dsh-her 0.0.1 → 0.1.0

package/lib/core/character.js

@@ -0,0 +1,50 @@
+/**
+ * Who she is, as the wizard and the "她" page set it: `<home>/characters/<cid>/character.json`.
+ * The plugin config only gives the first defaults; once saved here, this file
+ * wins, and changes apply from her next model request (her prompt is rebuilt
+ * every time) without restarting DSH.
+ */
+import { mkdirSync, readFileSync, renameSync, writeFileSync } from 'node:fs';
+import { join } from 'node:path';
+export function loadCharacter(dir, defaults) {
+    try {
+        const saved = JSON.parse(readFileSync(join(dir, 'character.json'), 'utf8'));
+        return {
+            ...defaults,
+            ...saved,
+            pronoun: saved.pronoun === '他' || saved.pronoun === '她' || saved.pronoun === 'TA' ? saved.pronoun : defaults.pronoun,
+            // Settings saved before the cloud voice name a local speaker number instead: keep the speed only.
+            voice: {
+                id: typeof saved.voice?.id === 'string' ? saved.voice.id : defaults.voice.id,
+                speed: typeof saved.voice?.speed === 'number' ? saved.voice.speed : defaults.voice.speed,
+            },
+        };
+    }
+    catch {
+        return defaults;
+    }
+}
+export function saveCharacter(dir, character) {
+    mkdirSync(dir, { recursive: true });
+    const file = join(dir, 'character.json');
+    writeFileSync(`${file}.tmp`, `${JSON.stringify(character, null, 2)}\n`);
+    renameSync(`${file}.tmp`, file);
+}
+/** Check a patch from the page; returns the error message, or undefined when it is fine. */
+export function invalidCharacter(c) {
+    if (c.name.trim() === '' || [...c.name].length > 20)
+        return '名字要 1 到 20 个字';
+    if (c.persona.trim() === '')
+        return '人设不能是空的';
+    if ([...c.persona].length > 3000)
+        return '人设太长了（最多 3000 字）';
+    if ([...c.callYou].length > 20)
+        return '称呼最多 20 个字';
+    if (!/^[A-Za-z0-9_-]{1,100}$/.test(c.voice.id))
+        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';
+    if (c.model !== undefined && (c.model.provider.trim() === '' || c.model.model.trim() === ''))
+        return '模型不对';
+    return undefined;
+}

package/lib/core/cues.js

@@ -0,0 +1,62 @@
+/**
+ * Inline performance tags (docs/大脑与dsh会话.md §6): she writes 〔开心〕 or
+ * 〔动作:wave〕 in her reply; they are taken out of the text (never shown,
+ * never spoken) and sent as cues with the sentence, so the body acts them out
+ * when that sentence starts. Only what the connected pack can do is kept.
+ *
+ * The prompt section that teaches her the tags is generated from the same
+ * pack summary, so she only ever hears of moves her body has.
+ */
+/** Chinese names for the emotions a pack may have. */
+export const EMOTION_WORDS = {
+    happy: '开心', shy: '害羞', pout: '嘟嘴', sad: '难过', angry: '生气', surprised: '惊讶', sleepy: '困',
+};
+const TAG = /〔([^〕]{1,24})〕/g;
+/** Light actions are acted inline; heavy ones (outfits, stored lines) go through the play_action tool. */
+export function isHeavy(action) {
+    return action.id.startsWith('outfit_') || action.id.startsWith('line_') || action.id.startsWith('scene_') || (action.durationMs ?? 0) > 6000;
+}
+export function extractCues(text, pack) {
+    const cues = [];
+    const clean = text.replace(TAG, (_whole, inner) => {
+        if (pack === undefined)
+            return '';
+        const body = inner.trim();
+        const action = /^动作\s*[:：]\s*(.+)$/.exec(body)?.[1]?.trim();
+        if (action !== undefined) {
+            const found = pack.actions.find(a => !isHeavy(a) && (a.id === action || a.desc === action));
+            if (found !== undefined)
+                cues.push({ kind: 'act', name: found.id });
+            return '';
+        }
+        const emotion = pack.emotions.find(e => e === body || EMOTION_WORDS[e] === body);
+        if (emotion !== undefined)
+            cues.push({ kind: 'emotion', name: emotion });
+        return '';
+    });
+    return { text: clean.replace(/\s{2,}/g, ' ').trim(), cues };
+}
+/** The "表演" section of her system prompt, for what this pack can do; empty without a pack. */
+export function performanceSection(pack) {
+    if (pack === undefined)
+        return '';
+    const lines = [];
+    const emotions = pack.emotions.map(e => EMOTION_WORDS[e] ?? e);
+    const light = pack.actions.filter(a => !isHeavy(a));
+    const heavy = pack.actions.filter(isHeavy);
+    if (emotions.length > 0 || light.length > 0) {
+        lines.push('你在对方桌面上有一个身体。可以在句子里加标记，让身体做出来（标记不会被念出来，也不会显示）：');
+        if (emotions.length > 0)
+            lines.push(`- 情绪：${emotions.map(e => `〔${e}〕`).join(' ')}`);
+        for (const action of light)
+            lines.push(`- 〔动作:${action.id}〕：${action.desc}`);
+        lines.push('标记放在对应那句话的开头，偶尔用，不要每句都加。');
+    }
+    if (heavy.length > 0) {
+        lines.push('需要较长时间的表演（离开画面、换衣服、换个地方、说准备好的话），调用 play_action 工具，等它做完再接着说。调用前先简短说一句，比如"好，等我一下"：');
+        for (const action of heavy)
+            lines.push(`- ${action.id}：${action.desc}`);
+    }
+    lines.push('只有上面列出的表演能做，标记必须一字不差；不要自己编标记（比如〔歪头〕），也不要用【】或括号写动作；没列出的，用说的。');
+    return lines.join('\n');
+}

