diff --git a/web/README.md b/web/README.md
index 189bee4..ec53be1 100644
--- a/web/README.md
+++ b/web/README.md
@@ -2,8 +2,9 @@
 
 Bàn số của trò chơi "Lô tô" — SvelteKit app.
 
-Two routes: `/` for players, `/master` for the host (quản trò) — calls
-numbers, shows a tracking board, and has its own player card to play along.
+Single page (`/`). The host enables "Chế độ quản trò" in settings to
+reveal the master panel inline below their player card; called numbers
+get spoken aloud in Vietnamese using bundled MP3 clips (no runtime TTS).
 
 See `docs/` for architecture, code standards, and deployment.
 
@@ -36,3 +37,18 @@ npm run build:gh      # /loto basePath, for tiennm99.github.io/loto manual expor
 
 Static export to `build/`. Deployed to Cloudflare Pages from `main`
 (set up via the CF dashboard — see `docs/deployment-guide.md`).
+
+## Regenerating audio
+
+Vietnamese voice clips live in `static/audio/{voiceId}/`. Generated
+once with [edge-tts](https://github.com/rany2/edge-tts) (free, no API
+key) — runtime never calls TTS. To regenerate (e.g., to add a new voice
+that Microsoft ships, or to change wording):
+
+```bash
+pip install edge-tts
+python3 scripts/generate-audio.py
+```
+
+The script auto-discovers every `vi-*` voice and writes a manifest the
+app reads on next reload.
diff --git a/web/docs/codebase-summary.md b/web/docs/codebase-summary.md
index 953e3ea..3c759eb 100644
--- a/web/docs/codebase-summary.md
+++ b/web/docs/codebase-summary.md
@@ -12,7 +12,7 @@
 | File | Purpose |
 |------|---------|
 | `src/lib/PlayerBoard.svelte` | Reusable player card (9×9 grid rendered as 3 stacked 3×9 mini-cards: Tân Tân / An khang thịnh vượng / Tân Tân tốt nhất). Tall (3:5 on mobile; wider on sm+) cells with condensed bold black numbers (`tan-tan-num` font stack), white number cells, purple empty cells by default. Handles crossed state, bingo popup, "Chờ X" toast. Two header actions: "Tạo bảng mới" (regenerate grid) and "Xoá đánh dấu" (clear all marks, keep grid — only shown when a grid exists). Accepts `storagePrefix` prop for multi-card isolation. Empty cells use `--empty-cell-bg` CSS var from settings store. |
-| `src/lib/SettingsButton.svelte` | Gear icon + modal. 4 fieldsets: Giao diện (theme auto/light/dark), Chế độ quản trò (master mode toggle), Tự động xổ (auto-call + speed 1–10s), Màu ô trống (10 Excel color swatches). Reset-to-default button. Mounted on `/`. |
+| `src/lib/SettingsButton.svelte` | Gear icon + modal. 5 fieldsets: Giao diện (theme auto/light/dark), Chế độ quản trò (master mode toggle), Tự động xổ (auto-call + speed 1–10s), Âm thanh (voice toggles for master + player + voice picker from manifest), Màu ô trống (10 Excel color swatches). Reset-to-default button. Mounted on `/`. |
 | `src/lib/MasterPanel.svelte` | Host controls. New game / draw, "Số vừa xổ" hero token, "Thứ tự đã xổ" history list, 11×9 last-digit-aligned tracking grid (with circular tokens + draw-order overlay). No host's own player card (the player already has one above). "Xổ số" / "Bắt đầu / Dừng" button bound to auto-call. Mounted conditionally on `/` when `settings.masterMode === true`. |
 | `src/lib/PageFooter.svelte` | Footer with tagline ("Made by miti99 with ❤️ SVG icon") + link. Mounted on `/`. |
 
@@ -20,7 +20,10 @@
 | File | Purpose |
 |------|---------|
 | `src/lib/game-logic.js` | Stateless utilities: generateGrid (constraint-aware picker — exact 5 per row & per col, ascending-sorted columns, soft "no 3 consecutive filled cols per row" via rejection sampling), saveGrid, loadGrid, saveCrossedState, loadCrossedState, isRowComplete, getWaitingNumber. |
-| `src/lib/settings-store.svelte.js` | Reactive global UI settings via Svelte 5 runes. Stores 5 keys: `theme` (enum: "auto" / "light" / "dark"), `masterMode` (bool), `autoCallEnabled` (bool), `autoCallSpeed` (1–10), `emptyCellColor` (hex). Persisted to localStorage `loto_settings`. Pushes values to CSS vars and `<html class="dark">` on `:root`. Per-key validators preserve old data. |
+| `src/lib/vietnamese-number.js` | `numberToVietnamese(n)` — pure utility mapping 0..90 to spoken Vietnamese, with tonal exceptions (15 → "mười lăm", 21 → "hai mươi mốt", 25 → "hai mươi lăm"). Out-of-range falls back to `String(n)`. |
+| `src/lib/voice.js` | Bundled-MP3 playback. Exports `playNumber(n)`, `playWaiting(n)` (sequences cho + N), `playBingo()`, `cancelPlayback()`. Lazy `<audio>` cache, cancel-then-play, token-based cancel ensures stale promises can't resume after a new event. Reads active voice from `settings.voice`; URLs go through `import { base } from "$app/paths"` for basePath safety. |
+| `src/lib/audio-manifest.js` | Re-exports `static/audio/manifest.json` as `VOICES` (array) + `VOICE_IDS` (Set) + `DEFAULT_VOICE`. Manifest is generated by `scripts/generate-audio.py`. |
+| `src/lib/settings-store.svelte.js` | Reactive global UI settings via Svelte 5 runes. Stores 8 keys: `theme` (enum: "auto" / "light" / "dark"), `masterMode` (bool), `autoCallEnabled` (bool), `autoCallSpeed` (1–10), `emptyCellColor` (hex), `voiceEnabledMaster` (bool), `voiceEnabledPlayer` (bool), `voice` (string id matching audio manifest). Persisted to localStorage `loto_settings`. Pushes values to CSS vars and `<html class="dark">` on `:root`. Per-key validators preserve old data. |
 
 ### Styling
 | File | Purpose |
@@ -30,8 +33,9 @@
 ### Tests
 | File | Purpose |
 |------|---------|
-| `src/lib/game-logic.test.js` | 26 unit tests: generateGrid shape (9×9, 5 per row/col, no duplicates), column ranges & ascending sort, row completion, waiting number detection, persistence (saveGrid/loadGrid/saveCrossedState/loadCrossedState with validators). |
-| `src/lib/settings-store.test.js` | 27 unit tests: defaults, loadSettings (restore 5 keys from localStorage, apply CSS vars, toggle dark class, handle empty/corrupt), saveSettings, resetSettings, theme toggle (auto → OS pref detection), master mode, auto-call + speed, color validation. |
+| `src/lib/game-logic.test.js` | 27 unit tests: generateGrid shape (9×9, 5 per row/col, no duplicates), column ranges & ascending sort, no-3-consecutive soft constraint, row completion, waiting number detection, persistence (saveGrid/loadGrid/saveCrossedState/loadCrossedState with validators). |
+| `src/lib/settings-store.test.js` | 31 unit tests: defaults (incl. voice keys), loadSettings (restore 8 keys, apply CSS vars, toggle dark class, handle empty/corrupt), saveSettings, resetSettings, theme toggle (auto → OS pref detection), master mode, auto-call + speed, color validation, voice round-trip + invalid-id fallback. |
+| `src/lib/vietnamese-number.test.js` | 40 unit tests: ones (0–9), teens (10–19 incl. mười lăm), 20–90 incl. mốt and lăm exceptions, out-of-range fall-through. |
 
