2 的模型提示词明确:`emotion` / `instruct` 传自然语言,由服务映射到 `emotion_text`;不要在正文写 `emotion_text` 或方括号标签。服务 0.4.2 起这两个 options 不再由模型页覆盖。
+
+## 0.3.1
+
+- 自动朗读不再把 `end_turn` 当成实质工具调用。模型说完正文再调 `end_turn` 结束这一轮时,正文照常按概率朗读;仍会压掉与 `web_search` 等实际工具同轮出现的中间正文。
+- 网关 0.1.1 已把 World 的 `outputTap` 扇进主 session;此前网关 bot 里概率朗读拿不到正文。
+
+## 0.3.0
+
+- 系统 Config tab 移除「音色与情绪」组;音色、情绪控制、默认情绪、默认表演指令、语速与副语言标签移到「模型」页。
+- 朗读风格按模型保存进 `worlds.voice.modelStyles[model]`;切换模型时面板读回该模型的音色、风格与参数,不用先保存。旧全局字段保留为没有模型设置时的回落。
+- `voice_style` 在选定模型后只写当前模型的风格;没有模型时仍写全局回落值。
+- 服务返回的 Ming 3.0 模型参数(模板、语言、指令语言、音高、音量、生成与分块参数)在模型页显示;IndexTTS2 的中文参数名继续保留。
+- 文本过滤、合成与 `tts.voice` 事件都使用当前模型生效的风格。
+
 ## 0.2.1
 
 - 新增 World 页的「模型」面板:模型与音色在面板里选,不再占系统 Config tab 的字段。

package/README.md

@@ -34,7 +34,6 @@ corepack pnpm add "D:/path/to/cortico-world-voice" --ignore-workspace
       "serviceUrl": "http://127.0.0.1:9883",
       "serviceTokenSecret": "",
       "model": "ming3-tts",
-      "voice": "4016",
       "playback": "local",
       "autoSpeak": true,
       "autoProbability": 0.3
@@ -47,7 +46,7 @@ corepack pnpm add "D:/path/to/cortico-world-voice" --ignore-workspace
 
 ## 模型面板
 
-World 页的「模型」tab 选模型与音色,并按服务 `/v1/models` 返回的参数元数据渲染字段。保存写进 `worlds.voice.modelOptions[model]`,合成时作为 `options` 发给服务;空参数表删掉该模型的条目。模型专属的文本与情绪写法显示在同页的「模型提示词」折叠区。系统 Config tab 只留服务地址、风格、触发、过滤与输出。
+World 页的「模型」tab 选模型与音色,并按服务 `/v1/models` 返回的参数元数据渲染字段。同一页的「声音与情绪」按模型保存情绪控制方式、默认情绪、默认表演指令、语速与副语言标签策略;切换模型时读回该模型自己的设置。模型设置存 `worlds.voice.modelStyles[model]`,模型参数存 `worlds.voice.modelOptions[model]`;合成时作为 `options` 发给服务。模型专属的文本与情绪写法显示在同页的「模型提示词」折叠区。系统 Config tab 只留服务地址、触发、过滤与输出。
 
 ## 工具
 
@@ -78,11 +77,11 @@ World 页的「模型」tab 选模型与音色,并按服务 `/v1/models` 返回
 
 只想响应工具与显式请求、不要随机朗读:保留 `autoSpeak: true`,把 `autoProbability` 设为 `0`。把 `autoSpeak` 整个关掉会让 tap 不再挂载,显式关键词也就看不到正文了;这时只保留 `voice_speak` 工具触发。
 
-自动朗读依赖 Persona 把已挂载 World 的 `outputTap()` 扇进主 session;仓内 `cormini` 与 Coopanion 都这样做。只想要工具触发时,把 `autoSpeak` 关掉即可。
+自动朗读依赖 Persona 把已挂载 World 的 `outputTap()` 扇进主 session;`cormini`、Coopanion 与 `Kmch AI · 网关` 0.1.1 起都这样做。只想要工具触发时,把 `autoSpeak` 关掉即可。模型说完正文再调 `end_turn` 不影响朗读;与搜索等实际工具同轮的中间正文仍不念。
 
 ## 情绪、语速与标签
 
-语音只把语义字段发给服务,由服务的模型 profile 决定怎么落到具体模型:
+情绪与语速设置按模型保存在「模型」页,不再出现在系统 Config tab。语音只把语义字段发给服务,由服务的模型 profile 决定怎么落到具体模型:
 
 | `emotionMode` | 语音发送 | `audiocpp-ming3` profile 的落点 |
 |---|---|---|
@@ -92,12 +91,12 @@ World 页的「模型」tab 选模型与音色,并按服务 `/v1/models` 返回
 | `tags` | 不发情绪参数,只用正文标签。 | 正文标签。 |
 | `none` | 都不发。 | — |
 
-OpenAI 兼容条目把 `instructions` / `emotion` / `speed` 直接放进请求。其他模型可在服务面板里覆盖 profile 的字段映射,语音不用改配置。
+OpenAI 兼容条目把 `instructions` / `emotion` / `speed` 直接放进请求。IndexTTS2 把 `instructions` 或 `emotion` 映射成 `use_emotion_text: true` 与 `emotion_text`;这两个参数由语音的「默认情绪 / 默认表演指令」或单句 `emotion` / `instruct` 驱动,不在模型参数面板里单独出现。其他模型可在服务面板里覆盖 profile 的字段映射,语音不用改配置。
 
 `response_format` 每次都带;`model` 与 `voice` 配了才带,工具参数里的音色优先于配置默认值。
 
-- 情绪来源:工具参数 > 正文里的 `[emotion:开心]` 标记 > `defaultEmotion`。
-- `voice_style` 读写长期风格(情绪、指令、音色、语速),写回 `worlds.voice`。
+- 情绪来源:工具参数 > 正文里的 `[emotion:开心]` 标记 > 当前模型的 `defaultEmotion`。
+- `voice_style` 读写当前模型的长期风格(情绪、指令、音色、语速),写进 `modelStyles[model]`;还没选模型时写到全局回落值。
 - 副语言标签:`[Breathing] [Cough] [Sigh] [Laughter] [Surprise-ah] [Surprise-oh] [Dissatisfaction-hnn] [Uhm] [Confirmation-en] [Question-ei] [Question-ah] [Question-en] [Question-oh] [Question-huh]`。
   `paralinguistic: auto` 只在当前模型支持标签(Ming 3.0 系)时保留,IndexTTS2 与通用模型剥掉;`keep` 一律保留,`strip` 一律删除。表外 `[...]` 会从朗读文本里剥掉,防止被念出来。
 
@@ -141,7 +140,7 @@ OpenAI 兼容条目把 `instructions` / `emotion` / `speed` 直接放进请求
 | `httpApi` / `httpPort` / `httpToken` / `httpPlay` | 删除;服务的网关取代本地语音接口。 |
 | `singUrl` / `singToken`、`voice_sing`、`【唱】` | 删除;唱歌由服务的 `/v1/audio/sing` 与网关接入。 |
 | `model` / `voice` / `format` / `timeoutMs` | 保留,含义不变。 |
-| `emotionMode` / `defaultEmotion` / `defaultInstruct` / `speed` / `paralinguistic` | 保留;语音只发语义字段,由服务的模型 profile 映射到具体模型。 |
+| `emotionMode` / `defaultEmotion` / `defaultInstruct` / `speed` / `paralinguistic` | 移到「模型」页,按模型保存;旧全局值仍作为没有模型设置时的回落。 |
 | 触发、过滤、输出配置 | 保留。 |
 
