dsh-her 0.1.0 → 0.2.0

package/lib/core/character.js

@@ -40,7 +40,7 @@ export function invalidCharacter(c) {
         return '人设太长了（最多 3000 字）';
     if ([...c.callYou].length > 20)
         return '称呼最多 20 个字';
-    if (!/^[A-Za-z0-9_-]{1,100}$/.test(c.voice.id))
+    if (!/^(?:[A-Za-z0-9_-]{1,100}|ming:[\u4e00-\u9fff_]{1,20}|ming-own:v[0-9a-z]{1,20})$/.test(c.voice.id))
         return '音色不对';
     if (!(c.voice.speed >= 0.5 && c.voice.speed <= 2))
         return '语速要在 0.5 到 2 之间';

package/lib/core/index.js

@@ -5,6 +5,8 @@ export { judgeBargeIn } from "./echo.js";
 export { EMOTION_WORDS, extractCues, isHeavy, performanceSection } from "./cues.js";
 export { buildSystemPrompt, memoryContext, SAMPLE_PERSONA } from "./prompt.js";
 export { herPaths } from "./paths.js";
-export { DEFAULT_VOICE, SERVICE_NAMES, serviceOf, VOICE_MODEL, VOICES, voiceName } from "./voices.js";
+export { DEFAULT_VOICE, MING_MOODS, mingSpeech, OWN_PREFIX, SERVICE_NAMES, serviceOf, VOICE_MODEL, VOICES, voiceName } from "./voices.js";
+export { extractSounds, SOUND_SECTION, withSounds } from "./sounds.js";
+export { spokenNumbers } from "./spoken-numbers.js";
 export { invalidCharacter, loadCharacter, saveCharacter } from "./character.js";
 export { decodeWav, encodeWav, wavFromPcm16 } from "./wav.js";

package/lib/core/sounds.js

@@ -0,0 +1,31 @@
+/**
+ * Sounds she can make with the Ming voice (docs/Ming本地声音设计.md §5): she
+ * writes 〔叹气〕 or 〔笑〕 in a sentence; the tag is taken out of the subtitle
+ * and Ming's own marker ([Sigh], [Laughter]) goes at the start or the end of
+ * what is spoken. A marker in the middle of a sentence swallows the word next
+ * to it, so it is moved to whichever end the tag was nearer.
+ *
+ * With another voice the tags are simply dropped (extractCues removes every
+ * tag it does not know), and her prompt never mentions them.
+ */
+const SOUNDS = { 叹气: '[Sigh]', 笑: '[Laughter]' };
+const SOUND_TAG = /〔\s*(叹气|笑)\s*〕/g;
+export function extractSounds(text) {
+    const before = [];
+    const after = [];
+    const length = [...text.replace(SOUND_TAG, '')].length;
+    let at = 0;
+    const clean = text.replace(SOUND_TAG, (whole, name, offset) => {
+        const position = [...text.slice(0, offset)].length - at;
+        at += [...whole].length;
+        (position * 2 <= length ? before : after).push(SOUNDS[name]);
+        return '';
+    });
+    return { text: clean, before, after };
+}
+/** What Ming says: the sentence with its sound markers at the ends. */
+export function withSounds(spoken, sounds) {
+    return [...sounds.before, spoken, ...sounds.after].filter(part => part !== '').join(' ');
+}
+/** The lines of her prompt that teach the sounds (only while she speaks with Ming). */
+export const SOUND_SECTION = '你的声音能真的叹气和笑出声：在句子开头或结尾写〔叹气〕或〔笑〕（不会显示）。只在真的想叹气、笑出来时用，一次回复最多一次。';