 ### Configuration
 | File | Purpose |
@@ -44,6 +48,9 @@
 | `.gitignore` | Excludes node_modules, build, .env.local, etc. |
 | `.env.example` | codeserver profile vars (CODESERVER_HOST, CODESERVER_PORT). |
 | `static/_redirects` | Cloudflare Pages: `/* / 301` — every unknown path 301-redirects to homepage. Static assets are served first so this only fires on misses. |
+| `static/audio/{voiceId}/*.mp3` | Pre-generated Vietnamese voice clips (1–90 + cho + kinh per voice). Generated once by `scripts/generate-audio.py` (edge-tts). Runtime never calls TTS. |
+| `static/audio/manifest.json` | Voice list `{ id, edgeName, label, gender }[]`. Written by the generator, imported by `audio-manifest.js`. |
+| `scripts/generate-audio.py` | One-shot Python script: discovers every `vi-*` edge-tts voice, generates 92 clips per voice, writes manifest. Run after voice/wording changes. |
 
 ## Key Data Structures
 
diff --git a/web/docs/development-roadmap.md b/web/docs/development-roadmap.md
index 0564c6a..cf92195 100644
--- a/web/docs/development-roadmap.md
+++ b/web/docs/development-roadmap.md
@@ -6,9 +6,6 @@ now", see `docs/project-overview-pdr.md` and `docs/codebase-summary.md`.
 
 ## Idea Phase
 
-### Sound Effects on Bingo
-Play celebratory chime or "Kinh!" voice snippet when row completes. Could use Web Audio API or `<audio>` tag. **Status**: Idea (no demand yet)
-
 ### Undo Last Cell
 Allow player to undo the most recent cross/uncross action. Requires change history or state snapshot. **Status**: Idea (low priority)
 
diff --git a/web/docs/project-overview-pdr.md b/web/docs/project-overview-pdr.md
index 7e950e7..1db750d 100644
--- a/web/docs/project-overview-pdr.md
+++ b/web/docs/project-overview-pdr.md
@@ -35,10 +35,12 @@ That format is intentionally out of scope.
   a master board with circular tokens (pink ring for 1–49, green for 50–90,
   cream fill on called, dim on uncalled, red ring + scale on the most
   recent draw), and may also play their own card.
-- **Settings**: gear icon in headers opens a modal with 4 fieldsets:
+- **Settings**: gear icon in headers opens a modal with 5 fieldsets:
   - Giao diện: theme switcher (auto / light / dark; auto mirrors OS pref)
   - Chế độ quản trò: toggle to show MasterPanel on `/`
   - Tự động xổ: enable auto-call; speed slider (1–10 seconds)
+  - Âm thanh: voice toggles for master + player + Vietnamese voice picker
+    (every `vi-*` edge-tts voice; default Hoài Mỹ)
   - Màu ô trống: color picker + 10 Excel preset swatches
   Persisted to `loto_settings`.
 
@@ -70,6 +72,7 @@ That format is intentionally out of scope.
 - **Runtime**: Svelte 5 runes ($state, $derived, $effect, $props)
 - **Styling**: Tailwind CSS 4 (utility-first, animations)
 - **Persistence**: localStorage (no backend)
+- **Audio**: pre-generated MP3 clips bundled under `static/audio/{voiceId}/`. Built once by `scripts/generate-audio.py` using free `edge-tts` (Microsoft Neural). Runtime plays via plain HTML5 Audio — no TTS API at runtime.
 - **Deploy**: Cloudflare Pages (root domain), GitHub Pages fallback (`/loto`)
 - **Dev Profile**: code-server compatible via `/absproxy/{port}` basePath + HMR proxy config
 
