| 1 | import type { SoundCue, SoundEvent } from './engine'; |
| 2 | |
| 3 | /** |
| 4 | * Table noise, synthesised rather than sampled — a handful of oscillators is |
| 5 | * smaller than one .wav and needs no loading, which matters when the whole |
| 6 | * point is that the cue lands the instant the tile hits the table. |
| 7 | * |
| 8 | * Nothing here touches the audio hardware until `play` is first called, so the |
| 9 | * module is safe to import in tests and safe to load before the user has |
| 10 | * gestured at the page (browsers refuse to start audio before that). |
| 11 | */ |
| 12 | |
| 13 | let ctx: AudioContext | null = null; |
| 14 | let master: GainNode | null = null; |
| 15 | let noise: AudioBuffer | null = null; |
| 16 | let enabled = false; |
| 17 | let volume = 0.7; |
| 18 | |
| 19 | function audio(): AudioContext | null { |
| 20 | if (typeof window === 'undefined') return null; |
| 21 | if (!ctx) { |
| 22 | const Ctor = window.AudioContext ?? (window as { webkitAudioContext?: typeof AudioContext }).webkitAudioContext; |
| 23 | if (!Ctor) return null; |
| 24 | try { |
| 25 | ctx = new Ctor(); |
| 26 | } catch { |
| 27 | return null; |
| 28 | } |
| 29 | master = ctx.createGain(); |
| 30 | master.gain.value = volume; |
| 31 | master.connect(ctx.destination); |
| 32 | } |
| 33 | // Chrome parks the context until a gesture, and again when the tab sleeps. |
| 34 | if (ctx.state === 'suspended') void ctx.resume(); |
| 35 | return ctx; |
| 36 | } |
| 37 | |
| 38 | /** One second of white noise, reused by every clack. */ |
| 39 | function noiseBuffer(c: AudioContext): AudioBuffer { |
| 40 | if (!noise) { |
| 41 | noise = c.createBuffer(1, c.sampleRate, c.sampleRate); |
| 42 | const d = noise.getChannelData(0); |
| 43 | for (let i = 0; i < d.length; i++) d[i] = Math.random() * 2 - 1; |
| 44 | } |
| 45 | return noise; |
| 46 | } |
| 47 | |
| 48 | // ---- voice --------------------------------------------------------------- |
| 49 | /** |
| 50 | * 報牌 — the calls said out loud, the way they are shouted at a table: 碰, 吃, |
| 51 | * 槓, 胡了, and the name of every tile as it is discarded (三條, 五萬, 東風…). |
| 52 | * |
| 53 | * A table only ever says about fifty things, so the whole vocabulary is |
| 54 | * rendered ahead of time by `scripts/voice.mjs` into `public/voice/` and |
| 55 | * shipped as audio, rather than left to whatever speech synthesis the browser |
| 56 | * has lying around — which on Linux is usually espeak-ng, and sounds like it. |
| 57 | * Playback then goes through the same Web Audio graph as the chimes: instant, |
| 58 | * identical everywhere, and it can be cut off mid-word when the next call |
| 59 | * lands. |
| 60 | */ |
| 61 | const CALL_CLIP: Partial<Record<SoundEvent, string>> = { |
| 62 | pung: 'pung', |
| 63 | chow: 'chow', |
| 64 | kong: 'kong', |
| 65 | hu: 'hu', |
| 66 | selfDraw: 'selfDraw', |
| 67 | drawGame: 'drawGame', |
| 68 | flower: 'flower', |
| 69 | }; |
| 70 | |
| 71 | const VOICE_DIR = `${import.meta.env.BASE_URL}voice/`; |
| 72 | const encoded = new Map<string, ArrayBuffer>(); |
| 73 | const clips = new Map<string, AudioBuffer>(); |
| 74 | let voiceOn = false; |
| 75 | let fetching: Promise<void> | null = null; |