 ## 验证
@@ -153,7 +152,7 @@ pnpm smoke:player out.wav                 # 本机播放器,音量 0
 pnpm smoke:chain http://127.0.0.1:9883    # 触发→过滤→经服务合成→tts.voice 全链路
 ```
 
-装进 Cortico 后应看到:World 页多一张「Kmch AI · 语音」卡,服务灯、队列灯与 `service` / `model` / `voice` 徽标;「模型」tab 能选模型与音色、按服务参数渲染字段并保存;服务在线时灯亮;让 bot 调 `voice_speak` 能听见;说「念给我听」会绕过概率;`playback: event` 时时间线出现 `tts.voice`。
+装进 Cortico 后应看到:World 页多一张「Kmch AI · 语音」卡,服务灯、队列灯与 `service` / `model` / `voice` 徽标;「模型」tab 能选模型与音色,按模型调整声音与情绪,并按服务参数渲染字段、保存;服务在线时灯亮;让 bot 调 `voice_speak` 能听见;说「念给我听」会绕过概率;`playback: event` 时时间线出现 `tts.voice`。
 
 ## 许可
 

package/dist/webui.js

@@ -1 +1 @@
-var W={zh:{title:"\u6A21\u578B",model:"\u6A21\u578B",voice:"\u97F3\u8272",refresh:"\u5237\u65B0",save:"\u4FDD\u5B58",saved:"\u5DF2\u4FDD\u5B58",loading:"\u8BFB\u53D6\u4E2D\u2026",promptTitle:"\u6A21\u578B\u63D0\u793A\u8BCD"},en:{title:"\u6A21\u578B",model:"Model",voice:"Voice",refresh:"Refresh",save:"Save",saved:"Saved",loading:"Loading\u2026",promptTitle:"Model prompt"}};function O(e){return e.toLowerCase().startsWith("zh")?W.zh:W.en}var U={emotion_alpha:"\u60C5\u7EEA\u5F3A\u5EA6",emotion_text:"\u60C5\u7EEA\u63CF\u8FF0",use_emotion_text:"\u4F7F\u7528\u6587\u5B57\u60C5\u7EEA",use_random_emotion:"\u968F\u673A\u60C5\u7EEA",duration_scale:"\u65F6\u957F\u7F29\u653E",interval_silence_ms:"\u5206\u6BB5\u9759\u97F3",do_sample:"\u91C7\u6837",top_p:"top_p",top_k:"top_k",temperature:"\u6E29\u5EA6",length_penalty:"\u957F\u5EA6\u60E9\u7F5A",num_beams:"beam \u6570",repetition_penalty:"\u91CD\u590D\u60E9\u7F5A",max_tokens:"\u6700\u5927 token",seed:"\u79CD\u5B50",text_chunk_size:"\u6587\u672C\u5206\u5757\u5927\u5C0F",text_chunk_mode:"\u5206\u5757\u6A21\u5F0F"};function j(e,t){return Object.prototype.hasOwnProperty.call(t,e.name)?t[e.name]:e.default}function C(e){return e==null?"":Array.isArray(e)?e.map(t=>String(t)).join(", "):String(e)}function H(e){return typeof e=="number"&&Number.isFinite(e)?String(e):typeof e=="string"&&e.trim()&&Number.isFinite(Number(e))?e.trim():""}function I(e,t,s){let i=e.ui,r=i.h("div","vx-field"),m=i.h("div","vx-head"),p=U[t.name];p&&p!==t.name?m.append(i.h("span","vx-label",p),i.h("span","vx-name",t.name)):m.append(i.h("span","vx-label",t.name)),r.append(m);let f=j(t,s),u;switch(t.type){case"bool":{let a=i.checkbox("",{checked:f===!0});r.append(a.el),u=()=>a.checked;break}case"enum":{let a=t.values??[],d=C(f),c=[{value:"",label:"\u2014"},...a.map(k=>({value:k,label:k}))];d&&!a.includes(d)&&c.push({value:d,label:d});let b=i.select({options:c,value:d});r.append(b),u=()=>b.value||void 0;break}case"int":case"float":{let a=i.input({type:"number",value:H(f)});t.min!==void 0&&(a.min=String(t.min)),t.max!==void 0&&(a.max=String(t.max)),a.step=t.type==="int"?"1":"any",r.append(a),u=()=>{let d=a.value.trim();if(!d)return;let c=Number(d);if(Number.isFinite(c))return t.type==="int"?Math.trunc(c):c};break}case"string_list":case"path_list":{let a=i.input({type:"text",value:C(f)});r.append(a),u=()=>{let d=a.value.split(/[,，]/).map(c=>c.trim()).filter(Boolean);return d.length>0?d:void 0};break}default:{let a=i.input({type:"text",value:C(f)});r.append(a),u=()=>a.value.trim()||void 0;break}}return t.description&&r.append(i.h("div","vx-desc",t.description)),{param:t,el:r,read:u}}function X(e){let t=O(e.language),s=e.ui,i=null,r="",m=[],p=!1,f=s.sheet({title:t.title}),u=s.h("div","vx-top"),a=s.select({onChange:()=>B(a.value)}),d=s.select({}),c=s.h("label","vx-pick");c.append(s.h("span","vx-pick-label",t.model),a);let b=s.h("label","vx-pick");b.append(s.h("span","vx-pick-label",t.voice),d);let k=s.button(t.refresh,{size:"sm",onClick:()=>{E("refresh")}});u.append(c,b,k);let y=s.h("div","vx-params"),S=s.button(t.save,{variant:"primary",size:"sm",onClick:()=>{F()}}),w=s.msgline(),T=s.actions();T.append(S),f.body.append(u,y,T,w),e.root.append(f.el);let _=s.foldSheet("voice-model-prompt",{title:t.promptTitle,defaultOpen:!1}),z=s.h("pre","vx-prompt");_.body.append(z),e.root.append(_.el);function x(n,o=n){let l=s.h("option");return l.value=n,l.textContent=o,l}function h(n,o){w.textContent=n,w.classList.toggle("bad",o)}function v(){let n=i!==null&&i.models.length>0;a.disabled=p||!n,d.disabled=p||!n||!r,S.disabled=p||!n||!r,k.disabled=p}function L(){let o=i?.models.find(M=>M.id===r)?.voices??[],l=i?.model===r?i.voice:"",g=[...o];l&&!g.includes(l)&&g.unshift(l);let P=g.map(M=>x(M));l||P.unshift(x("","\u2014")),d.replaceChildren(...P.length>0?P:[x("","\u2014")]),d.value=l}function R(){let n=i?.models.find(l=>l.id===r),o=i&&i.model===r?i.options:{};m=[],y.replaceChildren();for(let l of n?.parameters??[]){let g=I(e,l,o);m.push(g),y.append(g.el)}}function B(n){r=n,L(),R(),v()}function A(n){i=n,r=n.model,N(),R(),z.textContent=n.prompt,_.note.textContent=n.model||"\u2014",v(),h(n.error??"",n.error!==void 0)}function N(){let n=i?.models??[],o=n.map(l=>x(l.id,l.label||l.id));!r&&n.length>0?o.unshift(x("","\u2014")):r&&!n.some(l=>l.id===r)&&o.unshift(x(r,r)),a.replaceChildren(...o),a.value=r,L()}function V(){let n={};for(let o of m){let l=o.read();l!==void 0&&(n[o.param.name]=l)}return n}async function E(n){if(!p){p=!0,v(),h(t.loading,!1);try{let o=await e.invoke(n,[]);e.signal.aborted||A(o)}catch(o){e.signal.aborted||h(o instanceof Error?o.message:String(o),!0)}finally{p=!1,e.signal.aborted||v()}}}async function F(){if(p||!i)return;let n={model:r,voice:d.value,options:V()};p=!0,v(),h("",!1);try{let o=await e.invoke("save",[n]);e.signal.aborted||(A(o),h(t.saved,!1))}catch(o){e.signal.aborted||h(o instanceof Error?o.message:String(o),!0)}finally{p=!1,e.signal.aborted||v()}}v(),E("state")}var q={panels:{model:{mount:X}}},Y=q;export{Y as default};