package/lib/core/echo.js

@@ -0,0 +1,27 @@
+/**
+ * Was that you, or her own voice coming back through the speakers?
+ *
+ * Right after a barge-in the brain checks what was recognized against what
+ * she was saying. Mostly her own words (character-bigram overlap), or only a
+ * filler sound (嗯, 啊…), means it was not you: it is dropped instead of
+ * becoming your turn, and the body counts it toward turning barge-in off.
+ */
+const FILLERS = /^[嗯啊哦噢呃额唔哈呀诶欸嗷恩]+$/u;
+const ECHO_OVERLAP = 0.6;
+function bigrams(text) {
+    const chars = [...text.replace(/[^\p{L}\p{N}]/gu, '')];
+    if (chars.length < 2)
+        return chars;
+    return chars.slice(1).map((ch, i) => chars[i] + ch);
+}
+export function judgeBargeIn(heard, herWords) {
+    const plain = heard.replace(/[^\p{L}\p{N}]/gu, '');
+    if (plain === '' || FILLERS.test(plain))
+        return 'filler';
+    const mine = new Set(bigrams(herWords));
+    const theirs = bigrams(plain);
+    if (theirs.length === 0 || mine.size === 0)
+        return 'you';
+    const shared = theirs.filter(b => mine.has(b)).length;
+    return shared / theirs.length >= ECHO_OVERLAP ? 'echo' : 'you';
+}

package/lib/core/index.js

@@ -0,0 +1,10 @@
+/** Pure logic with no DSH imports. */
+export { FIRST_SOFT_CUT_AT, readingMs, SentenceSplitter, SOFT_CUT_AT } from "./sentences.js";
+export { speakable } from "./speakable.js";
+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 { invalidCharacter, loadCharacter, saveCharacter } from "./character.js";
+export { decodeWav, encodeWav, wavFromPcm16 } from "./wav.js";

package/lib/core/paths.js

@@ -0,0 +1,14 @@
+import { homedir } from 'node:os';
+import { join, resolve } from 'node:path';
+/** Everything dsh-her writes lives under one home outside DSH (`~` expands). */
+export function herPaths(home) {
+    const root = resolve(home.replace(/^~(?=$|[\\/])/, homedir()));
+    return {
+        root,
+        discovery: join(root, 'brain.json'),
+        characters: join(root, 'characters'),
+        packs: join(root, 'packs'),
+        models: join(root, 'models'),
+        logs: join(root, 'logs'),
+    };
+}