@@ -94,6 +97,9 @@ State is entirely client-side. Each card / panel instance uses a unique localSto
 - [x] Player can click cells to toggle crossed state.
 - [x] Player can clear all marks on the current card without regenerating it (confirm prompt when marks exist).
 - [x] No row has 3 consecutive filled columns (soft constraint, rejection-sampled).
+- [x] Master speaks the called number aloud in Vietnamese (bundled MP3, no runtime TTS API).
+- [x] Player hears "Chờ N" when a row is one away and "Kinh" on bingo.
+- [x] Voice picker in settings — switch among every Vietnamese edge-tts voice.
 - [x] Bingo popup triggers when row is complete, shows row number and "Kinh!" message.
 - [x] Player may keep marking after a Kinh — no game-end lock.
 - [x] Toast notifications show "Chờ X" before bingo (one number remaining).
diff --git a/web/plans/260427-0759-vietnamese-voice-calls/phase-01-vietnamese-number-utility.md b/web/plans/260427-0759-vietnamese-voice-calls/phase-01-vietnamese-number-utility.md
new file mode 100644
index 0000000..05a31b8
--- /dev/null
+++ b/web/plans/260427-0759-vietnamese-voice-calls/phase-01-vietnamese-number-utility.md
@@ -0,0 +1,92 @@
+# Phase 1 — Vietnamese number words utility
+
+## Context
+- [plan.md](plan.md). No dependencies; self-contained pure module.
+
+## Overview
+- Priority: P0 (everything else depends on it)
+- Status: TODO
+- Effort: ~15 min
+
+## Goal
+
+Pure function `numberToVietnamese(n)` mapping `1..90` to spoken Vietnamese,
+honoring the tonal exceptions that show up in lô tô numbers.
+
+## Tonal rules (the parts that matter)
+
+| Position | Rule | Example |
+|---|---|---|
+| Standalone unit | "không" / "một" / ... / "chín" | `5 = năm` |
+| 10 alone | "mười" | `10 = mười` |
+| 11–19, units 1–9 | "mười X" | `12 = mười hai` |
+| 11–19, unit = 5 | **"mười lăm"** (not "năm") | `15 = mười lăm` |
+| 20–90 tens place | "X mươi" (not "mười") | `30 = ba mươi` |
+| 21+ unit = 1 | **"…mốt"** (not "một") | `21 = hai mươi mốt` |
+| 21+ unit = 5 | **"…lăm"** (not "năm") | `25 = hai mươi lăm` |
+| Tens 0 | drop unit | `40 = bốn mươi` |
+
+(`tư` for 4 in unit position is regional — we stay with `bốn` for
+consistency.)
+
+## Files
+
+| File | Change |
+|---|---|
+| `src/lib/vietnamese-number.js` | NEW — exports `numberToVietnamese(n)` |
+| `src/lib/vietnamese-number.test.js` | NEW — covers edge cases below |
+
+## Implementation sketch
+
+```js
+const ONES = [
+  "không", "một", "hai", "ba", "bốn",
+  "năm", "sáu", "bảy", "tám", "chín",
+];
+
+/** @param {number} n integer 0..90 */
+export function numberToVietnamese(n) {
+  if (!Number.isInteger(n) || n < 0 || n > 90) return String(n);
+  if (n < 10) return ONES[n];
+  if (n === 10) return "mười";
+  if (n < 20) {
+    const u = n - 10;
+    return u === 5 ? "mười lăm" : `mười ${ONES[u]}`;
+  }
+  const t = Math.floor(n / 10);
+  const u = n % 10;
+  const tens = `${ONES[t]} mươi`;
+  if (u === 0) return tens;
+  if (u === 1) return `${tens} mốt`;
+  if (u === 5) return `${tens} lăm`;
+  return `${tens} ${ONES[u]}`;
+}
+```
+
+## Tests (must cover)
+
+| n | Expected |
+|---|---|
+| 1 | "một" |
+| 5 | "năm" |
+| 10 | "mười" |
+| 11 | "mười một" |
+| 15 | "mười lăm" |
+| 19 | "mười chín" |
+| 20 | "hai mươi" |
+| 21 | "hai mươi mốt" |
+| 25 | "hai mươi lăm" |
+| 45 | "bốn mươi lăm" |
+| 81 | "tám mươi mốt" |
+| 90 | "chín mươi" |
+
+Plus: out-of-range fall-through (`91 → "91"`, `0.5 → "0.5"`, `-1 → "-1"`).
+
+## Success criteria
+
+- All ~12 table cases pass.
+- `numberToVietnamese` is pure (no side effects, idempotent, no DOM).
+- Test file slots into existing Vitest setup with no config changes.
+
+## Next
+- Phase 2 imports it from the voice module.
diff --git a/web/plans/260427-0759-vietnamese-voice-calls/phase-02-audio-generation-script.md b/web/plans/260427-0759-vietnamese-voice-calls/phase-02-audio-generation-script.md
new file mode 100644
index 0000000..5a3f56d
--- /dev/null
+++ b/web/plans/260427-0759-vietnamese-voice-calls/phase-02-audio-generation-script.md
@@ -0,0 +1,204 @@
+# Phase 2 — Audio generation script (build-time, all Vietnamese voices)
+
+## Context
+- [plan.md](plan.md). Output: per-voice MP3 sets committed under
+  `static/audio/{voiceId}/`, ready for Phase 3's playback module.
+
+## Overview
+- Priority: P0
+- Status: TODO
+- Effort: ~30 min (incl. ~5-10 min generation runtime)
+- Run frequency: once per voice change (rare).
+
+## What gets generated
+
+```
+static/audio/
+├── hoai-my/                       (vi-VN-HoaiMyNeural — female)
+│   ├── 1.mp3        "một"
+│   ├── 2.mp3        "hai"
+│   ├── ... (90)
+│   ├── 90.mp3       "chín mươi"
+│   ├── cho.mp3      "Chờ"
+│   └── kinh.mp3     "Kinh"
+├── nam-minh/                      (vi-VN-NamMinhNeural — male)
+│   ├── 1.mp3
+│   ├── ... (same 92 names)
+│   └── kinh.mp3
+└── manifest.json                  (voice list, generated)
+```
+
+92 clips × N voices. As of writing, edge-tts ships 2 Vietnamese voices
+(female + male) — total ~184 files, ~2.2 MB. Script auto-discovers
+voices via `edge-tts.list_voices()`, so new voices added by Microsoft
+get picked up on the next regen.
+
+## Voice ID mapping
+
+| edge-tts voice | Folder ID | Display label |
+|---|---|---|
+| `vi-VN-HoaiMyNeural` | `hoai-my` | "Hoài Mỹ (nữ)" |
+| `vi-VN-NamMinhNeural` | `nam-minh` | "Nam Minh (nam)" |
+| (future `vi-VN-XxxNeural`) | `xxx` | "Xxx" — script slug-ifies |
+
+Folder ID = lowercase + kebab-case + drop locale prefix + drop
+"Neural" suffix. Conversion is deterministic; script writes the
+manifest so the JS side never has to mirror Python.
+
+## Tooling
+
+| Tool | Why |
+|---|---|
+| Python 3 | already on dev machine |
+| `edge-tts` (`pip install edge-tts`) | free, no API key, MS Neural quality |
+
+User installs `edge-tts` once on the dev machine; project doesn't
+add Python deps to `package.json` (build-time only).
+
+## File: `scripts/generate-audio.py`
+
+```python
+#!/usr/bin/env python3
+"""Generate Vietnamese audio clips (1-90 + Chờ + Kinh) for every
+edge-tts Vietnamese voice. Output committed to static/audio/{voiceId}/
+and shipped with the app — runtime never calls TTS."""
+import asyncio, json, os, re, sys
+
+OUT_ROOT = os.path.join(os.path.dirname(__file__), "..", "static", "audio")
+
+ONES = ["không", "một", "hai", "ba", "bốn",
+        "năm", "sáu", "bảy", "tám", "chín"]
+
+def number_to_vietnamese(n: int) -> str:
+    if n < 10: return ONES[n]
+    if n == 10: return "mười"
+    if n < 20:
+        u = n - 10
+        return "mười lăm" if u == 5 else f"mười {ONES[u]}"
+    t, u = divmod(n, 10)
+    tens = f"{ONES[t]} mươi"
+    if u == 0: return tens
+    if u == 1: return f"{tens} mốt"
+    if u == 5: return f"{tens} lăm"
+    return f"{tens} {ONES[u]}"
+
+def voice_id(short_name: str) -> str:
+    """vi-VN-HoaiMyNeural -> hoai-my"""
+    name = short_name.split("-")[-1]                # HoaiMyNeural
+    name = re.sub(r"Neural$", "", name)             # HoaiMy
+    name = re.sub(r"(?<!^)(?=[A-Z])", "-", name)    # Hoai-My
+    return name.lower()                             # hoai-my
+
+async def synth(text: str, voice: str, out: str):
+    import edge_tts
+    await edge_tts.Communicate(text, voice).save(out)
+    print(f"  {out}  ←  \"{text}\"")
+
+async def main():
+    import edge_tts