+var B={zh:{title:"\u6A21\u578B",model:"\u6A21\u578B",voice:"\u97F3\u8272",refresh:"\u5237\u65B0",save:"\u4FDD\u5B58",saved:"\u5DF2\u4FDD\u5B58",loading:"\u8BFB\u53D6\u4E2D\u2026",promptTitle:"\u6A21\u578B\u63D0\u793A\u8BCD",styleTitle:"\u58F0\u97F3\u4E0E\u60C5\u7EEA",paramsTitle:"\u6A21\u578B\u53C2\u6570",emotionMode:"\u60C5\u7EEA\u63A7\u5236",defaultEmotion:"\u9ED8\u8BA4\u60C5\u7EEA",defaultInstruct:"\u9ED8\u8BA4\u8868\u6F14\u6307\u4EE4",speed:"\u8BED\u901F",paralinguistic:"\u526F\u8BED\u8A00\u6807\u7B7E",emotionModeOptions:[{value:"instruct",label:"\u81EA\u7136\u8BED\u8A00"},{value:"categorical",label:"\u60C5\u7EEA\u6807\u7B7E"},{value:"basic",label:"\u8BED\u901F"},{value:"tags",label:"\u6B63\u6587\u6807\u7B7E"},{value:"none",label:"\u4E0D\u53D1\u9001"}],paralinguisticOptions:[{value:"auto",label:"\u81EA\u52A8"},{value:"keep",label:"\u4FDD\u7559"},{value:"strip",label:"\u5220\u9664"}]},en:{title:"\u6A21\u578B",model:"Model",voice:"Voice",refresh:"Refresh",save:"Save",saved:"Saved",loading:"Loading\u2026",promptTitle:"Model prompt",styleTitle:"Voice and emotion",paramsTitle:"Model parameters",emotionMode:"Emotion control",defaultEmotion:"Default emotion",defaultInstruct:"Default direction",speed:"Speed",paralinguistic:"Paralinguistic tags",emotionModeOptions:[{value:"instruct",label:"Natural language"},{value:"categorical",label:"Emotion label"},{value:"basic",label:"Speed"},{value:"tags",label:"Inline tags"},{value:"none",label:"Send nothing"}],paralinguisticOptions:[{value:"auto",label:"Auto"},{value:"keep",label:"Keep"},{value:"strip",label:"Strip"}]}};function j(e){return e.toLowerCase().startsWith("zh")?B.zh:B.en}var D={template_name:"\u6A21\u677F",language:"\u8BED\u8A00",instruction_schema:"\u6307\u4EE4\u8BED\u8A00",pitch:"\u97F3\u9AD8",volume:"\u97F3\u91CF",guidance_scale:"\u5F15\u5BFC\u5F3A\u5EA6",sigma:"\u566A\u58F0\u5C3A\u5EA6",raw_text:"\u6574\u6BB5\u53D1\u9001",trim_silence:"\u4FEE\u526A\u9759\u97F3",emotion_alpha:"\u60C5\u7EEA\u5F3A\u5EA6",emotion_text:"\u60C5\u7EEA\u63CF\u8FF0",use_emotion_text:"\u4F7F\u7528\u6587\u5B57\u60C5\u7EEA",use_random_emotion:"\u968F\u673A\u60C5\u7EEA",duration_scale:"\u65F6\u957F\u7F29\u653E",interval_silence_ms:"\u5206\u6BB5\u9759\u97F3",do_sample:"\u91C7\u6837",top_p:"top_p",top_k:"top_k",temperature:"\u6E29\u5EA6",length_penalty:"\u957F\u5EA6\u60E9\u7F5A",num_beams:"beam \u6570",repetition_penalty:"\u91CD\u590D\u60E9\u7F5A",max_tokens:"\u6700\u5927 token",seed:"\u79CD\u5B50",text_chunk_size:"\u6587\u672C\u5206\u5757\u5927\u5C0F",text_chunk_mode:"\u5206\u5757\u6A21\u5F0F"};var Q={voice:"",emotionMode:"instruct",defaultEmotion:"",defaultInstruct:"",speed:1,paralinguistic:"auto"};function Z(e,t){return Object.prototype.hasOwnProperty.call(t,e.name)?t[e.name]:e.default}function O(e){return e==null?"":Array.isArray(e)?e.map(t=>String(t)).join(", "):String(e)}function $(e){return typeof e=="number"&&Number.isFinite(e)?String(e):typeof e=="string"&&e.trim()&&Number.isFinite(Number(e))?e.trim():""}function ee(e,t,l){let n=e.ui,i=n.h("div","vx-field"),f=n.h("div","vx-head"),p=D[t.name];p&&p!==t.name?f.append(n.h("span","vx-label",p),n.h("span","vx-name",t.name)):f.append(n.h("span","vx-label",t.name)),i.append(f);let m=Z(t,l),c;switch(t.type){case"bool":{let a=n.checkbox("",{checked:m===!0});i.append(a.el),c=()=>a.checked;break}case"enum":{let a=t.values??[],r=O(m),u=[{value:"",label:"\u2014"},...a.map(y=>({value:y,label:y}))];r&&!a.includes(r)&&u.push({value:r,label:r});let x=n.select({options:u,value:r});i.append(x),c=()=>x.value||void 0;break}case"int":case"float":{let a=n.input({type:"number",value:$(m)});t.min!==void 0&&(a.min=String(t.min)),t.max!==void 0&&(a.max=String(t.max)),a.step=t.type==="int"?"1":"any",i.append(a),c=()=>{let r=a.value.trim();if(!r)return;let u=Number(r);if(Number.isFinite(u))return t.type==="int"?Math.trunc(u):u};break}case"string_list":case"path_list":{let a=n.input({type:"text",value:O(m)});i.append(a),c=()=>{let r=a.value.split(/[,，]/).map(u=>u.trim()).filter(Boolean);return r.length>0?r:void 0};break}default:{let a=n.input({type:"text",value:O(m)});i.append(a),c=()=>a.value.trim()||void 0;break}}return t.description&&i.append(n.h("div","vx-desc",t.description)),{param:t,el:i,read:c}}function M(e,t,l){let n=e.ui.h("div","vx-field"),i=e.ui.h("div","vx-head");return i.append(e.ui.h("span","vx-label",t)),n.append(i,l),n}function te(e,t,l){let n=e.ui,i=n.h("div","vx-params"),f=n.select({options:t.emotionModeOptions.map(({value:r,label:u})=>({value:r,label:u})),value:l.emotionMode}),p=n.input({type:"text",value:l.defaultEmotion}),m=n.input({type:"text",value:l.defaultInstruct}),c=n.input({type:"number",value:String(l.speed)});c.min="0.5",c.max="2",c.step="0.05";let a=n.select({options:t.paralinguisticOptions.map(({value:r,label:u})=>({value:r,label:u})),value:l.paralinguistic});return i.append(M(e,t.emotionMode,f),M(e,t.defaultEmotion,p),M(e,t.defaultInstruct,m),M(e,t.speed,c),M(e,t.paralinguistic,a)),{el:i,read:()=>{let r=Number(c.value);return{voice:l.voice,emotionMode:f.value,defaultEmotion:p.value.trim(),defaultInstruct:m.value.trim(),speed:Number.isFinite(r)?Math.min(2,Math.max(.5,r)):l.speed,paralinguistic:a.value}}}}function ne(e){let