package/lib/core/spoken-numbers.js

@@ -0,0 +1,90 @@
+/**
+ * Digits and a few symbols written out the way they are said, for the Ming
+ * voice, which reads "2个" as "二个" (its MLX port has no text
+ * normalization). Only what turns up in her chat: counts, times, dates,
+ * years, percentages, decimals, phone numbers and codes, ranges, degrees.
+ */
+const DIGITS = '零一二三四五六七八九';
+const SMALL_UNITS = ['', '十', '百', '千'];
+const BIG_UNITS = ['', '万', '亿'];
+/** Measure words after which 2 is 两 (两个, 两点, 两次 …). */
+const MEASURE = /^(?:个|位|点|次|天|只|本|件|张|条|岁|种|遍|杯|分钟|小时|周|斤|块|台|辆|家|句|下|口|双|对|份|层|片|场|首|部|回|趟|步|顿|碗|瓶|盒|袋|根|颗|粒|间|座|所|名|人|样|批|年|星期|礼拜|倍)/;
+/** Words before a number that is read digit by digit. */
+const DIGIT_CONTEXT = /(?:验证码|号码|电话|手机|编号|尾号|房间|门牌|邮编|工号|学号|密码|单号)[是为:：\s]*$/;
+/** "2026" → 二零二六. */
+export function readDigits(digits) {
+    return [...digits].map(d => DIGITS[Number(d)]).join('');
+}
+/** "1205" → 一千二百零五; "12" → 十二; "20000" → 两万. Longer than 12 digits: one by one. */
+export function readInteger(number) {
+    const digits = number.replace(/^0+(?=\d)/, '');
+    if (digits.length > 12)
+        return readDigits(digits);
+    if (/^0+$/.test(digits))
+        return '零';
+    const groups = [];
+    for (let end = digits.length; end > 0; end -= 4)
+        groups.unshift(Number(digits.slice(Math.max(0, end - 4), end)));
+    let out = '';
+    let zero = false;
+    groups.forEach((group, i) => {
+        const big = BIG_UNITS[groups.length - 1 - i];
+        if (group === 0) {
+            zero = out !== '';
+            return;
+        }
+        if (out !== '' && (zero || group < 1000))
+            out += '零';
+        zero = false;
+        out += readGroup(group) + big;
+    });
+    return out.replace(/^一十/, '十').replace(/^二(?=[百千万亿])/, '两');
+}
+function readGroup(group) {
+    let out = '';
+    let zero = false;
+    for (let pos = 3; pos >= 0; pos--) {
+        const d = Math.floor(group / 10 ** pos) % 10;
+        if (d === 0) {
+            zero = out !== '';
+            continue;
+        }
+        if (zero)
+            out += '零';
+        zero = false;
+        out += DIGITS[d] + SMALL_UNITS[pos];
+    }
+    return out;
+}
+export function spokenNumbers(text) {
+    return text
+        // 30% → 百分之三十; 2.5% → 百分之二点五
+        .replace(/(\d+(?:\.\d+)?)\s*[%％]/g, (_w, n) => `百分之${readNumber(n)}`)
+        // 15:30 → 十五点三十分; 9:05 → 九点零五分; 8:00 → 八点
+        .replace(/(\d{1,2})[:：](\d{2})(?!\d)/g, (_w, h, m) => `${hour(h)}点${m === '00' ? '' : `${m.startsWith('0') ? '零' : ''}${readInteger(m)}分`}`)
+        // 2026-10-31 → 二零二六年十月三十一日
+        .replace(/(?<!\d)(\d{4})[-/](\d{1,2})[-/](\d{1,2})(?!\d)/g, (_w, y, m, d) => `${readDigits(y)}年${readInteger(m)}月${readInteger(d)}日`)
+        // 2026年 → 二零二六年
+        .replace(/(\d{4})(?=\s*年)/g, (_w, y) => readDigits(y))
+        // 20~25度, 3-5个 → 到 (short numbers only: 2026-10-31 is a date)
+        .replace(/(?<![\d-])(\d{1,3})\s*[~～—–-]\s*(\d{1,3})(?![\d-])/g, '$1到$2')
+        .replace(/\s*(?:℃|°C|°)/g, '度')
+        .replace(/(\d+)\.(\d+)/g, (_w, i, f) => `${readInteger(i)}点${readDigits(f)}`)
+        .replace(/\d+/g, (n, offset, whole) => {
+        const before = whole.slice(0, offset);
+        const after = whole.slice(offset + n.length);
+        if (n.length >= 7 || (n.length > 1 && n.startsWith('0')) || DIGIT_CONTEXT.test(before))
+            return readDigits(n);
+        if (n === '2' && MEASURE.test(after))
+            return '两';
+        return readInteger(n);
+    });
+}
+function readNumber(n) {
+    const [i, f] = n.split('.');
+    return f === undefined ? readInteger(i) : `${readInteger(i)}点${readDigits(f)}`;
+}
+/** 2 o'clock is 两点. */
+function hour(h) {
+    return Number(h) === 2 ? '两' : readInteger(h);
+}