+    all_voices = await edge_tts.list_voices()
+    vi_voices = [v for v in all_voices if v["Locale"].startswith("vi-")]
+    if not vi_voices:
+        sys.exit("No Vietnamese voices found in edge-tts.")
+
+    manifest = {"voices": []}
+    for v in vi_voices:
+        vid = voice_id(v["ShortName"])
+        out_dir = os.path.join(OUT_ROOT, vid)
+        os.makedirs(out_dir, exist_ok=True)
+        print(f"\n→ {v['ShortName']}  →  static/audio/{vid}/")
+
+        tasks = []
+        for n in range(1, 91):
+            tasks.append(synth(number_to_vietnamese(n),
+                               v["ShortName"],
+                               os.path.join(out_dir, f"{n}.mp3")))
+        tasks.append(synth("Chờ",  v["ShortName"], os.path.join(out_dir, "cho.mp3")))
+        tasks.append(synth("Kinh", v["ShortName"], os.path.join(out_dir, "kinh.mp3")))
+        await asyncio.gather(*tasks)
+
+        # Display label: gender + given name
+        gender_vi = "nữ" if v["Gender"].lower() == "female" else "nam"
+        given = re.sub(r"(?<!^)(?=[A-Z])", " ", re.sub(r"Neural$", "",
+                       v["ShortName"].split("-")[-1])).strip()
+        manifest["voices"].append({
+            "id": vid,
+            "edgeName": v["ShortName"],
+            "label": f"{given} ({gender_vi})",
+            "gender": v["Gender"].lower(),
+        })
+
+    with open(os.path.join(OUT_ROOT, "manifest.json"), "w", encoding="utf-8") as f:
+        json.dump(manifest, f, ensure_ascii=False, indent=2)
+    print(f"\nWrote manifest with {len(manifest['voices'])} voice(s).")
+
+if __name__ == "__main__":
+    try:
+        asyncio.run(main())
+    except ImportError:
+        sys.exit("Install dep first: pip install edge-tts")
+```
+
+## How to run
+
+```bash
+pip install edge-tts                          # one-time, dev-machine only
+python3 scripts/generate-audio.py
+git add static/audio/
+git commit -m "chore(audio): regenerate Vietnamese clips"
+```
+
+## manifest.json — the JS side reads this
+
+```json
+{
+  "voices": [
+    { "id": "hoai-my",  "edgeName": "vi-VN-HoaiMyNeural",  "label": "Hoai My (nữ)",  "gender": "female" },
+    { "id": "nam-minh", "edgeName": "vi-VN-NamMinhNeural", "label": "Nam Minh (nam)", "gender": "male" }
+  ]
+}
+```
+
+Phase 4's settings UI fetches `manifest.json` once (or imports it via
+Vite's `?json` if we want it baked into the bundle) to populate the
+voice picker. Adding a new voice = re-run the script, ship the new
+folder + manifest, no code change required.
+
+## Repository policy / README addendum
+
+```md
+### Regenerating audio
+Vietnamese voice clips live in `static/audio/{voiceId}/`. To regenerate
+(e.g., to add a new edge-tts voice or change wording):
+
+    pip install edge-tts
+    python3 scripts/generate-audio.py
+```
+
+## Edge cases
+
+| Case | Handling |
+|---|---|
+| Network fails mid-generation | `asyncio.gather` raises; rerun (idempotent) |
+| edge-tts upstream removes a voice | Committed MP3s still play; only regen blocks. Manifest stays accurate to what's on disk. |
+| New Vietnamese voice arrives | Auto-included in next regen; Phase 4 UI picks it up via manifest |
+| Voice name collision after slug-ify | Defensive: script aborts with message if two `id`s collide |
+
+## Success criteria
+
+- `python3 scripts/generate-audio.py` writes 92 MP3s per Vietnamese
+  voice into `static/audio/{voiceId}/` plus `manifest.json`.
+- Each `45.mp3` says "bốn mươi lăm" in the corresponding voice.
+- `cho.mp3` + `45.mp3` in sequence sounds like "Chờ bốn mươi lăm".
+- Repo grows by ~2.2 MB (acceptable).
+
+## Risks
+
+| Risk | Mitigation |
+|---|---|
+| edge-tts endpoint changes break regen later | MP3s already committed; runtime unaffected |
+| Manifest drifts from disk (manual edit) | Script always rewrites manifest; treat as generated artifact |
+| Bandwidth — 184 files lazy-loaded | Browser caches; player only ever uses one voice's 92 files at a time |
+
+## Next
+- Phase 3 reads `manifest.json` + active-voice setting to build URLs.
diff --git a/web/plans/260427-0759-vietnamese-voice-calls/phase-03-audio-playback-module.md b/web/plans/260427-0759-vietnamese-voice-calls/phase-03-audio-playback-module.md
new file mode 100644
index 0000000..37d4e7c
--- /dev/null
+++ b/web/plans/260427-0759-vietnamese-voice-calls/phase-03-audio-playback-module.md
@@ -0,0 +1,171 @@
+# Phase 3 — Audio playback module
+
+## Context
+- [plan.md](plan.md). Depends on Phase 2 (committed MP3s).
+- Single module both Master and Player import from.
+
+## Overview
+- Priority: P0
+- Status: TODO
+- Effort: ~25 min
+
+## Goal
+
+Three exports both call sites use:
+
+```
+playNumber(n)    // master: plays {voice}/{n}.mp3
+playWaiting(n)   // player: plays {voice}/cho.mp3 then {voice}/{n}.mp3
+playBingo()      // player: plays {voice}/kinh.mp3
+cancelPlayback() // pauses any in-flight audio (used on new game / new card)
+```
+
+Active voice resolved at call time from `settings.voice`. URLs go
+through `import { base } from "$app/paths"` for basePath safety.
+
+## Files
+
+| File | Change |
+|---|---|
+| `src/lib/voice.js` | NEW — playback functions, `<audio>` cache, sequencer |
+
+## Implementation sketch
+
+```js
+import { base } from "$app/paths";
+import { settings } from "$lib/settings-store.svelte.js";
+
+/** @type {Map<string, HTMLAudioElement>} */
+const cache = new Map();
+
+/** @type {HTMLAudioElement | null} */
+let activePrimary = null; // currently-playing or chained chain leader
+
+/** @param {string} url */
+function getAudio(url) {
+  let a = cache.get(url);
+  if (!a) {
+    a = new Audio(url);
+    a.preload = "auto";
+    cache.set(url, a);
+  }
+  return a;
+}
+
+function clipUrl(name) {
+  return `${base}/audio/${settings.voice}/${name}.mp3`;
+}
+
+export function cancelPlayback() {
+  if (activePrimary) {
+    activePrimary.onended = null;
+    activePrimary.pause();
+    activePrimary.currentTime = 0;
+    activePrimary = null;
+  }
+}
+
+/**
+ * Play a single clip; resolves when it ends (or errors / is canceled).
+ * @param {string} url
+ */
+function playClip(url) {
+  return new Promise((resolve) => {
+    const a = getAudio(url);
+    a.currentTime = 0;
+    activePrimary = a;
+    const done = () => {
+      a.onended = null;
+      a.onerror = null;
+      if (activePrimary === a) activePrimary = null;
+      resolve();
+    };
+    a.onended = done;
+    a.onerror = done;
+    a.play().catch(done); // autoplay blocked, etc.
+  });
+}
+
+/** @param {number} n */
+export function playNumber(n) {
+  cancelPlayback();
+  void playClip(clipUrl(String(n)));
+}
+
+/** @param {number} n */
+export function playWaiting(n) {
+  cancelPlayback();
+  // Sequence: cho → number. Re-check cancelPlayback between clips so
+  // a fast user click can interrupt mid-sequence.
+  (async () => {
+    const cho = getAudio(clipUrl("cho"));
+    const num = getAudio(clipUrl(String(n)));
+    activePrimary = cho;
+    await playClip(clipUrl("cho"));
+    if (activePrimary !== null && activePrimary !== num) return; // canceled
+    await playClip(clipUrl(String(n)));
+  })();
+}
+
+export function playBingo() {
+  cancelPlayback();
+  void playClip(clipUrl("kinh"));
+}
+```
+
+## Manifest loading
+
+Phase 4 imports `manifest.json` directly via Vite's `?json` query so