t=j(e.language),l=e.ui,n=null,i="",f=[],p=!1,m=l.sheet({title:t.title}),c=l.h("div","vx-top"),a=l.select({onChange:()=>K(a.value)}),r=l.select({}),u=l.h("label","vx-pick");u.append(l.h("span","vx-pick-label",t.model),a);let x=l.h("label","vx-pick");x.append(l.h("span","vx-pick-label",t.voice),r);let y=l.button(t.refresh,{size:"sm",onClick:()=>{H("refresh")}});c.append(u,x,y);let Y=l.section(t.styleTitle),V=l.h("div","vx-section"),k=l.section(t.paramsTitle);k.hidden=!0;let S=l.h("div","vx-params"),L=l.button(t.save,{variant:"primary",size:"sm",onClick:()=>{J()}}),P=l.msgline(),F=l.actions();F.append(L),m.body.append(c,Y,V,k,S,F,P),e.root.append(m.el);let _=l.foldSheet("voice-model-prompt",{title:t.promptTitle,defaultOpen:!1}),I=l.h("pre","vx-prompt");_.body.append(I),e.root.append(_.el);let T=null;function g(o,s=o){let d=l.h("option");return d.value=o,d.textContent=s,d}function b(o,s){P.textContent=o,P.classList.toggle("bad",s)}function v(){let o=n!==null&&n.models.length>0;a.disabled=p||!o,r.disabled=p||!o||!i,L.disabled=p||!o||!i,y.disabled=p}function w(o){return n?.settings?.[o]}function z(o){return w(o)?.style??(n?.model===o?n.style:void 0)??Q}function A(){let s=n?.models.find(E=>E.id===i)?.voices??[],d=w(i)?.voice??(n?.model===i?n.voice:""),h=[...s];d&&!h.includes(d)&&h.unshift(d);let C=h.map(E=>g(E));d||C.unshift(g("","\u2014")),r.replaceChildren(...C.length>0?C:[g("","\u2014")]),r.value=d}function R(){T=te(e,t,z(i)),V.replaceChildren(T.el)}function W(){let o=n?.models.find(d=>d.id===i),s=w(i)?.options??(n?.model===i?n.options:{});f=[],S.replaceChildren();for(let d of o?.parameters??[]){let h=ee(e,d,s);f.push(h),S.append(h.el)}k.hidden=(o?.parameters?.length??0)===0}function N(){let o=n?.prompts?.[i]??(n?.model===i?n.prompt:"");I.textContent=o,_.note.textContent=i||"\u2014"}function K(o){i=o,A(),R(),W(),N(),v()}function U(o){n=o,i=o.model,X(),R(),W(),N(),v(),b(o.error??"",o.error!==void 0)}function X(){let o=n?.models??[],s=o.map(d=>g(d.id,d.label||d.id));!i&&o.length>0?s.unshift(g("","\u2014")):i&&!o.some(d=>d.id===i)&&s.unshift(g(i,i)),a.replaceChildren(...s),a.value=i,A()}function q(){let o={};for(let s of f){let d=s.read();d!==void 0&&(o[s.param.name]=d)}return o}function G(){return{...T?.read()??z(i),voice:r.value}}async function H(o){if(!p){p=!0,v(),b(t.loading,!1);try{let s=await e.invoke(o,[]);e.signal.aborted||U(s)}catch(s){e.signal.aborted||b(s instanceof Error?s.message:String(s),!0)}finally{p=!1,e.signal.aborted||v()}}}async function J(){if(p||!n)return;let o={model:i,voice:r.value,style:G(),options:q()};p=!0,v(),b("",!1);try{let s=await e.invoke("save",[o]);e.signal.aborted||(U(s),b(t.saved,!1))}catch(s){e.signal.aborted||b(s instanceof Error?s.message:String(s),!0)}finally{p=!1,e.signal.aborted||v()}}v(),H("state")}var ie={panels:{model:{mount:ne}}},de=ie;export{de as default};

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@kmchai/cortico-world-voice",
-  "version": "0.2.1",
+  "version": "0.3.2",
   "type": "module",
   "description": "Cortico World: voice tools, probabilistic reading, emotion and paralinguistic tags, text filtering and local playback; speech is synthesized by the Kmch AI service over HTTP",
   "author": "kmchai",

package/src/config.ts

@@ -1,7 +1,7 @@
 /** `worlds.voice` 的默认值、配置组与界面文案。 */
 import type { ConfigGroup } from 'cortico/core/types.ts';
 import { pick, type Language, type LanguageTable } from 'cortico/core/language.ts';
-import type { VoiceConfigSection } from './types.ts';
+import type { VoiceConfigSection, VoiceModelStyle } from './types.ts';
 
 export const VOICE_ID = 'voice';
 /** World 定义、WebUI 与配置页共用同一个显示名。 */
@@ -15,6 +15,7 @@ export const VOICE_DEFAULTS: VoiceConfigSection = {
   model: '',
   voice: '',
   modelOptions: {},
+  modelStyles: {},
   format: 'wav',
   timeoutMs: 180_000,
 
@@ -49,6 +50,27 @@ export const VOICE_DEFAULTS: VoiceConfigSection = {
   shareAudio: false,
 };
 
+/**
+ * 当前模型的朗读风格。模型页保存的值优先,缺省字段回落到旧的全局值,
+ * 这样升级前已有的配置仍然有效。
+ */
+export function modelStyle(cfg: VoiceConfigSection, model = cfg.model): VoiceModelStyle {
+  const saved = cfg.modelStyles?.[model] ?? {};
+  return {
+    voice: saved.voice ?? cfg.voice,
+    emotionMode: saved.emotionMode ?? cfg.emotionMode,
+    defaultEmotion: saved.defaultEmotion ?? cfg.defaultEmotion,
+    defaultInstruct: saved.defaultInstruct ?? cfg.defaultInstruct,
+    speed: saved.speed ?? cfg.speed,
+    paralinguistic: saved.paralinguistic ?? cfg.paralinguistic,
+  };
+}
+
+/** 把模型风格盖到配置上,给过滤、合成与提示词读取。 */
+export function styledConfig(cfg: VoiceConfigSection, model = cfg.model): VoiceConfigSection {
+  return { ...cfg, ...modelStyle(cfg, model) };
+}
+
 const TEXT = {
   zh: {
     label: 'Kmch AI · 语音',
@@ -60,6 +82,7 @@ const TEXT = {
     modelSaveShape: '保存失败:参数格式不对。',
     modelSaveModel: '保存失败:模型 id 不能为空。',
     modelSaveVoice: '保存失败:音色必须是字符串。',
+    modelSaveStyle: '保存失败:朗读风格格式不对。',
     modelSaveOptions: '保存失败:模型参数必须是对象。',
     modelUnknownMethod: '未知的面板方法。',
     connection: '连接',
@@ -70,15 +93,6 @@ const TEXT = {
     serviceTokenSecretDesc: '部署 .env 里的变量名,留空表示服务不校验。',
     format: '音频格式',
     timeoutMs: '超时(毫秒)',
-    style: '音色与情绪',
-    styleDesc: '',
-    emotionMode: '情绪控制',
-    emotionModeDesc: 'instruct 折成自然语言指令;categorical 只发情绪标签;basic 只发语速;tags 只用正文标签;none 不发。',
-    defaultEmotion: '默认情绪',
-    defaultInstruct: '默认表演指令',
-    speed: '语速',
-    paralinguistic: '副语言标签',
-    paralinguisticDesc: 'auto 只在当前模型支持标签时保留;keep 一律保留;strip 一律从朗读文本里删掉。',
     triggers: '触发',
     triggersDesc: '',