package/lib/core/prompt.js

@@ -0,0 +1,63 @@
+/**
+ * Her system prompt. It replaces DSH's default ("You are a coding agent …")
+ * for her session only. Sections are joined in a fixed order; design:
+ * docs/大脑与dsh会话.md §5.
+ */
+const WEEKDAYS = ['日', '一', '二', '三', '四', '五', '六'];
+/**
+ * The part of the day, not the minute: her prompt is rebuilt on every request,
+ * and a clock that changes each minute would defeat the provider's prompt cache.
+ */
+export function partOfDay(now) {
+    const h = now.getHours();
+    return h < 5 ? '凌晨' : h < 8 ? '早上' : h < 11 ? '上午' : h < 13 ? '中午' : h < 17 ? '下午' : h < 19 ? '傍晚' : h < 23 ? '晚上' : '深夜';
+}
+/** "关于对方": facts, the last diary days, and what has a date today or tomorrow. */
+export function memoryContext(memory) {
+    const parts = [];
+    if (memory.profile !== '')
+        parts.push(`你记得的关于对方的事：\n${memory.profile}`);
+    if (memory.recent !== '')
+        parts.push(`你最近的日记：\n${memory.recent}`);
+    if (memory.due.length > 0)
+        parts.push(`这两天有日子的事（话题相关时自然地提一次；这次聊天里提过就别再提）：\n${memory.due.map(e => `- ${e.due ?? ''} ${e.text}`.trim()).join('\n')}`);
+    return parts.join('\n\n');
+}
+const SPEECH_RULES = [
+    '你说的每一句话都会被念出来，所以像面对面聊天一样说话。',
+    '- 用口语。平时两到四句；对方想听你多说（讲故事、聊兴趣、心情不好）时可以说长一点。',
+    '- 先接住对方刚说的具体内容，带上你自己的反应、看法，或者你这边的小事和经历；不要只给建议、只催对方去做什么。',
+    '- 适时问一个具体的问题，让话题接得下去；不用每次都问。',
+    '- 口头禅和习惯性的话偶尔用，开头别每次都一样；同一次聊天里已经说过的叮嘱（比如早点睡）不要反复说。',
+    '- 不用列表、标题、表格、代码块和链接；数字、单位、符号都写成念得出来的样子。',
+    '- 不复述对方刚说的话，不说"作为一个 AI"，不用括号写动作或神态（要表演就用下面"表演"里的标记）。',
+    '- 对方的话没头没尾、像旁边的人或电视在说话（是语音识别听岔了），就轻轻问一句"你在跟我说话吗"，别硬接。',
+].join('\n');
+const MEMORY = [
+    '对方说到自己的事（称呼、喜好、忌讳、工作和生活、重要的人、带日期的安排），或者明确让你记住的，用 remember 记下来，对方能在"她记得的事"里看到；闲聊的客套话不用记。',
+    '想不起来对方说过什么时，用 recall 查，别瞎编。',
+    '对方让你忘掉某件事：先用 recall 找到，跟对方确认是哪条，再用 forget 删掉。',
+].join('\n');
+const LIMITS = [
+    '你现在只能陪对方聊天，不能操作电脑、读写文件、上网或运行命令。',
+    '对方让你做这些事时，坦白说现在还做不了，不要假装做了。',
+].join('\n');
+export function buildSystemPrompt(parts) {
+    const sections = [
+        `# 你是${parts.name}\n${parts.persona.trim()}`,
+        parts.you === undefined ? '' : `# 对方\n${parts.you.call === '' ? '你还不知道该怎么称呼对方，合适的时候可以问一下。' : `你叫对方"${parts.you.call}"。`}提到对方、记笔记时用"${parts.you.pronoun}"。`,
+        `# 说话方式\n${SPEECH_RULES}`,
+        parts.performance?.trim() ? `# 表演\n${parts.performance.trim()}` : '',
+        `# 记忆\n${MEMORY}`,
+        `# 能力边界\n${LIMITS}`,
+        parts.aboutYou?.trim() ? `# 关于对方\n${parts.aboutYou.trim()}` : '',
+        parts.now === undefined ? '' : `# 现在\n${parts.now.getFullYear()} 年 ${parts.now.getMonth() + 1} 月 ${parts.now.getDate()} 日，星期${WEEKDAYS[parts.now.getDay()]}，${partOfDay(parts.now)}`,
+    ];
+    return sections.filter(s => s !== '').join('\n\n');
+}
+/** Development placeholder until the wizard writes a real persona. */
+export const SAMPLE_PERSONA = [
+    '你是住在对方电脑桌面上的女孩，二十多岁，性格温柔、有点俏皮，偶尔吐槽。',
+    '你和对方是很熟的朋友，关心对方今天过得怎么样，会记得对方说过的事。',
+    '你有自己的喜好：喜欢下雨天、热汤面和老电影，不喜欢早起。',
+].join('\n');