| 76 | let decoding = false; |
| 77 | let voiceGain: GainNode | null = null; |
| 78 | let speaking: AudioBufferSourceNode[] = []; |
| 79 | |
| 80 | /** |
| 81 | * Fetching needs no AudioContext, which matters: sound is on from the start, |
| 82 | * but a browser will not let the audio hardware wake until the page has been |
| 83 | * clicked. So the bytes come down straight away and are decoded later, on the |
| 84 | * first gesture — by which time the whole pack (~330 kB) is already in hand. |
| 85 | */ |
| 86 | function fetchVoicePack(): Promise<void> { |
| 87 | if (fetching) return fetching; |
| 88 | fetching = (async () => { |
| 89 | const names: string[] = await fetch(`${VOICE_DIR}manifest.json`).then((r) => r.json()); |
| 90 | await Promise.all( |
| 91 | names.map(async (name) => { |
| 92 | try { |
| 93 | encoded.set(name, await fetch(`${VOICE_DIR}${name}.mp3`).then((r) => r.arrayBuffer())); |
| 94 | } catch { |
| 95 | // One missing clip just means that one call stays silent. |
| 96 | } |
| 97 | }), |
| 98 | ); |
| 99 | })().catch(() => { |
| 100 | fetching = null; // let a later toggle try again |
| 101 | }); |
| 102 | return fetching; |
| 103 | } |
| 104 | |
| 105 | /** Decode what was fetched. Runs once, the first time there is a live context. */ |
| 106 | function decodeVoicePack(c: AudioContext) { |
| 107 | if (decoding || clips.size || !encoded.size) return; |
| 108 | decoding = true; |
| 109 | for (const [name, bytes] of encoded) { |
| 110 | c.decodeAudioData(bytes) |
| 111 | .then((buf) => clips.set(name, buf)) |
| 112 | .catch(() => {}); |
| 113 | } |
| 114 | encoded.clear(); // decodeAudioData takes ownership of the buffers |
| 115 | } |
| 116 | |
| 117 | export function setVoiceEnabled(on: boolean) { |
| 118 | voiceOn = on; |
| 119 | if (on) void fetchVoicePack(); |
| 120 | } |
| 121 | |
| 122 | /** |
| 123 | * Speak one or more clips back to back. Calls land on top of each other at |
| 124 | * table pace, so anything still talking is cut off: the newest call is the |
| 125 | * only one that matters. |
| 126 | * |
| 127 | * Returns how long it will be talking for, in seconds — zero if it is not |
| 128 | * going to say anything at all. That is what lets the table hold still while |
| 129 | * somebody is speaking, and only for as long as they are. |
| 130 | */ |
| 131 | function say(...names: string[]): number { |
| 132 | const c = audio(); |
| 133 | if (!voiceOn || !enabled || !c || !master) return 0; |
| 134 | if (!clips.size) { |
| 135 | // Nothing to say yet: either still coming down the wire, or waiting for |
| 136 | // this very gesture to be allowed to decode. |
| 137 | decodeVoicePack(c); |
| 138 | void fetchVoicePack(); |
| 139 | return 0; |
| 140 | } |
| 141 | |
| 142 | for (const src of speaking) { |
| 143 | try { |
| 144 | src.stop(); |
| 145 | } catch { |
| 146 | // already finished |
| 147 | } |
| 148 | } |
| 149 | speaking = []; |
| 150 | |
| 151 | if (!voiceGain) { |
| 152 | // A little above the chimes: the chime is a nudge, the call is the message. |
| 153 | voiceGain = c.createGain(); |
| 154 | voiceGain.gain.value = 1.4; |
| 155 | voiceGain.connect(master); |
| 156 | } |
| 157 | |
| 158 | const from = c.currentTime + 0.02; |
| 159 | let at = from; |