package/lib/core/voices.js

@@ -5,7 +5,10 @@
  *   needs the internet; voice/edge.ts);
  * - cosyvoice: Alibaba Cloud Model Studio (百炼) CosyVoice, model
  *   cosyvoice-v3-flash (needs a key and paid quota; voice/cosyvoice.ts). Its
- *   ids are the API's `voice` parameter, exactly as listed.
+ *   ids are the API's `voice` parameter, exactly as listed;
+ * - ming: Ming-3.0-tts on this computer (Apple chips; voice/ming.ts). Built-in
+ *   voices are `ming:<name>`; voices made on the "她" page (designed from a
+ *   description, or cloned from a recording) are `ming-own:<id>`.
  */
 export const VOICE_MODEL = 'cosyvoice-v3-flash';
 export const VOICES = [
@@ -28,6 +31,13 @@ export const VOICES = [
     { id: 'longyingling_v3', name: '龙应聆', desc: '温和，会共情', service: 'cosyvoice' },
     { id: 'longanwen_v3', name: '龙安温', desc: '优雅，知性', service: 'cosyvoice' },
     { id: 'longxiaochun_v3', name: '龙小淳', desc: '知性，积极', service: 'cosyvoice' },
+    // Ming's built-in voices that suit her: grown-up women, warm or sweet (the catalog's "girlfriend" voices).
+    { id: 'ming:灵风信', name: '灵风信', desc: '体贴，温暖安心，带情绪', service: 'ming', moods: ['中性', '高兴', '悲伤', '生气', '厌恶'] },
+    { id: 'ming:灵雨汐', name: '灵雨汐', desc: '软糯，有气声，带情绪', service: 'ming', moods: ['中性', '高兴', '悲伤', '惊讶', '恐惧'] },
+    { id: 'ming:灵梦', name: '灵梦', desc: '很甜，爱撒娇', service: 'ming' },
+    { id: 'ming:灵恬', name: '灵恬', desc: '甜蜜，温柔', service: 'ming' },
+    { id: 'ming:灵小汐', name: '灵小汐', desc: '清甜，自然', service: 'ming' },
+    { id: 'ming:灵小芸', name: '灵小芸', desc: '温和，像知心朋友', service: 'ming' },
 ];
 export const DEFAULT_VOICE = 'zh-CN-XiaoxiaoNeural';
 export function voiceName(id) {
@@ -35,9 +45,29 @@ export function voiceName(id) {
 }
 /** Which service a voice belongs to; an id not in the list goes by its shape (Edge ids end in "Neural"). */
 export function serviceOf(id) {
+    if (id.startsWith('ming:') || id.startsWith(OWN_PREFIX))
+        return 'ming';
     return VOICES.find(v => v.id === id)?.service ?? (/Neural$/.test(id) ? 'edge' : 'cosyvoice');
 }
+/** A Ming voice made on the "她" page. */
+export const OWN_PREFIX = 'ming-own:';
+/** Her emotions (core/cues.ts) as Ming moods; the others are said plainly. */
+export const MING_MOODS = { happy: '高兴', sad: '悲伤', angry: '生气', surprised: '惊讶' };
+/** How Ming says a sentence in a voice: a built-in voice (its mood version when it has one) or one of your own, plus mood and speed. */
+export function mingSpeech(id, emotion, speed) {
+    const mood = emotion === undefined ? undefined : MING_MOODS[emotion];
+    const pace = speed < 0.9 ? '慢速' : speed > 1.1 ? '快速' : undefined;
+    const extra = { ...(pace === undefined ? {} : { speed: pace }) };
+    if (id.startsWith(OWN_PREFIX))
+        return { own: id.slice(OWN_PREFIX.length), ...(mood === undefined ? {} : { emotion: mood }), ...extra };
+    const name = id.slice('ming:'.length);
+    const moods = VOICES.find(v => v.id === id)?.moods;
+    if (moods !== undefined)
+        return { voice: `${name}_${mood !== undefined && moods.includes(mood) ? mood : '中性'}`, ...extra };
+    return { voice: name, ...(mood === undefined ? {} : { emotion: mood }), ...extra };
+}
 export const SERVICE_NAMES = {
     edge: '微软 Edge（免费，要联网）',
     cosyvoice: '阿里云百炼 CosyVoice（要 key，按字数付费）',
+    ming: '本机 Ming（离线、免费、能表达情绪；Apple 芯片，下载约 2 GB）',
 };