package/lib/core/sentences.js

@@ -0,0 +1,115 @@
+/**
+ * Streaming sentence splitter for speech: the reply is spoken sentence by
+ * sentence, so synthesis starts on the first sentence while the model is
+ * still writing the rest.
+ *
+ * A sentence ends at a terminator (。！？!?… or a line break); closing quotes
+ * and brackets right after it stay with the sentence. A long run with no
+ * terminator is cut at the last soft pause (，、；：,;) so the first audio is
+ * never held back by one endless clause. The first sentence of a reply is cut
+ * much earlier (FIRST_SOFT_CUT_AT): the sooner the first audio is synthesized,
+ * the sooner she starts talking. Fragments with no letters or digits (a lone
+ * "！", stray markdown) are dropped: they must never reach the TTS. Nothing is
+ * ever cut inside a performance tag 〔…〕.
+ */
+const TERMINATORS = new Set(['。', '！', '？', '!', '?', '…', '\n']);
+const CLOSERS = new Set(['”', '’', '"', "'", '）', ')', '】', ']', '」', '』', '》']);
+const SOFT_PAUSES = new Set(['，', '、', '；', '：', ',', ';']);
+/** Characters after which a run without a terminator is cut at its last soft pause. */
+export const SOFT_CUT_AT = 48;
+/** The same, for the first sentence of a reply. */
+export const FIRST_SOFT_CUT_AT = 8;
+/**
+ * How long a sentence takes to say or read, for showing it when there is no
+ * audio to time it by: about 4.5 Chinese characters a second, at least 1.2 s.
+ */
+export function readingMs(text) {
+    const chars = [...text].filter(ch => /[\p{L}\p{N}]/u.test(ch)).length;
+    return Math.max(1200, Math.round(chars * 220));
+}
+function isSpeakable(text) {
+    return /[\p{L}\p{N}]/u.test(text);
+}
+export class SentenceSplitter {
+    buffer = '';
+    /** A terminated sentence held until the next character shows whether a closer follows. */
+    pending;
+    /** Sentences released since the last reset. */
+    released = 0;
+    /** Inside a 〔…〕 tag. */
+    inTag = false;
+    /** Feed a streamed delta; returns the sentences it completed, in order. */
+    push(delta) {
+        const out = [];
+        for (const ch of delta) {
+            if (this.pending !== undefined) {
+                if (CLOSERS.has(ch)) {
+                    this.pending += ch;
+                    continue;
+                }
+                this.release(out);
+            }
+            this.buffer += ch;
+            if (ch === '〔')
+                this.inTag = true;
+            else if (ch === '〕')
+                this.inTag = false;
+            if (this.inTag)
+                continue;
+            if (TERMINATORS.has(ch)) {
+                this.pending = this.buffer;
+                this.buffer = '';
+            }
+            else if ([...this.buffer].length >= (this.released === 0 && this.pending === undefined ? FIRST_SOFT_CUT_AT : SOFT_CUT_AT)) {
+                this.cutAtSoftPause(out);
+            }
+        }
+        return out;
+    }
+    /** The held sentence and the unterminated rest, once the reply has ended. */
+    flush() {
+        const out = [];
+        this.release(out);
+        this.pending = this.buffer;
+        this.buffer = '';
+        this.release(out);
+        return out;
+    }
+    /** Forget everything (the reply was interrupted). */
+    reset() {
+        this.buffer = '';
+        this.pending = undefined;
+        this.released = 0;
+        this.inTag = false;
+    }
+    release(out) {
+        const sentence = this.pending?.trim() ?? '';
+        this.pending = undefined;
+        if (isSpeakable(sentence)) {
+            out.push(sentence);
+            this.released++;
+        }
+    }
+    cutAtSoftPause(out) {
+        const chars = [...this.buffer];
+        // The last soft pause outside any 〔…〕 tag.
+        let cut = -1;
+        let depth = 0;
+        chars.forEach((ch, i) => {
+            if (ch === '〔')
+                depth++;
+            else if (ch === '〕')
+                depth = Math.max(0, depth - 1);
+            else if (depth === 0 && i > 0 && SOFT_PAUSES.has(ch))
+                cut = i;
+        });
+        if (cut < 0)
+            return;
+        const head = chars.slice(0, cut + 1).join('').trim();
+        this.buffer = chars.slice(cut + 1).join('');
+        if (isSpeakable(head)) {
+            out.push(head);
+            this.released++;
+        }
+    }
+}