+it lands in the JS bundle (no fetch needed):
+
+```js
+// In settings-store.svelte.js or SettingsButton.svelte
+import manifest from "../../static/audio/manifest.json";
+```
+
+If `static/audio/manifest.json` doesn't exist yet (script not run on
+fresh checkout), import fails at build time. Phase 7 mentions this in
+the README.
+
+Alternative if static-import is awkward: copy the manifest into `src/lib/`
+during the script run. Defer that micro-optimization.
+
+## Edge cases
+
+| Case | Behavior |
+|---|---|
+| Voice clip 404 (e.g., user upgraded but didn't pull MP3s) | `audio.onerror` fires, promise resolves silently — game keeps working |
+| User changes voice mid-game | Next call uses new URL; cache holds both voices' clips. Memory cost: each Audio node ~few KB; acceptable. |
+| Rapid manual draws or auto-call at 1s | Each call hits `cancelPlayback` first; only the latest plays |
+| iOS Safari autoplay block | All call sites trigger from a click handler (Tạo bảng / Xổ số / cell click) — autoplay policy satisfied |
+| `playWaiting` interrupted mid-sequence | Activity flag check between clips bails out; prevents stale tail playing on top of a newer event |
+| Browser tab backgrounded | Browser pauses Audio elements automatically; nothing for us to do |
+
+## Tests
+
+Skip unit tests for this module — it's a thin wrapper around the
+DOM Audio element. Phase 1 covers the only branchy logic
+(`numberToVietnamese`). Manual smoke test in browser is the validation.
+
+If we ever want to test it: stub `Audio` with a mock that records
+`play()` / `pause()` calls. Defer until there's a real bug to chase.
+
+## Success criteria
+
+- `import { playNumber, playWaiting, playBingo } from "$lib/voice.js"`
+  works in components.
+- `playNumber(45)` plays the chosen voice's "bốn mươi lăm" clip.
+- `playWaiting(45)` plays "Chờ" then "bốn mươi lăm" with no audible
+  gap > 200 ms.
+- Rapid calls cancel prior playback (no overlap, no backlog).
+- Module no-ops cleanly if a clip is missing (logs only).
+
+## Risks
+
+| Risk | Mitigation |
+|---|---|
+| Cache grows unbounded across voices | At most 92 × N voices ≈ 184 nodes after full warmup. Bounded; acceptable. |
+| `base` not resolved at module-init time on SSR | We're SSR-disabled (`ssr: false` in `+layout.js`); `base` is always set when these functions run |
+
+## Next
+- Phase 4 wires settings + voice picker.
diff --git a/web/plans/260427-0759-vietnamese-voice-calls/phase-04-settings-integration.md b/web/plans/260427-0759-vietnamese-voice-calls/phase-04-settings-integration.md
new file mode 100644
index 0000000..0312d0a
--- /dev/null
+++ b/web/plans/260427-0759-vietnamese-voice-calls/phase-04-settings-integration.md
@@ -0,0 +1,139 @@
+# Phase 4 — Settings: 3 keys + voice picker UI
+
+## Context
+- [plan.md](plan.md). Adds 3 keys to `settings-store` and a new
+  "Âm thanh" fieldset with toggles + a voice list.
+
+## Overview
+- Priority: P1
+- Status: TODO
+- Effort: ~25 min
+
+## New settings keys
+
+| Key | Type | Default | Effect |
+|---|---|---|---|
+| `voiceEnabledMaster` | boolean | `true` | Speak called number when master draws |
+| `voiceEnabledPlayer` | boolean | `true` | Speak "Chờ N" / "Kinh" on player events |
+| `voice` | string (voiceId) | first manifest entry | Which voice to play |
+
+All persisted under existing `loto_settings` blob, same merge-and-validate
+pattern as the other keys.
+
+## Manifest source
+
+```js
+// In settings-store.svelte.js (or a small lib/audio-manifest.js):
+import manifest from "../../static/audio/manifest.json";
+export const VOICES = manifest.voices;
+// e.g. [{ id: "hoai-my", edgeName: "...", label: "Hoai My (nữ)", gender: "female" }, ...]
+```
+
+Vite imports JSON natively — manifest is part of the JS bundle, no
+fetch needed at runtime.
+
+## Files
+
+| File | Change |
+|---|---|
+| `src/lib/audio-manifest.js` (NEW, optional) | Re-exports manifest as `VOICES` so settings-store and SettingsButton share the source |
+| `src/lib/settings-store.svelte.js` | Add 3 keys to `DEFAULTS`, validate in `loadSettings` (boolean for two, string-in-VOICES-ids for `voice`) |
+| `src/lib/settings-store.test.js` | 4-5 tests: defaults, persistence round-trip, corruption fallback, invalid voiceId fallback |
+| `src/lib/SettingsButton.svelte` | NEW "Âm thanh" fieldset: 2 toggles + voice radio/pill list |
+
+## settings-store.svelte.js — pattern
+
+```js
+import { VOICES } from "$lib/audio-manifest.js";
+
+const VOICE_IDS = new Set(VOICES.map((v) => v.id));
+const DEFAULT_VOICE = VOICES[0]?.id ?? "hoai-my"; // graceful fallback
+
+const DEFAULTS = {
+  // ...existing 5 keys...
+  voiceEnabledMaster: true,
+  voiceEnabledPlayer: true,
+  voice: DEFAULT_VOICE,
+};
+
+// In loadSettings(), per-key validate:
+//   voiceEnabledMaster:  typeof v === "boolean"  → keep, else default
+//   voiceEnabledPlayer:  typeof v === "boolean"  → keep, else default
+//   voice:               VOICE_IDS.has(v)        → keep, else default
+```
+
+No CSS variables, no `<html>` class — pure values read by other modules.
+
+## SettingsButton.svelte — UI fieldset
+
+Place between "Tự động xổ" and "Màu ô trống" (logical grouping with
+feedback features).
+
+```svelte
+<fieldset class="mb-5">
+  <legend class="text-sm font-semibold ...">Âm thanh</legend>
+  <p class="text-xs text-slate-400 dark:text-slate-500 mb-2">
+    Đọc số bằng tiếng Việt
+  </p>
+
+  <button onclick={() => toggleVoiceMaster()} class="...pill...">
+    Quản trò: {settings.voiceEnabledMaster ? "Bật" : "Tắt"}
+  </button>
+
+  <button onclick={() => toggleVoicePlayer()} class="...pill mt-2...">
+    Người chơi (Chờ / Kinh): {settings.voiceEnabledPlayer ? "Bật" : "Tắt"}
+  </button>
+
+  <div class="mt-3">
+    <p class="text-xs text-slate-500 dark:text-slate-400 mb-1">Giọng đọc</p>
+    <div class="grid grid-cols-2 gap-2">
+      {#each VOICES as v}
+        <button
+          onclick={() => saveSettings({ voice: v.id })}
+          class="px-3 py-2 rounded-lg border-2 text-sm
+                 {settings.voice === v.id
+                   ? 'border-emerald-500 bg-emerald-50 dark:bg-emerald-950/30'
+                   : 'border-slate-300 dark:border-slate-600'}"
+        >
+          {v.label}
+        </button>
+      {/each}
+    </div>
+  </div>
+</fieldset>
+```
+
+Optional polish: tiny "▶" preview button next to each voice that calls
+`playNumber(45)` so users can hear before committing. Skip in v1; add
+if asked.
+
+## Tests
+
+In `settings-store.test.js`:
+
+1. Defaults: `voiceEnabledMaster === true`, `voiceEnabledPlayer === true`,
+   `voice === DEFAULT_VOICE` (first manifest entry).
+2. Round-trip: `saveSettings({ voice: "nam-minh" })` persists and reloads.
+3. Corrupt boolean → falls back to default for that key, others survive.
+4. Invalid voice id (`"made-up"`) → falls back to default voice.
+5. Empty manifest case (theoretical — if somehow no voices) → default voice
+   string still set without crashing other settings.
+
+## Success criteria
+
+- Toggling either pill flips the boolean immediately (rune-reactive).
+- Picking a voice updates `settings.voice` and next call uses new clips.
+- Page reload preserves all 3 settings.
+- Vitest still green; settings-store tests grow by 4-5.
+
+## Risks
+
+| Risk | Mitigation |
+|---|---|