package/lib/dsh/api.js

@@ -69,6 +69,59 @@ export function apiHandlers(rt) {
             }),
         }),
         '/voice/preview': async (body) => rt.previewVoice(str(body.voice, '音色'), num(body.speed ?? 1, '语速'), str(body.text ?? '你好呀，我在你桌面上等你好久啦。', '试听的话')),
+        '/voice/ming/install': async () => {
+            rt.mingInstall();
+            return withState();
+        },
+        '/voice/ming/uninstall': async () => {
+            rt.mingUninstall();
+            return withState();
+        },
+        '/voice/ming/restart': async () => {
+            await rt.mingRestart();
+            return withState();
+        },
+        '/voice/ming/settings': async (body) => {
+            const patch = {};
+            if (body.quality === 'quality' || body.quality === 'speed')
+                patch.quality = body.quality;
+            if (typeof body.moods === 'boolean')
+                patch.moods = body.moods;
+            if (body.idleMinutes !== undefined) {
+                // -1: back to the default for this Mac's memory.
+                const minutes = num(body.idleMinutes, '停止前的分钟数');
+                if (minutes !== -1 && (minutes < 0 || minutes > 24 * 60))
+                    throw new ApiError('停止前的分钟数不对');
+                Object.assign(patch, { idleMinutes: minutes === -1 ? undefined : minutes });
+            }
+            rt.mingSettings(patch);
+            return withState();
+        },
+        '/voice/ming/design': async (body) => {
+            const description = str(body.description, '声音描述').trim();
+            if (description.length < 4 || description.length > 300)
+                throw new ApiError('声音描述写 4 到 300 个字');
+            return { drafts: await rt.mingDesign(description) };
+        },
+        '/voice/ming/clone': async (body) => {
+            const audio = str(body.audio, '录音');
+            // About 20 MB of recording: far more than the 15 seconds that are used.
+            if (audio.length > 28_000_000)
+                throw new ApiError('录音文件太大了，剪一段 15 秒以内的');
+            return { draft: await rt.mingClone(audio, str(body.name ?? 'recording.wav', '文件名')) };
+        },
+        '/voice/ming/keep': async (body) => {
+            const change = {};
+            if (body.text !== undefined)
+                change.text = str(body.text, '录音里说的话');
+            if (body.consent !== undefined)
+                change.consent = body.consent === true;
+            return withState({ voice: rt.mingKeep(str(body.draft, '试听'), str(body.name, '名字'), change) });
+        },
+        '/voice/ming/remove': async (body) => {
+            rt.mingRemove(str(body.id, '声音'));
+            return withState();
+        },
         '/body/start': async () => withState({ message: rt.startBody() }),
         '/body/mode': async (body) => {
             const mode = str(body.mode, '模式');