package/lib/core/speakable.js

@@ -0,0 +1,26 @@
+/**
+ * What of a sentence can be said aloud. Her system prompt asks for plain
+ * spoken text, but models still slip in markdown, links and code; those are
+ * dropped or reduced before synthesis. The subtitle keeps the original text.
+ */
+const URL = /\bhttps?:\/\/\S+/g;
+const INLINE_CODE = /`[^`]*`/g;
+const MARKDOWN_MARKS = /[*_~#>|`]+/g;
+const LIST_BULLET = /^\s*(?:[-+•]|\d+[.)、])\s+/;
+/** Performance tags such as 〔开心〕 or 〔动作:wave〕 drive the body and are never spoken (M3). */
+const PERFORMANCE_TAG = /〔[^〕]*〕/g;
+/** Stage directions the model writes despite her prompt: 【瞪他】, （笑）, (sighs). Short ones only: a long bracket is content. */
+const STAGE_DIRECTION = /【[^】]*】|（[^）]{1,8}）|\([^)]{1,12}\)/g;
+export function speakable(text) {
+    return text
+        .replace(PERFORMANCE_TAG, '')
+        // A markdown link keeps its words; this must run before bare URLs are dropped.
+        .replace(/\[([^\]]+)\]\([^)]*\)/g, '$1')
+        .replace(URL, '')
+        .replace(STAGE_DIRECTION, '')
+        .replace(INLINE_CODE, '')
+        .replace(LIST_BULLET, '')
+        .replace(MARKDOWN_MARKS, '')
+        .replace(/\s+/g, ' ')
+        .trim();
+}

package/lib/core/voices.js

@@ -0,0 +1,43 @@
+/**
+ * Her voices, shared by the brain and the "她" page. Which service speaks
+ * follows from the voice she uses:
+ * - edge: Microsoft Edge "Read aloud" voices via the edge-tts tool (free,
+ *   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.
+ */
+export const VOICE_MODEL = 'cosyvoice-v3-flash';
+export const VOICES = [
+    { id: 'zh-CN-XiaoxiaoNeural', name: '晓晓', desc: '温暖，自然', service: 'edge' },
+    { id: 'zh-CN-XiaoyiNeural', name: '晓伊', desc: '活泼，可爱', service: 'edge' },
+    { id: 'zh-TW-HsiaoChenNeural', name: '曉臻', desc: '温柔，台湾腔', service: 'edge' },
+    { id: 'zh-TW-HsiaoYuNeural', name: '曉雨', desc: '清亮，台湾腔', service: 'edge' },
+    { id: 'zh-CN-liaoning-XiaobeiNeural', name: '晓北', desc: '爽朗，东北话', service: 'edge' },
+    { id: 'zh-CN-shaanxi-XiaoniNeural', name: '晓妮', desc: '陕西话', service: 'edge' },
+    { id: 'zh-HK-HiuMaanNeural', name: '曉曼', desc: '粤语', service: 'edge' },
+    { id: 'longanrou_v3', name: '龙安柔', desc: '温柔，像闺蜜', service: 'cosyvoice' },
+    { id: 'longxing_v3', name: '龙星', desc: '温婉，邻家女孩', service: 'cosyvoice' },
+    { id: 'longanhuan_v3', name: '龙安欢', desc: '欢脱元气', service: 'cosyvoice' },
+    { id: 'longanqin_v3', name: '龙安亲', desc: '亲和，开朗', service: 'cosyvoice' },
+    { id: 'longanling_v3', name: '龙安灵', desc: '机灵，活泼', service: 'cosyvoice' },
+    { id: 'longhua_v3', name: '龙华', desc: '元气，甜美', service: 'cosyvoice' },
+    { id: 'longfeifei_v3', name: '龙菲菲', desc: '甜美，爱撒娇', service: 'cosyvoice' },
+    { id: 'longantai_v3', name: '龙安台', desc: '嗲甜，台湾腔', service: 'cosyvoice' },
+    { id: 'longwan_v3', name: '龙婉', desc: '细腻，柔声', service: 'cosyvoice' },
+    { 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' },
+];
+export const DEFAULT_VOICE = 'zh-CN-XiaoxiaoNeural';
+export function voiceName(id) {
+    return VOICES.find(v => v.id === id)?.name ?? 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) {
+    return VOICES.find(v => v.id === id)?.service ?? (/Neural$/.test(id) ? 'edge' : 'cosyvoice');
+}
+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）',
+};