+| Manifest absent on fresh checkout (Phase 2 not run) | Build fails fast at import — README tells contributors to run the script first |
+| Manifest renamed a voice id; saved id no longer valid | `loadSettings` falls back to default voice (covered by test #4) |
+| User picks a voice but its clips are missing on disk | Phase 3's `audio.onerror` keeps app functional; UI still works (silent) |
+
+## Next
+- Phase 5 reads `voiceEnabledMaster` to gate master speech.
+- Phase 6 reads `voiceEnabledPlayer` to gate player speech.
diff --git a/web/plans/260427-0759-vietnamese-voice-calls/phase-05-master-integration.md b/web/plans/260427-0759-vietnamese-voice-calls/phase-05-master-integration.md
new file mode 100644
index 0000000..875289f
--- /dev/null
+++ b/web/plans/260427-0759-vietnamese-voice-calls/phase-05-master-integration.md
@@ -0,0 +1,76 @@
+# Phase 5 — Master: play called number
+
+## Context
+- [plan.md](plan.md). Depends on Phases 2–4.
+- Single hook point: `handleDrawNext()` in `MasterPanel.svelte:119`.
+  Both manual click ("Xổ số") and auto-call interval funnel through it.
+
+## Overview
+- Priority: P1
+- Status: TODO
+- Effort: ~5 min
+
+## Files
+
+| File | Change |
+|---|---|
+| `src/lib/MasterPanel.svelte` | Import `playNumber`, `cancelPlayback`. Call `playNumber(next)` after state update; `cancelPlayback()` on new game. |
+
+## Diff (illustrative)
+
+```diff
++ import { playNumber, cancelPlayback } from "$lib/voice.js";
+  // ...
+
+  function handleNewGame() {
+    if (state && !confirm("Bạn có muốn tạo ván mới không?")) return;
++   cancelPlayback();
+    autoRunning = false;
+    state = createFreshState();
+    lastCalled = null;
+  }
+
+  function handleDrawNext() {
+    if (!state || state.remaining.length === 0) return;
+    const next = state.remaining[0];
+    state = {
+      called: [...state.called, next],
+      remaining: state.remaining.slice(1),
+    };
+    lastCalled = next;
++   if (settings.voiceEnabledMaster) playNumber(next);
+  }
+```
+
+## Why one hook covers both manual + auto
+
+The auto-call `$effect` (line ~96 in current file) calls
+`handleDrawNext()` directly. No duplicate hook needed. ✓
+
+## Edge cases
+
+| Case | Behavior |
+|---|---|
+| User mashes "Xổ số" rapidly | `playNumber` cancels prior playback (Phase 3) — only the latest plays. ✓ |
+| Auto-call at 1s with ~1.5s clips | Same — cancel-then-play. Truncation is the cost of fast mode; users adjust speed if needed. ✓ |
+| Toggle `voiceEnabledMaster` off mid-game | Next draw stays silent. Already-playing clip finishes (acceptable) or `cancelPlayback()` if we want immediate silence — KISS, skip. |
+| `state.remaining.length === 0` | Early-return guards; `playNumber` never called. ✓ |
+| Voice changed mid-game | Next draw uses new voice's clip. ✓ |
+
+## Success criteria
+
+- Click "Xổ số" with `voiceEnabledMaster=true` → audible Vietnamese
+  number in chosen voice.
+- Auto-call announces every drawn number.
+- Toggle off → next draws silent.
+- "Ván mới" → any pending playback stops.
+
+## Risks
+
+| Risk | Mitigation |
+|---|---|
+| Clip 404 (corrupt static dir) | Phase 3's `audio.onerror` no-ops silently |
+| iOS Safari autoplay block | First "Xổ số" click satisfies the user-gesture requirement |
+
+## Next
+- Phase 6 mirrors this in PlayerBoard.
diff --git a/web/plans/260427-0759-vietnamese-voice-calls/phase-06-player-integration.md b/web/plans/260427-0759-vietnamese-voice-calls/phase-06-player-integration.md
new file mode 100644
index 0000000..7d04b1d
--- /dev/null
+++ b/web/plans/260427-0759-vietnamese-voice-calls/phase-06-player-integration.md
@@ -0,0 +1,104 @@
+# Phase 6 — Player: play "Chờ N" and "Kinh"
+
+## Context
+- [plan.md](plan.md). Depends on Phases 2–4.
+- Hook into the existing waiting/celebrated `$effect` block in
+  `PlayerBoard.svelte` (lines 83–108). Sound is a sibling output to
+  the toast and bingo popup, gated by the existing `notifiedWaitingRows`
+  / `celebratedRows` sets to dedupe.
+
+## Overview
+- Priority: P1
+- Status: TODO
+- Effort: ~10 min
+
+## Files
+
+| File | Change |
+|---|---|
+| `src/lib/PlayerBoard.svelte` | Import `playWaiting`, `playBingo`, `cancelPlayback`. Call beside existing toast + popup triggers. Cancel on `handleGenerate` / `handleClear`. |
+
+## Diff (illustrative)
+
+```diff
++ import { playWaiting, playBingo, cancelPlayback } from "$lib/voice.js";
++ import { settings } from "$lib/settings-store.svelte.js";
+  // ...
+
+  $effect(() => {
+    if (!grid || crossed.length === 0) return;
+
+    // Pass 1: at most one bingo popup per render
+    for (let i = 0; i < grid.length; i++) {
+      if (!celebratedRows.has(i) && isRowComplete(grid, crossed, i)) {
+        celebratedRows.add(i);
+        notifiedWaitingRows.add(i);
+        congratsRow = i + 1;
+        showCongrats = true;
++       if (settings.voiceEnabledPlayer) playBingo();
+        break;
+      }
+    }
+
+    // Pass 2: update waiting state for every non-celebrated row
+    for (let i = 0; i < grid.length; i++) {
+      if (celebratedRows.has(i)) continue;
+      const waitNum = getWaitingNumber(grid, crossed, i);
+      if (waitNum !== null && !notifiedWaitingRows.has(i)) {
+        notifiedWaitingRows.add(i);
+        showToast(`Chờ ${waitNum}`);
++       if (settings.voiceEnabledPlayer) playWaiting(waitNum);
+      } else if (waitNum === null && notifiedWaitingRows.has(i)) {
+        notifiedWaitingRows.delete(i);
+      }
+    }
+  });
+
+  function handleGenerate() {
+    if (grid && !confirm("Bạn có muốn tạo lại bảng không?")) return;
++   cancelPlayback();
+    // ...existing...
+  }
+
+  function handleClear() {
+    if (!grid) return;
+    const hasMarks = crossed.some((row) => row.some(Boolean));
+    if (hasMarks && !confirm("Bạn có muốn xoá tất cả đánh dấu không?")) return;
++   cancelPlayback();
+    // ...existing...
+  }
+```
+
+## De-duplication is already there
+
+The existing `notifiedWaitingRows` and `celebratedRows` sets gate the
+toast + popup; the audio call hangs on the same gates so we never
+play twice for the same row. ✓
+
+## Edge cases
+
+| Case | Behavior |
+|---|---|
+| Multiple rows enter "waiting" same render | Each calls `playWaiting`, but cancel-then-play (Phase 3) means only the LAST plays. Acceptable; queueing is a future tune-up. |
+| Bingo + waiting on same render | Impossible — `if (celebratedRows.has(i)) continue` short-circuits the waiting check. ✓ |
+| Player crosses fast, toggles waiting on/off | `notifiedWaitingRows.delete(i)` lets re-firing happen next entry into waiting. ✓ |
+| Player uncrosses winning cell | No "un-bingo" sound. Acceptable — the popup already requires explicit dismiss. |
+| Master + player speech race | Latest call wins (cancel-then-play). Player events are low-frequency, so this is rarely user-visible. |
+
+## Success criteria
+
+- Cross 4 of 5 cells in a row → "Chờ {N}" plays.
+- Cross the 5th → bingo popup + "Kinh" plays.
+- "Tạo bảng mới" / "Xoá đánh dấu" stop any pending playback.
+- `voiceEnabledPlayer=false` → silent on both events; toast and popup
+  still appear.
+
+## Risks
+
+| Risk | Mitigation |
+|---|---|
+| Master and player speech overlap (e.g., player marks a cell during master draw) | Cancel-then-play documented behavior; latest event wins |