package/lib/dsh/plugin.js

@@ -32,7 +32,7 @@
 import { spawn } from 'node:child_process';
 import { appendFileSync, mkdirSync, readFileSync, writeFileSync } from 'node:fs';
 import { join } from 'node:path';
-import { buildSystemPrompt, extractCues, herPaths, invalidCharacter, judgeBargeIn, loadCharacter, memoryContext, performanceSection, readingMs, saveCharacter, serviceOf, speakable } from "../core/index.js";
+import { buildSystemPrompt, DEFAULT_VOICE, extractCues, extractSounds, herPaths, invalidCharacter, judgeBargeIn, loadCharacter, memoryContext, OWN_PREFIX, performanceSection, readingMs, saveCharacter, serviceOf, SOUND_SECTION, speakable, withSounds } from "../core/index.js";
 import { digestPrompt, notePrompt, parseDigest, transcriptOf } from "../memory/digest.js";
 import { dayOf, FileMemory } from "../memory/store.js";
 import { PROTOCOL_VERSION } from "../protocol/index.js";
@@ -46,7 +46,30 @@ import { exampleFile, examples, exportProject, findProject, importProject, listP
 import { HerSession } from "./her-session.js";
 import { heardText, SpeechQueue } from "../voice/speech-queue.js";
 import { Voice } from "../voice/voice.js";
+import { defaultIdleMinutes, MingVoice } from "../voice/ming.js";
+import { MING_STEPS, OwnVoices, prepareRecording } from "../voice/ming-voices.js";
 const SERVICES = ['agents', 'agentDefaultModel', 'workspaceRegistry', 'sessionTitle', 'sessionQuery', 'llm'];
+/** The error a model stream chunk carries, if any (an `error` chunk, or a finish whose reason is an error). */
+function chunkError(chunk) {
+    const reason = chunk.reason;
+    if (chunk.type === 'error')
+        return String(chunk.error?.message ?? chunk.message ?? '模型出错了');
+    if (reason?.kind === 'error')
+        return reason.error?.message ?? '模型出错了';
+    return undefined;
+}
+/** What to tell you when her model failed: the likely fix, in plain words. */
+export function explainModelError(message, model) {
+    if (/no API key|api[_ ]?key|401|unauthori[sz]ed/i.test(message))
+        return `她用的模型（${model}）没有 API key，没法回答。到「她 → 设置 → 模型」换一个能用的模型，或者在 dsh 的设置里给它填上 key。`;
+    if (/quota|insufficient|balance|余额|429/i.test(message))
+        return `她用的模型（${model}）额度不够或请求太多，暂时没法回答。可以等一会儿，或在「她 → 设置 → 模型」换一个。`;
+    if (/timeout|timed out|ECONN|network|fetch failed/i.test(message))
+        return `连不上她用的模型（${model}），检查一下网络，或者换一个模型。`;
+    return `她用的模型（${model}）出错了：${message.slice(0, 160)}`;
+}
+/** After a failed digest, wait this long before trying again (the model is probably not usable yet). */
+const DIGEST_RETRY_MS = 60 * 60_000;
 /** The digest runs this long after you last talked, or from DIGEST_HOUR:DIGEST_MINUTE. */
 const DIGEST_IDLE_MS = 30 * 60_000;
 const DIGEST_HOUR = 23;
@@ -83,6 +106,10 @@ export function apply(ctx, config) {
     /** Resolves on the first sentence of the next reply. */
     let onFirstSentence;
     let voice;
+    /** The local Ming voice (made with the voice) and her own Ming voices. */
+    let ming;
+    const ownVoices = new OwnVoices(join(paths.root, 'voices', 'ming'));
+    const usesMing = () => serviceOf(character.voice.id) === 'ming';
     let speech;
     let nextStream = 1;
     /** What you heard of her last cut-off reply; told to her before your next turn. */
@@ -101,9 +128,22 @@ export function apply(ctx, config) {
     /** When you last talked to her, and the day of it (for the digest). */
     let lastTalkAt = 0;
     let digesting;
+    /** A digest that failed is not tried again before this (it ran every few minutes against a model with no key). */
+    let digestRetryAt = 0;
+    /** The last reply that came out empty because her model failed; cleared by the next sentence she says. */
+    let problem;
     /** Write her diary page for `day` and keep what is worth keeping; a short report. */
     const digest = (day) => {
-        digesting ??= runDigest(day).finally(() => { digesting = undefined; });
+        digesting ??= runDigest(day)
+            .then(report => {
+            if (report.includes('失败'))
+                digestRetryAt = Date.now() + DIGEST_RETRY_MS;
+            return report;
+        }, (error) => {
+            digestRetryAt = Date.now() + DIGEST_RETRY_MS;
+            throw error;
+        })
+            .finally(() => { digesting = undefined; });
         return digesting;
     };
     /** One model call outside her session (digest, persona drafts): the whole answer as text. */
@@ -119,6 +159,9 @@ export function apply(ctx, config) {
         })) {
             if (chunk.type === 'text-delta' && typeof chunk.text === 'string')
                 answer += chunk.text;
+            const error = chunkError(chunk);
+            if (error !== undefined)
+                throw new Error(error);
         }
         return answer.trim();
     };
@@ -182,6 +225,8 @@ export function apply(ctx, config) {
     const digestTick = () => {
         const now = new Date();
         memory.settle(now);
+        if (now.getTime() < digestRetryAt)
+            return;
         const done = memory.lastDigestDay();
         if (lastTalkAt === 0) {
             // DSH restarted since you last talked: catch up on yesterday, if it was never digested.
@@ -337,13 +382,20 @@ export function apply(ctx, config) {
     };
     const broadcast = (message) => server?.broadcast(message);
     /** A reply with no words: record what the turn did instead (tool calls, reasoning only, errors). */
+    /** Log the end of her session after a reply with no words; the model's error, when the turn ended on one. */
     const logQuietTurn = async () => {
         const id = session?.sessionId;
         const query = services?.sessionQuery;
         if (id === undefined || query === undefined)
-            return;
+            return undefined;
         try {
-            const events = (await query.readSession(id)).events.slice(-12);
+            // The reply's end can come before DSH has written the turn's end (and its error): wait for it a little.
+            let events = (await query.readSession(id)).events.slice(-12);
+            for (let i = 0; i < 10 && !events.some(e => e.type === 'turn/end'); i++) {
+                await new Promise(resolve => setTimeout(resolve, 300));
+                events = (await query.readSession(id)).events.slice(-12);
+            }
+            const failed = events.findLast(e => e.type === 'turn/end' && e.data.reason?.kind === 'error');
             log({
                 event: 'quiet-turn',
                 events: events.map(e => {
@@ -357,9 +409,11 @@ export function apply(ctx, config) {
                     return e.type;
                 }),
             });
+            return failed === undefined ? undefined : failed.data.reason.error?.message ?? '模型出错了';
         }
         catch (error) {
             log({ event: 'quiet-turn', error: error.message });
+            return undefined;
         }
     };
     /** Her model: the "她" page's choice, else the plugin config's, else DSH's default. */
@@ -391,6 +445,8 @@ export function apply(ctx, config) {
         cut();
         sentAt = Date.now();
         lastTalkAt = sentAt;
+        // A voice that has to start up (Ming after a quiet while) starts while she thinks.
+        voice?.warm();
         broadcast({ t: 'state', cid, state: 'think' });
         clearTimeout(stuckTimer);
         stuckTimer = setTimeout(() => {
@@ -537,7 +593,10 @@ export function apply(ctx, config) {
                 platform: process.platform,
                 character,
                 model: target === undefined ? null : { provider: target.provider, model: target.model, ...(target.reasoningEffort === undefined ? {} : { effort: target.reasoningEffort }), chosen: character.model !== undefined },
-                voice: { status: voice?.status ?? 'off', speaking: voice?.speaking ?? { status: 'off', service: serviceOf(character.voice.id) }, id: character.voice.id, speed: character.voice.speed },
+                voice: {