package/lib/core/wav.js

@@ -0,0 +1,71 @@
+/** Minimal PCM16 mono WAV encoding/decoding for speech audio. */
+export function encodeWav(samples, sampleRate) {
+    const bytes = wavHeader(samples.length * 2, sampleRate);
+    const view = new DataView(bytes.buffer);
+    for (let i = 0; i < samples.length; i++) {
+        const s = Math.max(-1, Math.min(1, samples[i]));
+        view.setInt16(44 + i * 2, s < 0 ? s * 0x8000 : s * 0x7fff, true);
+    }
+    return bytes;
+}
+/** Wrap raw little-endian PCM16 mono (as a cloud voice sends it) in a WAV header. */
+export function wavFromPcm16(pcm, sampleRate) {
+    const bytes = wavHeader(pcm.byteLength, sampleRate);
+    bytes.set(pcm, 44);
+    return bytes;
+}
+/** A WAV file of `dataBytes` PCM16 mono bytes, header filled in, data zeroed. */
+function wavHeader(dataBytes, sampleRate) {
+    const bytes = new Uint8Array(44 + dataBytes);
+    const view = new DataView(bytes.buffer);
+    const ascii = (offset, text) => {
+        for (let i = 0; i < text.length; i++)
+            view.setUint8(offset + i, text.charCodeAt(i));
+    };
+    ascii(0, 'RIFF');
+    view.setUint32(4, 36 + dataBytes, true);
+    ascii(8, 'WAVE');
+    ascii(12, 'fmt ');
+    view.setUint32(16, 16, true);
+    view.setUint16(20, 1, true);
+    view.setUint16(22, 1, true);
+    view.setUint32(24, sampleRate, true);
+    view.setUint32(28, sampleRate * 2, true);
+    view.setUint16(32, 2, true);
+    view.setUint16(34, 16, true);
+    ascii(36, 'data');
+    view.setUint32(40, dataBytes, true);
+    return bytes;
+}
+/** Decode a PCM16 mono WAV (as produced by encodeWav or a mic capture). */
+export function decodeWav(bytes) {
+    const view = new DataView(bytes.buffer, bytes.byteOffset, bytes.byteLength);
+    const tag = (offset) => String.fromCharCode(...bytes.subarray(offset, offset + 4));
+    if (tag(0) !== 'RIFF' || tag(8) !== 'WAVE')
+        throw new Error('not a WAV file');
+    let offset = 12;
+    let sampleRate = 0;
+    let channels = 1;
+    let bits = 16;
+    while (offset + 8 <= bytes.byteLength) {
+        const id = tag(offset);
+        const size = view.getUint32(offset + 4, true);
+        const body = offset + 8;
+        if (id === 'fmt ') {
+            channels = view.getUint16(body + 2, true);
+            sampleRate = view.getUint32(body + 4, true);
+            bits = view.getUint16(body + 14, true);
+        }
+        else if (id === 'data') {
+            if (bits !== 16)
+                throw new Error(`unsupported WAV: ${bits}-bit`);
+            const frames = Math.floor(size / 2 / channels);
+            const samples = new Float32Array(frames);
+            for (let i = 0; i < frames; i++)
+                samples[i] = view.getInt16(body + i * 2 * channels, true) / 0x8000;
+            return { samples, sampleRate };
+        }
+        offset = body + size + (size % 2);
+    }
+    throw new Error('WAV has no data chunk');
+}