| 160 | for (const name of names) { |
| 161 | const buf = clips.get(name); |
| 162 | if (!buf) continue; |
| 163 | const src = c.createBufferSource(); |
| 164 | src.buffer = buf; |
| 165 | src.connect(voiceGain); |
| 166 | src.start(at); |
| 167 | speaking.push(src); |
| 168 | at += buf.duration + 0.05; |
| 169 | } |
| 170 | return at - from; |
| 171 | } |
| 172 | |
| 173 | export function setSoundEnabled(on: boolean) { |
| 174 | enabled = on; |
| 175 | // Reach for the hardware only once the page has been interacted with — |
| 176 | // sound is on from the start, and a context built before that is refused |
| 177 | // anyway. When this runs from the click that toggled it, it unlocks there. |
| 178 | if (on && navigator.userActivation?.hasBeenActive !== false) { |
| 179 | const c = audio(); |
| 180 | if (c) decodeVoicePack(c); |
| 181 | } |
| 182 | } |
| 183 | |
| 184 | export function setVolume(v: number) { |
| 185 | volume = v; |
| 186 | if (master && ctx) master.gain.setTargetAtTime(v, ctx.currentTime, 0.01); |
| 187 | } |
| 188 | |
| 189 | /** Envelope helper: a tone that starts at `gain` and decays away over `dur`. */ |
| 190 | function tone( |
| 191 | c: AudioContext, |
| 192 | opts: { |
| 193 | freq: number; |
| 194 | at: number; |
| 195 | dur: number; |
| 196 | gain: number; |
| 197 | type?: OscillatorType; |
| 198 | /** Slide to this frequency over the life of the note. */ |
| 199 | to?: number; |
| 200 | }, |
| 201 | ) { |
| 202 | const o = c.createOscillator(); |
| 203 | const g = c.createGain(); |
| 204 | o.type = opts.type ?? 'sine'; |
| 205 | const t = c.currentTime + opts.at; |
| 206 | o.frequency.setValueAtTime(opts.freq, t); |
| 207 | if (opts.to !== undefined) o.frequency.exponentialRampToValueAtTime(opts.to, t + opts.dur); |
| 208 | g.gain.setValueAtTime(0.0001, t); |
| 209 | g.gain.exponentialRampToValueAtTime(opts.gain, t + 0.004); |
| 210 | g.gain.exponentialRampToValueAtTime(0.0001, t + opts.dur); |
| 211 | o.connect(g).connect(master!); |
| 212 | o.start(t); |
| 213 | o.stop(t + opts.dur + 0.02); |
| 214 | } |
| 215 | |
| 216 | /** Bamboo-on-bamboo: a filtered noise burst with a short woody body under it. */ |
| 217 | function clack(c: AudioContext, at: number, gain = 1) { |
| 218 | const t = c.currentTime + at; |
| 219 | const src = c.createBufferSource(); |
| 220 | src.buffer = noiseBuffer(c); |
| 221 | src.playbackRate.value = 1; |
| 222 | const bp = c.createBiquadFilter(); |
| 223 | bp.type = 'bandpass'; |
| 224 | bp.frequency.value = 2100; |
| 225 | bp.Q.value = 1.1; |
| 226 | const g = c.createGain(); |
| 227 | g.gain.setValueAtTime(0.55 * gain, t); |
| 228 | g.gain.exponentialRampToValueAtTime(0.0001, t + 0.075); |
| 229 | src.connect(bp).connect(g).connect(master!); |
| 230 | src.start(t, Math.random() * 0.5); |
| 231 | src.stop(t + 0.1); |
| 232 | tone(c, { freq: 320, to: 160, at, dur: 0.09, gain: 0.28 * gain, type: 'triangle' }); |
| 233 | } |
| 234 | |
| 235 | /** Struck-metal chime: a fundamental with two inharmonic partials over it. */ |
| 236 | function bell(c: AudioContext, freq: number, at: number, gain = 1, dur = 0.55) { |
| 237 | tone(c, { freq, at, dur, gain: 0.3 * gain, type: 'sine' }); |