+                    status: voice?.status ?? 'off', speaking: voice?.speaking ?? { status: 'off', service: serviceOf(character.voice.id) }, id: character.voice.id, speed: character.voice.speed,
+                    ming: { ...(ming?.status ?? { state: 'stopped' }), installed: ming?.installed ?? false, settings: ownVoices.settings(), own: ownVoices.list() },
+                },
                 body: {
                     connected: body !== undefined,
                     running: body !== undefined || bodyRunning(paths.root),
@@ -553,6 +612,7 @@ export function apply(ctx, config) {
                 },
                 import: importStatus,
                 ...(lastHeard === undefined ? {} : { heard: lastHeard }),
+                ...(problem === undefined ? {} : { problem }),
                 memory: { facts, episodes, diary, lastDigestDay: memory.lastDigestDay() ?? null },
             };
         },
@@ -695,6 +755,49 @@ export function apply(ctx, config) {
                 throw new Error('没有这个场景');
             broadcast({ t: 'look.set', ...change });
         },
+        mingInstall: () => {
+            if (ming === undefined)
+                throw new Error('声音还没准备好');
+            void ming.install().catch(error => ctx.logger.warn(`dsh-her: ${error.message}`));
+        },
+        mingUninstall: () => {
+            if (usesMing())
+                api.saveCharacter({ voice: { id: DEFAULT_VOICE, speed: character.voice.speed } });
+            ming?.uninstall();
+        },
+        mingRestart: async () => {
+            await ming?.restart();
+        },
+        mingSettings: patch => {
+            const next = ownVoices.saveSettings(patch);
+            ming?.setIdleMinutes(next.idleMinutes ?? defaultIdleMinutes());
+        },
+        mingDesign: async (description) => {
+            if (ming === undefined)
+                throw new Error('声音还没准备好');
+            const text = `你好呀，我是${character.name}。今天过得怎么样？有什么想跟我说的吗？`;
+            const drafts = [];
+            // Three takes of one description: each comes out a little different; you keep the one you like.
+            for (const seed of [11, 22, 33]) {
+                const audio = await ming.synthesize({ text, design: description, steps: MING_STEPS.quality, seed });
+                const id = ownVoices.addDraft(audio.audio, { kind: 'design', text, description });
+                drafts.push({ id, audio: Buffer.from(audio.audio).toString('base64'), durationMs: audio.durationMs });
+            }
+            return drafts;
+        },
+        mingClone: async (audio, name) => {
+            const recording = await prepareRecording(Buffer.from(audio, 'base64'), name);
+            // What it says, to start from (Ming needs the words); you correct it on the page.
+            const heard = await voice?.recognize(recording.pcm16k, 16_000).catch(() => undefined);
+            const id = ownVoices.addDraft(recording.wav, { kind: 'clone', text: heard?.text ?? '' });
+            return { id, audio: Buffer.from(recording.wav).toString('base64'), durationMs: Math.round(recording.seconds * 1000), text: heard?.text ?? '' };
+        },
+        mingKeep: (draft, name, change) => ownVoices.keep(draft, name, change),
+        mingRemove: id => {
+            if (character.voice.id === `${OWN_PREFIX}${id}`)
+                api.saveCharacter({ voice: { id: 'ming:灵风信', speed: character.voice.speed } });
+            ownVoices.remove(id);
+        },
         memoryRemove: ids => memory.remove(ids),
         memoryConfirm: ids => memory.confirm(ids),
         digest: () => digest(dayOf(new Date())),
@@ -706,7 +809,7 @@ export function apply(ctx, config) {
         session = new HerSession({
             cid,
             character: () => character,
-            performance: () => performanceSection(bodyPack),
+            performance: () => [performanceSection(bodyPack), usesMing() ? SOUND_SECTION : ''].filter(Boolean).join('\n'),
             playAction,
             aboutYou,
             memory,
@@ -731,20 +834,30 @@ export function apply(ctx, config) {
                     }
                     onFirstSentence?.();
                     onFirstSentence = undefined;
+                    problem = undefined;
                     log({ event: 'sentence', attemptId, text, sinceSendMs: sentAt === undefined ? null : now - sentAt });
-                    const performed = extractCues(text, bodyPack);
+                    // Ming can sigh and laugh: those tags become its sound markers, out of the subtitle.
+                    const sounds = usesMing() ? extractSounds(text) : undefined;
+                    const performed = extractCues(sounds?.text ?? text, bodyPack);
+                    const spoken = sounds !== undefined && sounds.before.length + sounds.after.length > 0 ? withSounds(performed.text, sounds) : undefined;
                     if (performed.text !== '' || performed.cues.length > 0)
-                        speech?.say(attemptId, seq, performed.text, performed.cues);
+                        speech?.say(attemptId, seq, performed.text, performed.cues, spoken);
                 },
                 end(attemptId, outcome) {
                     log({ event: 'end', attemptId, outcome, sinceSendMs: sentAt === undefined ? null : Date.now() - sentAt });
                     if (current?.attemptId === attemptId) {
                         current.outcome = outcome;
-                        if (current.sentences.length === 0 && outcome !== 'committed') {
-                            broadcast({ t: 'bubble', cid, text: '（我刚才没说出话来，模型可能出错了）', ttlMs: 6000 });
+                        if (current.sentences.length === 0) {
+                            // No words at all: say why, on her and on the page (a model without a key used to fail silently).
+                            const model = route();
+                            void logQuietTurn().then(error => {
+                                if (error === undefined && outcome === 'committed')
+                                    return;
+                                const text = error === undefined ? '我刚才没说出话来，模型可能出错了' : explainModelError(error, model === undefined ? '没有模型' : `${model.provider} / ${model.model}`);
+                                problem = { text, at: Date.now() };
+                                broadcast({ t: 'bubble', cid, text: `（${text}）`, ttlMs: 12_000 });
+                            });
                         }
-                        if (current.sentences.length === 0)
-                            void logQuietTurn();
                         finished.push(current);
                         if (finished.length > 20)
                             finished.shift();
@@ -756,10 +869,11 @@ export function apply(ctx, config) {
         });
         if (services !== undefined)
             open();
-        voice = new Voice(paths.models, join(paths.root, 'edge-tts'), { ...config.voice, voice: character.voice.id, speed: character.voice.speed }, ctx.logger);
+        ming = new MingVoice(join(paths.root, 'ming'), join(paths.root, 'uv'), ctx.logger, ownVoices.settings().idleMinutes ?? defaultIdleMinutes());
+        voice = new Voice(paths.models, join(paths.root, 'edge-tts'), { ...config.voice, voice: character.voice.id, speed: character.voice.speed }, ctx.logger, { voice: ming, own: ownVoices });
         const herVoice = voice;
         void herVoice.start();
-        speech = new SpeechQueue(text => herVoice.synthesize(speakable(text)), {
+        speech = new SpeechQueue((text, cues) => herVoice.synthesize(speakable(text), cues), {
             sentence(uid, seq, text, audio, cues) {
                 const withCues = cues.length === 0 ? {} : { cues };
                 if (audio === undefined) {
@@ -839,6 +953,7 @@ export function apply(ctx, config) {
             clearTimeout(stuckTimer);
             herVoice.stop();
             voice = undefined;
+            ming = undefined;
             speech = undefined;
             const closing = session;
             session = undefined;