     autoSpeak: '自动朗读',
@@ -123,6 +137,7 @@ const TEXT = {
     modelSaveShape: 'Save failed: invalid arguments.',
     modelSaveModel: 'Save failed: the model id must not be empty.',
     modelSaveVoice: 'Save failed: voice must be a string.',
+    modelSaveStyle: 'Save failed: invalid voice style.',
     modelSaveOptions: 'Save failed: model options must be an object.',
     modelUnknownMethod: 'Unknown panel method.',
     connection: 'Connection',
@@ -133,15 +148,6 @@ const TEXT = {
     serviceTokenSecretDesc: 'Name in the deployment .env; empty means the service does not check.',
     format: 'Audio format',
     timeoutMs: 'Timeout (ms)',
-    style: 'Voice and emotion',
-    styleDesc: '',
-    emotionMode: 'Emotion control',
-    emotionModeDesc: 'instruct turns it into natural language; categorical sends a label; basic sends speed; tags uses inline tags only; none sends nothing.',
-    defaultEmotion: 'Default emotion',
-    defaultInstruct: 'Default direction',
-    speed: 'Speed',
-    paralinguistic: 'Paralinguistic tags',
-    paralinguisticDesc: 'auto keeps them only when the current model supports them; keep always keeps; strip removes them from the spoken text.',
     triggers: 'Triggers',
     triggersDesc: '',
     autoSpeak: 'Automatic speech',
@@ -204,21 +210,6 @@ export function voiceConfigGroups(language: Language): ConfigGroup[] {
         },
       },
     },
-    {
-      id: 'world:voice-style',
-      owner: 'world:voice',
-      schema: {
-        type: 'object',
-        title: t.style,
-        properties: {
-          [`${p}emotionMode`]: { type: 'string', title: t.emotionMode, description: t.emotionModeDesc, enum: ['instruct', 'categorical', 'basic', 'tags', 'none'], 'x-hot': true },
-          [`${p}defaultEmotion`]: { type: 'string', title: t.defaultEmotion, 'x-hot': true },
-          [`${p}defaultInstruct`]: { type: 'string', title: t.defaultInstruct, 'x-hot': true },
-          [`${p}speed`]: { type: 'number', title: t.speed, minimum: 0.5, maximum: 2, multipleOf: 0.05, 'x-hot': true },
-          [`${p}paralinguistic`]: { type: 'string', title: t.paralinguistic, description: t.paralinguisticDesc, enum: ['auto', 'keep', 'strip'], 'x-hot': true },
-        },
-      },
-    },
     {
       id: 'world:voice-triggers',
       owner: 'world:voice',

package/src/definition.ts

@@ -7,7 +7,7 @@ import { VoiceWorld } from './world.ts';
 export const VOICE: WorldDefinition<VoiceConfigSection> = {
   id: VOICE_ID,
   label: VOICE_LABEL,
-  defaults: () => ({ ...VOICE_DEFAULTS, modelOptions: {} }),
+  defaults: () => ({ ...VOICE_DEFAULTS, modelOptions: {}, modelStyles: {} }),
   create: (ctx) => new VoiceWorld({
     cfg: ctx.cfg,
     timezone: ctx.timezone,

package/src/models.ts

@@ -28,7 +28,7 @@ const INDEXTTS2: ModelInfo = {
   match: ['index-tts2', 'index_tts2', 'index-tts2.5', 'index_tts2_5', '2.5'],
   label: 'IndexTTS2',
   tags: false,
-  prompt: '情绪用自然语言(emotion_text),不要写 [Laughter] 这类方括号标签;采样/时长参数在语音模型页配置,不要写进正文.',
+  prompt: '情绪用 voice_speak 的 emotion 或 instruct 传自然语言,服务会映射到 emotion_text;不要在正文写 emotion_text 或 [Laughter] 这类方括号标签。采样/时长参数在语音模型页配置,不要写进正文.',
 };
 
 const GENERIC: ModelInfo = {

package/src/panel.ts

@@ -2,19 +2,32 @@
  * 「模型」面板的数据形状。服务端经由 `World.webUi().invoke` 构造,浏览器端只
  * `import type` 这些声明;面板不接触服务地址与令牌。
  */
-import type { VoiceModel } from './types.ts';
+import type { VoiceModel, VoiceModelStyle } from './types.ts';
+
+/** 一个模型在面板里的已保存状态。 */
+export interface PanelModelSetting {
+  voice: string;
+  style: VoiceModelStyle;
+  options: Record<string, unknown>;
+}
 
 export interface PanelModelState {
   /** `worlds.voice.model`。 */
   model: string;
-  /** `worlds.voice.voice`。 */
+  /** 当前模型的音色。 */
   voice: string;
   /** 当前模型的环境提示词片段,来自 `models.ts`。 */
   prompt: string;
+  /** 每个模型的环境提示词片段,切换模型时立即更新。 */
+  prompts: Record<string, string>;
   /** 当前模型已保存的请求参数。 */
   options: Record<string, unknown>;
+  /** 当前模型生效的朗读风格;没有专属设置时回落到全局值。 */
+  style: VoiceModelStyle;
   /** 服务目录;服务不可达时是上一次可用的结果。 */
   models: VoiceModel[];
+  /** 每个模型的音色、风格与请求参数;面板切模型时不用等保存。 */
+  settings: Record<string, PanelModelSetting>;
   /** 最近一次读取服务目录失败的原因。 */
   error?: string;
 }
@@ -23,6 +36,7 @@ export interface PanelModelState {
 export interface PanelModelSave {
   model: string;
   voice: string;
+  style?: Partial<VoiceModelStyle>;
   options: Record<string, unknown>;
 }
 

package/src/triggers.ts

@@ -8,6 +8,8 @@ import type { VoiceConfigSection } from './types.ts';
 
 /** 工具参数里可能放正文的字段,按顺序取第一个字符串。 */
 const TEXT_FIELDS = ['script', 'question', 'text', 'content', 'message', 'body'] as const;
+/** 只结束这一轮的控制工具,不算「这一轮做了别的事」。 */
+const FLOW_TOOLS = new Set(['end_turn']);
 
 export function splitList(raw: string): string[] {
   return raw
@@ -60,15 +62,27 @@ interface StreamLike {
 
 export class VoiceTap implements OutputTap {
   private text = '';
-  private calls = 0;
-  private readonly names = new Map<number, string>();
+  /** 这一轮出现过的函数调用:output_index -> name;名字未知时存空串。 */
+  private readonly callNames = new Map<number, string>();
+  /** 没有 output_index 的函数调用;无法判断是不是 end_turn,按实质调用处理。 */
+  private orphanCall = false;
   private readonly emitted = new Set<number>();
 
   constructor(private readonly opts: TapOptions) {}
 
+  private noteCall(index: number, name: string): void {
+    if (index < 0) {
+      if (name && FLOW_TOOLS.has(name)) return;
+      this.orphanCall = true;
+      return;
+    }
+    if (name) this.callNames.set(index, name);