| 238 | tone(c, { freq: freq * 2.02, at, dur: dur * 0.7, gain: 0.13 * gain, type: 'sine' }); |
| 239 | tone(c, { freq: freq * 3.01, at, dur: dur * 0.4, gain: 0.06 * gain, type: 'sine' }); |
| 240 | } |
| 241 | |
| 242 | /** |
| 243 | * Tile on tile, out in the middle. Not a `SoundEvent`: the engine never emits |
| 244 | * one of these, because it is not something that happens in the game — it is a |
| 245 | * thrown tile landing among the others, and only the physics knows when. |
| 246 | * |
| 247 | * `strength` is 0..1, straight off the impact, so a tile dropped in is a tick |
| 248 | * and one thrown hard is a crack. |
| 249 | */ |
| 250 | let lastKnock = 0; |
| 251 | export function knock(strength: number) { |
| 252 | if (!enabled) return; |
| 253 | const c = audio(); |
| 254 | if (!c || !master) return; |
| 255 | // A tile skittering across a pool of sixty can generate a dozen contacts in a |
| 256 | // tenth of a second, and hearing all of them is a rattle, not a table. |
| 257 | if (c.currentTime - lastKnock < 0.045) return; |
| 258 | lastKnock = c.currentTime; |
| 259 | clack(c, 0, 0.15 + strength * 0.6); |
| 260 | } |
| 261 | |
| 262 | /** |
| 263 | * A tile thrown so hard it comes off the pool and into somebody's tiles. |
| 264 | * |
| 265 | * The middle stops where the four players' hands start, so a tile that reaches |
| 266 | * the edge with any speed left has been thrown at whoever is sitting there — |
| 267 | * and at a real table that gets you told. The clack is the tile arriving; the |
| 268 | * grumble under it is the reaction, and 報牌 puts words to it. |
| 269 | * |
| 270 | * Only the physics knows this happened, so like `knock` it is not a |
| 271 | * `SoundEvent`: the engine neither knows nor cares where a discard ended up. |
| 272 | * |
| 273 | * Returns how long it will be talking for, so the table can wait it out — a |
| 274 | * computer player reaching for the tile over the top of somebody complaining |
| 275 | * about it is the whole thing landing on nobody. |
| 276 | */ |
| 277 | let lastBarge = 0; |
| 278 | export function barge(strength: number): number { |
| 279 | if (!enabled) return 0; |
| 280 | const c = audio(); |
| 281 | if (!c || !master) return 0; |
| 282 | // One complaint per throw, not one per bounce. |
| 283 | if (c.currentTime - lastBarge < 1.6) return 0; |
| 284 | lastBarge = c.currentTime; |
| 285 | |
| 286 | clack(c, 0, 0.5 + strength * 0.5); |
| 287 | // Two notes falling, which is what a grumble sounds like. |
| 288 | tone(c, { freq: 300, to: 190, at: 0.05, dur: 0.14, gain: 0.16, type: 'triangle' }); |
| 289 | tone(c, { freq: 240, to: 150, at: 0.2, dur: 0.22, gain: 0.13, type: 'triangle' }); |
| 290 | // Never the same line twice running: the repeat is what makes a stock phrase |
| 291 | // sound like a machine, and this one fires often enough for that to show. |
| 292 | let pick = Math.floor(Math.random() * WATCH_IT.length); |
| 293 | if (WATCH_IT[pick] === lastWords) pick = (pick + 1) % WATCH_IT.length; |
| 294 | lastWords = WATCH_IT[pick]; |
| 295 | return say(lastWords); |
| 296 | } |
| 297 | |
| 298 | /** |
| 299 | * What gets said. 喂，小心點 and 輕一點啦 straight, and then the way it would |
| 300 | * actually come out at a table — 是在哈囉, 你牌品很差欸, 這是麻將不是棒球. |