package/lib/types/core/character.d.ts

@@ -12,7 +12,7 @@ export interface Character {
     callYou: string;
     /** How she refers to you in her notes and memories. */
     pronoun: Pronoun;
-    /** A CosyVoice voice id (voices.ts), and how fast she talks. */
+    /** A voice id (voices.ts: Edge, CosyVoice or Ming), and how fast she talks. */
     voice: {
         id: string;
         speed: number;

package/lib/types/core/index.d.ts

@@ -5,6 +5,8 @@ export { judgeBargeIn, type EchoVerdict } from './echo.ts';
 export { EMOTION_WORDS, extractCues, isHeavy, performanceSection } from './cues.ts';
 export { buildSystemPrompt, memoryContext, SAMPLE_PERSONA, type MemoryContext, type PromptParts } from './prompt.ts';
 export { herPaths, type HerPaths } from './paths.ts';
-export { DEFAULT_VOICE, SERVICE_NAMES, serviceOf, VOICE_MODEL, VOICES, voiceName, type VoiceOption, type VoiceService } from './voices.ts';
+export { DEFAULT_VOICE, MING_MOODS, mingSpeech, OWN_PREFIX, SERVICE_NAMES, serviceOf, VOICE_MODEL, VOICES, voiceName, type VoiceOption, type VoiceService } from './voices.ts';
+export { extractSounds, SOUND_SECTION, withSounds } from './sounds.ts';
+export { spokenNumbers } from './spoken-numbers.ts';
 export { invalidCharacter, loadCharacter, saveCharacter, type Character, type Pronoun } from './character.ts';
 export { decodeWav, encodeWav, wavFromPcm16 } from './wav.ts';