+    else if (!this.callNames.has(index)) this.callNames.set(index, '');
+  }
+
   /** 工具调用完成:同一个 output_index 只交出去一次。 */
   private takeCall(index: number, name: string, args: string): void {
-    if (!name || name.startsWith('voice_')) return;
+    if (!name || name.startsWith('voice_') || FLOW_TOOLS.has(name)) return;
     if (this.emitted.has(index)) return;
     if (!toolNames(this.opts.cfg()).has(name)) return;
     const text = textOfToolArgs(args);
@@ -85,38 +99,39 @@ export class VoiceTap implements OutputTap {
       return;
     }
     if (e.type === 'response.output_item.added' && e.item?.type === 'function_call') {
-      this.calls += 1;
-      if (e.item.name) this.names.set(index, e.item.name);
+      this.noteCall(index, e.item.name ?? '');
       return;
     }
     // 有的传输只给参数结束事件,不带完整 item;名字从 added 里取。
     if (e.type === 'response.function_call_arguments.done') {
-      const name = this.names.get(index) ?? '';
+      const name = this.callNames.get(index) ?? '';
+      this.noteCall(index, name);
       if (name) this.takeCall(index, name, e.arguments ?? '');
       return;
     }
     if (e.type === 'response.output_item.done' && e.item?.type === 'function_call') {
-      this.calls += 1;
-      const name = e.item.name ?? this.names.get(index) ?? '';
+      const name = e.item.name ?? this.callNames.get(index) ?? '';
+      this.noteCall(index, name);
       if (name) this.takeCall(index, name, e.item.arguments ?? '');
-      this.names.delete(index);
     }
   }
 
   onRoundEnd(): void {
     const text = this.text;
     this.text = '';
-    const pureText = this.calls === 0;
-    this.calls = 0;
-    this.names.clear();
+    // 只有结束轮次的调用不算实质动作;纯文本回复照样可以自动朗读。
+    const nonFlow = this.orphanCall
+      || [...this.callNames.values()].some((name) => name === '' || !FLOW_TOOLS.has(name));
+    this.callNames.clear();
+    this.orphanCall = false;
     this.emitted.clear();
-    if (pureText && text.trim()) this.opts.onText(text, 'text');
+    if (!nonFlow && text.trim()) this.opts.onText(text, 'text');
   }
 
   onAbort(): void {
     this.text = '';
-    this.calls = 0;
-    this.names.clear();
+    this.callNames.clear();
+    this.orphanCall = false;
     this.emitted.clear();
     this.opts.onAbort?.();
   }

package/src/types.ts

@@ -7,6 +7,17 @@ export type EmotionMode = 'instruct' | 'categorical' | 'basic' | 'tags' | 'none'
 export type ParalinguisticMode = 'auto' | 'keep' | 'strip';
 export type PlaybackMode = 'local' | 'event' | 'none';
 
+/** 单个模型的朗读风格;缺省字段回落到 `worlds.voice` 上的全局同名值。 */
+export interface VoiceModelStyle {
+  voice: string;
+  emotionMode: EmotionMode;
+  defaultEmotion: string;
+  defaultInstruct: string;
+  /** 语速倍率;1 为原速。 */
+  speed: number;
+  paralinguistic: ParalinguisticMode;
+}
+
 /** `config.json` 的 `worlds.voice` 段。 */
 export interface VoiceConfigSection {
   enabled: boolean;
@@ -17,13 +28,16 @@ export interface VoiceConfigSection {
   /** 服务令牌的环境变量名;空表示服务不校验。 */
   serviceTokenSecret: string;
   model: string;
+  /** 旧的全局音色;只作为没有模型专属设置时的回落值。 */
   voice: string;
   /** 模型专属的 request 参数,按模型 id 分组;由「模型」面板写回,不进 ConfigGroup。 */
   modelOptions: Record<string, Record<string, unknown>>;
+  /** 模型专属的朗读风格;由「模型」面板写回,不进 ConfigGroup。 */
+  modelStyles: Record<string, Partial<VoiceModelStyle>>;
   format: AudioFormat;
   timeoutMs: number;
 
-  // 音色与情绪
+  // 朗读风格的全局回落值(旧配置仍可用;界面上已移到模型页)
   emotionMode: EmotionMode;
   defaultEmotion: string;
   defaultInstruct: string;

package/src/world.ts

@@ -14,12 +14,12 @@ import type {
   Logger, ToolDef, ToolOutcome, World, WorldHost, WorldLamp, WorldWebUiDecl,
 } from 'cortico/core/types.ts';
 import { nullLogger, nowIso } from 'cortico/core/util.ts';
-import { VOICE_DEFAULTS, VOICE_ID, voiceConfigGroups, voiceText } from './config.ts';
+import { modelStyle, styledConfig, VOICE_DEFAULTS, VOICE_ID, voiceConfigGroups, voiceText } from './config.ts';
 import type { AudioPlayer } from './audio/player.ts';
 import { createPlayer } from './audio/player.ts';
 import { SpeechQueue, type SpeechJob } from './audio/queue.ts';
 import { modelPrompt } from './models.ts';
-import { isPlainObject, type PanelModelState } from './panel.ts';
+import { isPlainObject, type PanelModelSetting, type PanelModelState } from './panel.ts';
 import { ServiceClient, type SpeechService } from './service.ts';
 import { splitSpeech, truncateSpeech } from './speech/chunk.ts';
 import { buildInstruct, resolveStyle } from './speech/style.ts';
@@ -27,7 +27,8 @@ import { filterSpeech } from './speech/text.ts';
 import { EventPoller, eventTypes, hasKeyword, latestUserText, splitList, VoiceTap } from './triggers.ts';
 import { VOICE_TOOL_DECLS } from './tools.ts';
 import type {
-  ServiceHealth, TtsAudio, TtsRequest, VoiceConfigSection, VoiceModel,
+  EmotionMode, ParalinguisticMode, ServiceHealth, TtsAudio, TtsRequest, VoiceConfigSection,
+  VoiceModel, VoiceModelStyle,
 } from './types.ts';
 
 const ENV_PROMPT_FILE = fileURLToPath(new URL('./ENV_PROMPT.md', import.meta.url));
@@ -142,7 +143,7 @@ export class VoiceWorld implements World {
     this.service = this.deps.service ?? new ServiceClient({
       fetch: this.deps.fetch,
       secret: (name) => this.opts.secret(name),
-      cfg: () => this.cfg,
+      cfg: () => this.styled,
     });
     this.player = this.deps.createPlayer(this.log);
     this.queue = new SpeechQueue({
@@ -166,14 +167,19 @@ export class VoiceWorld implements World {