| 301 | * Rendered by scripts/voice.mjs; a missing clip just means silence. |
| 302 | */ |
| 303 | const WATCH_IT = [ |
| 304 | 'watchIt', |
| 305 | 'watchIt2', |
| 306 | 'watchIt3', |
| 307 | 'watchIt4', |
| 308 | 'watchIt5', |
| 309 | 'watchIt6', |
| 310 | 'watchIt7', |
| 311 | 'watchIt8', |
| 312 | ]; |
| 313 | let lastWords = ''; |
| 314 | |
| 315 | /** |
| 316 | * Each cue has its own shape so it is identifiable without looking up — the |
| 317 | * whole reason for the feature is a player noticing across the table. |
| 318 | */ |
| 319 | export function play(cue: SoundCue | SoundEvent) { |
| 320 | const { kind, tile } = typeof cue === 'string' ? { kind: cue, tile: undefined } : cue; |
| 321 | if (!enabled) return; |
| 322 | |
| 323 | // A discard is announced by name; a claim by what was called; a flower by |
| 324 | // both, since "春" on its own tells you nothing about why you heard it. |
| 325 | if (kind === 'discard') { |
| 326 | if (tile !== undefined) say(`t${tile}`); |
| 327 | } else if (kind === 'flower' && tile !== undefined) { |
| 328 | say('flower', `t${tile}`); |
| 329 | } else if (CALL_CLIP[kind]) { |
| 330 | say(CALL_CLIP[kind]); |
| 331 | } |
| 332 | |
| 333 | const c = audio(); |
| 334 | if (!c || !master) return; |
| 335 | |
| 336 | switch (kind) { |
| 337 | // A tile laid down. Dry and short so it never masks the chime after it. |
| 338 | case 'discard': |
| 339 | clack(c, 0); |
| 340 | break; |
| 341 | // Somebody's claim window is open. Rising two-note bell, easy to hear over |
| 342 | // talk, and deliberately unlike the clack that precedes it. |
| 343 | case 'claimWindow': |
| 344 | bell(c, 784, 0.06, 1, 0.4); // G5 |
| 345 | bell(c, 1047, 0.19, 0.9, 0.55); // C6 |
| 346 | break; |
| 347 | // 碰 / 吃 — the tile going down alongside the pair it joins. |
| 348 | case 'pung': |
| 349 | case 'chow': |
| 350 | clack(c, 0, 0.8); |
| 351 | clack(c, 0.055, 0.9); |
| 352 | bell(c, kind === 'pung' ? 660 : 587, 0.02, 0.5, 0.3); |
| 353 | break; |
| 354 | // Four tiles going down, plus a replacement off the back of the wall. |
| 355 | case 'kong': |
| 356 | clack(c, 0, 0.7); |
| 357 | clack(c, 0.05, 0.8); |
| 358 | clack(c, 0.1, 0.9); |
| 359 | clack(c, 0.15, 1); |
| 360 | bell(c, 523, 0.16, 0.55, 0.5); |
| 361 | break; |
| 362 | case 'flower': |
| 363 | bell(c, 1319, 0, 0.45, 0.35); |
| 364 | break; |
| 365 | // 胡牌 — a rising figure, the only cue that takes its time. |
| 366 | case 'hu': |
| 367 | case 'selfDraw': |
| 368 | bell(c, 523, 0, 1, 0.5); |
| 369 | bell(c, 659, 0.11, 1, 0.5); |
| 370 | bell(c, 784, 0.22, 1, 0.6); |
| 371 | bell(c, 1047, 0.33, 1.1, 1.1); |
| 372 | break; |
| 373 | // 流局 — the same figure, falling and dulled. |
| 374 | case 'drawGame': |
| 375 | bell(c, 523, 0, 0.7, 0.5); |
| 376 | bell(c, 415, 0.13, 0.7, 0.6); |
| 377 | bell(c, 330, 0.26, 0.7, 0.9); |
| 378 | break; |
| 379 | // Sixteen tiles apiece, so a scatter rather than a beat. |
| 380 | case 'deal': |
| 381 | for (let i = 0; i < 9; i++) clack(c, i * 0.045 + Math.random() * 0.02, 0.4); |
| 382 | break; |
| 383 | case 'undo': |
| 384 | tone(c, { freq: 700, to: 330, at: 0, dur: 0.22, gain: 0.18, type: 'triangle' }); |
| 385 | break; |
| 386 | } |
| 387 | } |