package/lib/types/core/sounds.d.ts

@@ -0,0 +1,21 @@
+/**
+ * Sounds she can make with the Ming voice (docs/Ming本地声音设计.md §5): she
+ * writes 〔叹气〕 or 〔笑〕 in a sentence; the tag is taken out of the subtitle
+ * and Ming's own marker ([Sigh], [Laughter]) goes at the start or the end of
+ * what is spoken. A marker in the middle of a sentence swallows the word next
+ * to it, so it is moved to whichever end the tag was nearer.
+ *
+ * With another voice the tags are simply dropped (extractCues removes every
+ * tag it does not know), and her prompt never mentions them.
+ */
+export interface Sounds {
+    /** The sentence without the sound tags. */
+    text: string;
+    before: string[];
+    after: string[];
+}
+export declare function extractSounds(text: string): Sounds;
+/** What Ming says: the sentence with its sound markers at the ends. */
+export declare function withSounds(spoken: string, sounds: Pick<Sounds, 'before' | 'after'>): string;
+/** The lines of her prompt that teach the sounds (only while she speaks with Ming). */
+export declare const SOUND_SECTION = "\u4F60\u7684\u58F0\u97F3\u80FD\u771F\u7684\u53F9\u6C14\u548C\u7B11\u51FA\u58F0\uFF1A\u5728\u53E5\u5B50\u5F00\u5934\u6216\u7ED3\u5C3E\u5199\u3014\u53F9\u6C14\u3015\u6216\u3014\u7B11\u3015\uFF08\u4E0D\u4F1A\u663E\u793A\uFF09\u3002\u53EA\u5728\u771F\u7684\u60F3\u53F9\u6C14\u3001\u7B11\u51FA\u6765\u65F6\u7528\uFF0C\u4E00\u6B21\u56DE\u590D\u6700\u591A\u4E00\u6B21\u3002";

package/lib/types/core/spoken-numbers.d.ts

@@ -0,0 +1,11 @@
+/**
+ * Digits and a few symbols written out the way they are said, for the Ming
+ * voice, which reads "2个" as "二个" (its MLX port has no text
+ * normalization). Only what turns up in her chat: counts, times, dates,
+ * years, percentages, decimals, phone numbers and codes, ranges, degrees.
+ */
+/** "2026" → 二零二六. */
+export declare function readDigits(digits: string): string;
+/** "1205" → 一千二百零五; "12" → 十二; "20000" → 两万. Longer than 12 digits: one by one. */
+export declare function readInteger(number: string): string;
+export declare function spokenNumbers(text: string): string;

package/lib/types/core/voices.d.ts

@@ -5,19 +5,35 @@
  *   needs the internet; voice/edge.ts);
  * - cosyvoice: Alibaba Cloud Model Studio (百炼) CosyVoice, model
  *   cosyvoice-v3-flash (needs a key and paid quota; voice/cosyvoice.ts). Its
- *   ids are the API's `voice` parameter, exactly as listed.
+ *   ids are the API's `voice` parameter, exactly as listed;
+ * - ming: Ming-3.0-tts on this computer (Apple chips; voice/ming.ts). Built-in
+ *   voices are `ming:<name>`; voices made on the "她" page (designed from a
+ *   description, or cloned from a recording) are `ming-own:<id>`.
  */
 export declare const VOICE_MODEL = "cosyvoice-v3-flash";
-export type VoiceService = 'edge' | 'cosyvoice';
+export type VoiceService = 'edge' | 'cosyvoice' | 'ming';
 export interface VoiceOption {
     id: string;
     name: string;
     desc: string;
     service: VoiceService;
+    /** Ming voices recorded in several moods: the moods it has (`灵风信_高兴` …). */
+    moods?: readonly string[];
 }
 export declare const VOICES: readonly VoiceOption[];
 export declare const DEFAULT_VOICE = "zh-CN-XiaoxiaoNeural";
 export declare function voiceName(id: string): string;
 /** Which service a voice belongs to; an id not in the list goes by its shape (Edge ids end in "Neural"). */
 export declare function serviceOf(id: string): VoiceService;
+/** A Ming voice made on the "她" page. */
+export declare const OWN_PREFIX = "ming-own:";
+/** Her emotions (core/cues.ts) as Ming moods; the others are said plainly. */
+export declare const MING_MOODS: Record<string, string>;
+/** How Ming says a sentence in a voice: a built-in voice (its mood version when it has one) or one of your own, plus mood and speed. */
+export declare function mingSpeech(id: string, emotion: string | undefined, speed: number): {
+    voice?: string;
+    own?: string;
+    emotion?: string;
+    speed?: string;
+};
 export declare const SERVICE_NAMES: Record<VoiceService, string>;