     return this.opts.cfg;
   }
 
+  /** 当前模型的配置视图:模型专属的朗读风格盖在全局值上。 */
+  private get styled(): VoiceConfigSection {
+    return styledConfig(this.cfg, this.cfg.model);
+  }
+
   envPromptVars(): Record<string, string> {
-    const cfg = this.cfg;
+    const cfg = this.styled;
     return {
-      'voice.auto': cfg.autoSpeak ? `${Math.round(cfg.autoProbability * 100)}%` : 'off',
+      'voice.auto': this.cfg.autoSpeak ? `${Math.round(this.cfg.autoProbability * 100)}%` : 'off',
       'voice.voice': cfg.voice || 'default',
       'voice.emotion': cfg.defaultEmotion || 'neutral',
       'voice.instruct': cfg.defaultInstruct || 'none',
-      'voice.modelPrompt': modelPrompt(cfg.model),
+      'voice.modelPrompt': modelPrompt(this.cfg.model),
     };
   }
 
@@ -196,7 +202,7 @@ export class VoiceWorld implements World {
   webUi(language: Language = 'zh'): WorldWebUiDecl {
     const t = voiceText(language);
     this.maybeRefresh();
-    const cfg = this.cfg;
+    const cfg = this.styled;
     const stats = this.queue.stats();
     const serviceLamp: WorldLamp = this.serviceState === null
       ? { label: t.lampService, state: this.started ? 'loading' : 'offline' }
@@ -270,17 +276,41 @@ export class VoiceWorld implements World {
 
   private panelSnapshot(): PanelModelState {
     const cfg = this.cfg;
-    const saved = cfg.modelOptions?.[cfg.model];
+    const settings = this.panelSettings();
+    const current = settings[cfg.model];
+    const style = current?.style ?? modelStyle(cfg, cfg.model);
     return {
       model: cfg.model,
-      voice: cfg.voice,
+      voice: style.voice,
       prompt: modelPrompt(cfg.model),
-      options: isPlainObject(saved) ? { ...saved } : {},
+      prompts: Object.fromEntries(Object.keys(settings).map((id) => [id, modelPrompt(id)])),
+      options: current?.options ?? {},
+      style,
       models: this.catalog ?? [],
+      settings,
       ...(this.catalogError ? { error: this.catalogError } : {}),
     };
   }
 
+  /** 目录里每个模型的音色、风格与参数;当前模型与已保存过设置的模型也各带一条。 */
+  private panelSettings(): Record<string, PanelModelSetting> {
+    const ids = new Set((this.catalog ?? []).map((model) => model.id));
+    if (this.cfg.model.trim()) ids.add(this.cfg.model.trim());
+    for (const id of Object.keys(this.cfg.modelStyles ?? {})) if (id.trim()) ids.add(id);
+    for (const id of Object.keys(this.cfg.modelOptions ?? {})) if (id.trim()) ids.add(id);
+    const out: Record<string, PanelModelSetting> = {};
+    for (const id of ids) {
+      const style = modelStyle(this.cfg, id);
+      const saved = this.cfg.modelOptions?.[id];
+      out[id] = {
+        voice: style.voice,
+        style,
+        options: isPlainObject(saved) ? { ...saved } : {},
+      };
+    }
+    return out;
+  }
+
   private loadCatalog(selected: string): Promise<void> {
     if (this.catalogLoad) return this.catalogLoad;
     this.catalogLoad = this.fetchCatalog(selected).finally(() => {
@@ -338,18 +368,60 @@ export class VoiceWorld implements World {
     const model = typeof raw.model === 'string' ? raw.model.trim() : '';
     if (!model) throw new Error(t.modelSaveModel);
     if (typeof raw.voice !== 'string') throw new Error(t.modelSaveVoice);
+    if (raw.style !== undefined && !isPlainObject(raw.style)) throw new Error(t.modelSaveStyle);
     if (!isPlainObject(raw.options)) throw new Error(t.modelSaveOptions);
 
-    const next: VoiceConfigSection['modelOptions'] = { ...(this.cfg.modelOptions ?? {}) };
+    const style = this.readModelStyle(raw.style, model, raw.voice.trim(), t.modelSaveStyle);
+    const nextStyles: VoiceConfigSection['modelStyles'] = { ...(this.cfg.modelStyles ?? {}) };
+    nextStyles[model] = style;
+
+    const nextOptions: VoiceConfigSection['modelOptions'] = { ...(this.cfg.modelOptions ?? {}) };
     // 空对象表示删掉这个模型的条目,回到服务的默认参数。
-    if (Object.keys(raw.options).length === 0) delete next[model];
-    else next[model] = { ...raw.options };
+    if (Object.keys(raw.options).length === 0) delete nextOptions[model];
+    else nextOptions[model] = { ...raw.options };
     // 配置层对对象做深合并,删掉的键不会随 patch 传播;先让活引用指向新对象再写盘。
-    this.cfg.modelOptions = next;
-    this.opts.persist({ model, voice: raw.voice.trim(), modelOptions: next });
+    this.cfg.modelStyles = nextStyles;
+    this.cfg.modelOptions = nextOptions;
+    this.opts.persist({ model, modelStyles: nextStyles, modelOptions: nextOptions });
     return this.panelSnapshot();
   }
 
+  /** 面板传来的风格字段逐个校验;没给的字段沿用这个模型当前生效的值。 */
+  private readModelStyle(
+    raw: Record<string, unknown> | undefined,
+    model: string,
+    voice: string,
+    shapeError: string,
+  ): VoiceModelStyle {
+    const current = modelStyle(this.cfg, model);
+    if (!raw) return { ...current, voice };
+
+    const emotionModes: readonly EmotionMode[] = ['instruct', 'categorical', 'basic', 'tags', 'none'];
+    const paralinguisticModes: readonly ParalinguisticMode[] = ['auto', 'keep', 'strip'];
+    const emotionMode = raw.emotionMode ?? current.emotionMode;
+    if (typeof emotionMode !== 'string' || !emotionModes.includes(emotionMode as EmotionMode)) {
+      throw new Error(shapeError);
+    }
+    const paralinguistic = raw.paralinguistic ?? current.paralinguistic;
+    if (typeof paralinguistic !== 'string' || !paralinguisticModes.includes(paralinguistic as ParalinguisticMode)) {
+      throw new Error(shapeError);
+    }
+    const defaultEmotion = raw.defaultEmotion ?? current.defaultEmotion;
+    if (typeof defaultEmotion !== 'string') throw new Error(shapeError);
+    const defaultInstruct = raw.defaultInstruct ?? current.defaultInstruct;
+    if (typeof defaultInstruct !== 'string') throw new Error(shapeError);
+    const rawSpeed = raw.speed ?? current.speed;
+    if (typeof rawSpeed !== 'number' || !Number.isFinite(rawSpeed)) throw new Error(shapeError);
+    return {
+      voice,
+      emotionMode: emotionMode as EmotionMode,
+      defaultEmotion: defaultEmotion.trim().slice(0, 40),
+      defaultInstruct: defaultInstruct.trim().slice(0, 500),
+      speed: Math.min(2, Math.max(0.5, rawSpeed)),
+      paralinguistic: paralinguistic as ParalinguisticMode,
+    };
+  }
+
   outputTap() {
     const cfg = this.cfg;
     if (!cfg.autoSpeak) return undefined;
@@ -408,7 +480,7 @@ export class VoiceWorld implements World {
     const cfg = this.cfg;
     if (!cfg.autoSpeak) return;
 
-    const filtered = filterSpeech(raw, cfg);
+    const filtered = filterSpeech(raw, this.styled);
     if (!filtered.text) return;
 
     const userText = cfg.explicitTrigger
@@ -449,7 +521,7 @@ export class VoiceWorld implements World {
   // ———————— 工具 ————————
 
   private async handleSpeak(args: Record<string, unknown>): Promise<ToolOutcome> {
-    const filtered = filterSpeech(asString(args.text), this.cfg);
+    const filtered = filterSpeech(asString(args.text), this.styled);
     if (!filtered.text) return { text: 'voice_speak failed: there is nothing to say.', failed: true };
     const silent = args.silent === true || this.cfg.playback === 'none';
     const share = this.cfg.shareAudio || silent;
@@ -500,26 +572,47 @@ export class VoiceWorld implements World {
 
   private async handleStyle(args: Record<string, unknown>): Promise<ToolOutcome> {
     const has = (key: string): boolean => Object.prototype.hasOwnProperty.call(args, key);
-    const patch: Partial<VoiceConfigSection> = {};
-    if (has('emotion')) patch.defaultEmotion = asString(args.emotion).slice(0, 40);
-    if (has('instruct')) patch.defaultInstruct = asString(args.instruct).slice(0, 500);
-    if (has('voice')) patch.voice = asString(args.voice).slice(0, 200);
+    const model = this.cfg.model.trim();
+    if (!model) {
+      // 还没选模型时写到全局回落值;模型页仍可之后覆盖。
+      const patch: Partial<VoiceConfigSection> = {};
+      if (has('emotion')) patch.defaultEmotion = asString(args.emotion).slice(0, 40);
+      if (has('instruct')) patch.defaultInstruct = asString(args.instruct).slice(0, 500);
+      if (has('voice')) patch.voice = asString(args.voice).slice(0, 200);
+      if (has('speed')) {
+        const speed = asNumber(args.speed);
+        if (speed !== undefined) patch.speed = Math.min(2, Math.max(0.5, speed));
+      }
+      if (Object.keys(patch).length > 0) {
+        this.opts.persist(patch);
+        return { text: `Voice style updated. ${this.styleSummary()}` };
+      }
+      return { text: this.styleSummary() };
+    }
+
+    const current = modelStyle(this.cfg, model);
+    const next: VoiceModelStyle = { ...current };
+    if (has('emotion')) next.defaultEmotion = asString(args.emotion).slice(0, 40);
+    if (has('instruct')) next.defaultInstruct = asString(args.instruct).slice(0, 500);
+    if (has('voice')) next.voice = asString(args.voice).slice(0, 200);
     if (has('speed')) {
       const speed = asNumber(args.speed);
-      if (speed !== undefined) patch.speed = Math.min(2, Math.max(0.5, speed));
+      if (speed !== undefined) next.speed = Math.min(2, Math.max(0.5, speed));
     }
-    if (Object.keys(patch).length > 0) {
-      this.opts.persist(patch);
+    if (Object.keys(next).some((key) => next[key as keyof VoiceModelStyle] !== current[key as keyof VoiceModelStyle])) {
+      const modelStyles: VoiceConfigSection['modelStyles'] = { ...(this.cfg.modelStyles ?? {}), [model]: next };
+      this.cfg.modelStyles = modelStyles;
+      this.opts.persist({ modelStyles });
       return { text: `Voice style updated. ${this.styleSummary()}` };
     }
     return { text: this.styleSummary() };
   }
 
   private styleSummary(): string {
-    const cfg = this.cfg;
+    const cfg = this.styled;
     return `Voice: ${cfg.voice || 'default'}; emotion: ${cfg.defaultEmotion || 'neutral'}; `
       + `direction: ${cfg.defaultInstruct || 'none'}; speed: ${cfg.speed}; `
-      + `automatic: ${cfg.autoSpeak ? `${Math.round(cfg.autoProbability * 100)}%` : 'off'}.`;
+      + `automatic: ${this.cfg.autoSpeak ? `${Math.round(this.cfg.autoProbability * 100)}%` : 'off'}.`;
   }
 
   private spokenOutcome(audio: TtsAudio | undefined, seconds: number, silent: boolean): ToolOutcome {
@@ -553,7 +646,7 @@ export class VoiceWorld implements World {
     style: { emotion?: string; instruct?: string; voice?: string; speed?: number },
     signal: AbortSignal,
   ): Promise<TtsAudio> {
-    const cfg = this.cfg;
+    const cfg = this.styled;
     const resolved = resolveStyle(cfg, style);
     const request: TtsRequest = {
       text,
@@ -599,7 +692,7 @@ export class VoiceWorld implements World {
         origin: 'internal',
         senderKey: this.id,
         text: `Spoken line: ${chunk.slice(0, 100)}`,
-        meta: { source: job.source, voice: this.cfg.voice, emotion: this.cfg.defaultEmotion },
+        meta: { source: job.source, voice: this.styled.voice, emotion: this.styled.defaultEmotion },
         blobs: [{
           bytes: audio.bytes,
           mime: audio.mime,