+| iOS Safari first-event no-play (no prior gesture) | First waiting state requires 4 prior cell clicks → gesture already satisfied. ✓ |
+
+## Next
+- Phase 7 docs sync.
diff --git a/web/plans/260427-0759-vietnamese-voice-calls/phase-07-tests-and-docs.md b/web/plans/260427-0759-vietnamese-voice-calls/phase-07-tests-and-docs.md
new file mode 100644
index 0000000..93360f3
--- /dev/null
+++ b/web/plans/260427-0759-vietnamese-voice-calls/phase-07-tests-and-docs.md
@@ -0,0 +1,92 @@
+# Phase 7 — Tests pass + docs sync
+
+## Context
+- [plan.md](plan.md). Final phase. No new features — verification +
+  doc drift cleanup.
+
+## Overview
+- Priority: P2
+- Status: TODO
+- Effort: ~15 min
+
+## Verification
+
+```bash
+npm test
+npm run build
+```
+
+Expected:
+- 54 (current) + ~12 (Phase 1 number tests) + ~5 (Phase 4 settings
+  tests) ≈ 71 tests passing.
+- Build clean. `static/audio/` carried into `build/audio/` by
+  adapter-static automatically.
+
+Manual quick-check after build:
+```bash
+ls build/audio/hoai-my/ | head   # confirm clips bundled
+```
+
+## Manual smoke test (browser)
+
+| Step | Expected |
+|---|---|
+| Fresh load `/`, click "Tạo bảng mới" | grid renders, no audio |
+| Cross 4 cells in a row | "Chờ {N}" plays (chosen voice) |
+| Cross the 5th | "Kinh" plays + bingo popup |
+| Open Settings → toggle "Người chơi" off → cross to bingo | popup, no audio |
+| Toggle master mode on, click "Xổ số" | called number plays |
+| Settings → switch voice (Hoài Mỹ → Nam Minh) → click "Xổ số" | new voice plays |
+| Auto-call at 1s | each draw plays, no backlog |
+| Click "Ván mới" mid-clip | playback stops |
+| Hard refresh, settings persisted | voice selection survives |
+| DevTools → Network tab on first draw | one MP3 fetch from `/audio/{voice}/{n}.mp3`; subsequent draws use cache |
+
+## Docs to sync
+
+| File | Change |
+|---|---|
+| `docs/codebase-summary.md` | Add `vietnamese-number.js`, `voice.js`, `audio-manifest.js` rows. Update `SettingsButton.svelte` row to mention 5 fieldsets (was 4) + voice picker. Update `settings-store.svelte.js` description: 8 keys (was 5). New "Static Assets" subsection mentioning `static/audio/{voiceId}/`. |
+| `docs/project-overview-pdr.md` | Settings section: add "Âm thanh" fieldset (master toggle, player toggle, voice picker). Add acceptance items for voice-on-number, voice-on-Chờ/Kinh, voice picker. Tech Stack: note "Free build-time TTS via edge-tts". |
+| `docs/development-roadmap.md` | Drop "Sound Effects on Bingo" idea entry — done in better form (TTS, configurable voice). |
+| `docs/deployment-guide.md` | Brief note: regenerating audio = `pip install edge-tts && python3 scripts/generate-audio.py`; no impact on CF Pages build (MP3s are committed). |
+| `README.md` | "Regenerating audio" subsection (one short paragraph). |
+
+Skip changes to `system-architecture.md` and `code-standards.md` —
+voice is a leaf module with no architectural surface.
+
+## Roadmap
+
+The roadmap currently lists "Sound Effects on Bingo" as an Idea. After
+this lands, **delete that section** — it's done in the better form
+(neural TTS, configurable voice).
+
+## Cross-language `numberToVietnamese` divergence guard
+
+`scripts/generate-audio.py` and `src/lib/vietnamese-number.js` both
+implement the same Vietnamese number rules. They have to stay in sync.
+Mitigations:
+
+1. JS unit tests (Phase 1) cover all tonal exceptions; if JS drifts,
+   tests catch it.
+2. Manual: after editing either file, re-read both side by side.
+3. (Future) auto-check via a small CI script that calls Python +
+   loads the JS via Node and compares 1..90. Skip in v1.
+
+## Out of scope (intentionally)
+
+- `voice.test.js` — Phase 3 explained why (DOM Audio mock churn not
+  worth it; the only branchy logic is `numberToVietnamese`, already
+  covered by Phase 1 tests).
+- Component tests for SettingsButton voice toggles + voice picker —
+  manual smoke test plus settings-store unit tests cover it.
+- E2E tests for audio playback (requires Playwright + audio capture).
+
+## Success criteria
+
+- All tests pass.
+- `npm run build` clean. `build/audio/` populated.
+- Manual smoke checklist all green.
+- Docs updated to reflect new files, new settings keys, new voice
+  picker.
+- No stale "Sound Effects" idea in roadmap.
diff --git a/web/plans/260427-0759-vietnamese-voice-calls/plan.md b/web/plans/260427-0759-vietnamese-voice-calls/plan.md
new file mode 100644
index 0000000..108ffd0
--- /dev/null
+++ b/web/plans/260427-0759-vietnamese-voice-calls/plan.md
@@ -0,0 +1,116 @@
+---
+slug: vietnamese-voice-calls
+created: 2026-04-27
+updated: 2026-04-27
+status: completed
+completedAt: 2026-04-27
+mode: fast
+blockedBy: []
+blocks: []
+---
+
+# Vietnamese voice calls — bundled MP3 clips, no runtime APIs
+
+Speak the called number aloud (master), and "Chờ {N}" / "Kinh!"
+(player) when those events fire. **Free + offline + zero runtime API
+calls** — pre-generate Vietnamese audio once during dev (using the
+free `edge-tts` Microsoft neural voice) and ship the MP3s as static
+assets bundled with the app.
+
+## Why this approach (vs Web Speech API)
+
+User wants the audio bundled in the project — no runtime API or browser
+TTS dependency.
+
+| Concern | Web Speech API | Bundled MP3 (this plan) |
+|---|---|---|
+| Runtime cost | free | free |
+| Voice quality | OS-dependent (Linux Chrome often has no `vi`) | uniform, neural, Vietnamese |
+| Bundle weight | 0 KB | ~2.2 MB (92 clips × ~12 KB × 2 voices, lazy-loaded) |
+| Offline | yes | yes |
+| Privacy | TTS may upload text to OS service | fully self-hosted |
+| Build complexity | none | one-time `python` script |
+
+Trade-off accepted: ~1.1 MB added to repo + initial page weight, in
+exchange for consistent, high-quality Vietnamese audio everywhere.
+
+## Decisions (locked)
+
+- **TTS engine for generation**: `edge-tts` (Python). Free, no API key,
+  Microsoft Neural Voice quality.
+- **All Vietnamese voices generated**: script auto-discovers every
+  `vi-*` voice in edge-tts and produces a separate clip set per voice
+  under `static/audio/{voiceId}/`. As of writing: `vi-VN-HoaiMyNeural`
+  (female) and `vi-VN-NamMinhNeural` (male) — script picks up new
+  voices automatically on regen.
+- **Active voice configurable in settings**: new `voice` key (string,
+  matches a voice's `id` from `manifest.json`). Default = first voice
+  in manifest (`hoai-my`). User picks in the "Âm thanh" fieldset.
+- **Runtime playback**: HTML5 `<audio>` via plain `new Audio(url)`. No
+  Web Audio API. Files lazy-loaded on first request; browser caches
+  after that.
+- **Asset layout**: `static/audio/{voiceId}/{1..90}.mp3` +
+  `static/audio/{voiceId}/cho.mp3` + `static/audio/{voiceId}/kinh.mp3`,
+  plus `static/audio/manifest.json` (voice list). 92 files per voice.
+- **"Chờ N" composition**: play `cho.mp3` then `{N}.mp3` in sequence
+  via `audio.onended`. Saves 90 files vs pre-generating "Chờ 1.mp3"
+  through "Chờ 90.mp3".
+- **Cancel-then-play**: each request stops any in-flight audio first.
+  Auto-call at 1s with ~1.5s clips → without cancel we'd grow a
+  backlog. Latest number is what matters.
+- **basePath aware**: URLs go through `import { base } from "$app/paths"`
+  so the same code works on `loto.miti99.com`, `tiennm99.github.io/loto`,
+  and code-server `/absproxy/{port}`.
+- **Three settings keys**:
+  - `voiceEnabledMaster: boolean` (default `true`) — speak called number
+  - `voiceEnabledPlayer: boolean` (default `true`) — speak "Chờ N" / "Kinh"
+  - `voice: string` (default first manifest entry) — which voice to use
+- **No volume slider, no rate/pitch tuning** — device volume + a curated
+  voice list. KISS.
+- **Hook points**:
+  - Master: inside `handleDrawNext()` (covers manual click + auto-call)
+  - Player: inside the existing waiting/celebrated `$effect` in
+    `PlayerBoard.svelte` (uses existing dedupe sets)
+
+## Phases
+
+| # | Phase | File |
+|---|---|---|
+| 1 | Vietnamese number utility (1–90) — provides TTS prompts + tests | DONE | [phase-01-vietnamese-number-utility.md](phase-01-vietnamese-number-utility.md) |
+| 2 | Audio generation script (all `vi-*` voices via edge-tts) → commit `static/audio/{voiceId}/*.mp3` + `manifest.json` | DONE (script + placeholder manifest; user runs script to materialize MP3s) | [phase-02-audio-generation-script.md](phase-02-audio-generation-script.md) |
+| 3 | `voice.js` playback module — `<audio>` lazy cache, sequencer, cancel, voice-aware URL builder | DONE | [phase-03-audio-playback-module.md](phase-03-audio-playback-module.md) |
+| 4 | Settings: 3 keys (master/player toggles + voice picker) + "Âm thanh" fieldset + tests | DONE | [phase-04-settings-integration.md](phase-04-settings-integration.md) |
+| 5 | MasterPanel: announce drawn number in `handleDrawNext()` | DONE | [phase-05-master-integration.md](phase-05-master-integration.md) |
+| 6 | PlayerBoard: announce "Chờ N" / "Kinh" in waiting/bingo $effect | DONE | [phase-06-player-integration.md](phase-06-player-integration.md) |
+| 7 | Tests pass + docs sync | DONE (98/98 tests; reviewer fixes applied) | [phase-07-tests-and-docs.md](phase-07-tests-and-docs.md) |
+
+## Files Touched / Created
+
+| File | Phase |
+|---|---|
+| `src/lib/vietnamese-number.js` (NEW) | 1 |
+| `src/lib/vietnamese-number.test.js` (NEW) | 1 |
+| `scripts/generate-audio.py` (NEW) | 2 |
+| `scripts/README.md` (NEW, optional) | 2 |
+| `static/audio/*.mp3` (NEW, 92 files, committed) | 2 |
+| `src/lib/voice.js` (NEW) | 3 |
+| `src/lib/settings-store.svelte.js` | 4 |
+| `src/lib/settings-store.test.js` | 4 |
+| `src/lib/SettingsButton.svelte` | 4 |
+| `src/lib/MasterPanel.svelte` | 5 |
+| `src/lib/PlayerBoard.svelte` | 6 |
+| `docs/codebase-summary.md` | 7 |
+| `docs/project-overview-pdr.md` | 7 |
+| `docs/development-roadmap.md` | 7 (drop stale "Sound Effects" idea) |
+| `.gitattributes` (optional, for LFS-free MP3 handling) | 2 |
+| `README.md` (mention audio regen) | 2 |
+
+## Out of Scope
+
+- Audio sprite / single-file packing (HTTP/2 makes 92 small fetches fine)
+- Volume slider, voice picker, alt voices, male voice, child voice
+- Pre-rendered "Chờ 1..90" combo clips (sequence at runtime)
+- Sound effects on bingo (chimes / confetti audio)
+- Service Worker pre-caching (PWA territory; future plan)
+- Lô tô fairground rhythm / chant style — flat reading only
+- Runtime TTS APIs (edge-tts is build-time only; no runtime API)
diff --git a/web/scripts/generate-audio.py b/web/scripts/generate-audio.py
new file mode 100755
index 0000000..c4b33d3
--- /dev/null
+++ b/web/scripts/generate-audio.py
@@ -0,0 +1,111 @@
+#!/usr/bin/env python3
+"""Generate Vietnamese audio clips (1-90 + Chờ + Kinh) for every edge-tts
+Vietnamese voice. Output written to static/audio/{voiceId}/ and shipped
+with the app — runtime never calls TTS.
+
+Run once on a dev machine:
+
+    pip install edge-tts
+    python3 scripts/generate-audio.py
+"""
+import asyncio
+import json
+import os
+import re
+import sys
+
+OUT_ROOT = os.path.join(os.path.dirname(__file__), "..", "static", "audio")
+
+ONES = [
+    "không", "một", "hai", "ba", "bốn",
+    "năm", "sáu", "bảy", "tám", "chín",
+]
+
+
+def number_to_vietnamese(n: int) -> str:
+    """Mirror of src/lib/vietnamese-number.js for build-time use."""
+    if n < 10:
+        return ONES[n]
+    if n == 10:
+        return "mười"
+    if n < 20:
+        u = n - 10
+        return "mười lăm" if u == 5 else f"mười {ONES[u]}"
+    t, u = divmod(n, 10)
+    tens = f"{ONES[t]} mươi"
+    if u == 0:
+        return tens
+    if u == 1:
+        return f"{tens} mốt"
+    if u == 5:
+        return f"{tens} lăm"
+    return f"{tens} {ONES[u]}"
+
+
+def voice_id(short_name: str) -> str:
+    """vi-VN-HoaiMyNeural -> hoai-my"""
+    name = short_name.split("-")[-1]                 # HoaiMyNeural
+    name = re.sub(r"Neural$", "", name)              # HoaiMy
+    name = re.sub(r"(?<!^)(?=[A-Z])", "-", name)     # Hoai-My
+    return name.lower()                              # hoai-my
+
+
+def display_label(short_name: str, gender: str) -> str:
+    given = re.sub(r"Neural$", "", short_name.split("-")[-1])
+    given = re.sub(r"(?<!^)(?=[A-Z])", " ", given).strip()
+    gender_vi = "nữ" if gender.lower() == "female" else "nam"
+    return f"{given} ({gender_vi})"
+
+
+async def synth(text: str, voice: str, out: str) -> None:
+    import edge_tts
+    await edge_tts.Communicate(text, voice).save(out)
+    print(f"  {out}  ←  \"{text}\"")
+
+
+async def main() -> None:
+    import edge_tts
+    all_voices = await edge_tts.list_voices()
+    vi_voices = [v for v in all_voices if v["Locale"].startswith("vi-")]
+    if not vi_voices:
+        sys.exit("No Vietnamese voices found in edge-tts.")
+
+    seen_ids: set[str] = set()
+    manifest = {"voices": []}
+
+    for v in vi_voices:
+        vid = voice_id(v["ShortName"])
+        if vid in seen_ids:
+            sys.exit(f"Voice id collision: {vid} (from {v['ShortName']})")
+        seen_ids.add(vid)
+
+        out_dir = os.path.join(OUT_ROOT, vid)
+        os.makedirs(out_dir, exist_ok=True)
+        print(f"\n→ {v['ShortName']}  →  static/audio/{vid}/")
+
+        tasks = [
+            synth(number_to_vietnamese(n), v["ShortName"], os.path.join(out_dir, f"{n}.mp3"))
+            for n in range(1, 91)
+        ]
+        tasks.append(synth("Chờ",  v["ShortName"], os.path.join(out_dir, "cho.mp3")))
+        tasks.append(synth("Kinh", v["ShortName"], os.path.join(out_dir, "kinh.mp3")))
+        await asyncio.gather(*tasks)
+
+        manifest["voices"].append({
+            "id": vid,
+            "edgeName": v["ShortName"],
+            "label": display_label(v["ShortName"], v["Gender"]),
+            "gender": v["Gender"].lower(),
+        })
+
+    manifest_path = os.path.join(OUT_ROOT, "manifest.json")
+    with open(manifest_path, "w", encoding="utf-8") as f:
+        json.dump(manifest, f, ensure_ascii=False, indent=2)
+    print(f"\nWrote manifest with {len(manifest['voices'])} voice(s) → {manifest_path}")
+
+
+if __name__ == "__main__":
+    try:
+        asyncio.run(main())
+    except ImportError:
+        sys.exit("Install dep first: pip install edge-tts")
diff --git a/web/src/lib/MasterPanel.svelte b/web/src/lib/MasterPanel.svelte
index 4b4d736..6019292 100644
--- a/web/src/lib/MasterPanel.svelte
+++ b/web/src/lib/MasterPanel.svelte
@@ -64,6 +64,7 @@
 
 <script>
   import { settings } from "$lib/settings-store.svelte.js";
+  import { cancelPlayback, playNumber } from "$lib/voice.js";
 
   let state = $state(
     /** @type {{called: number[], remaining: number[]} | null} */ (null),
@@ -109,8 +110,13 @@
     return () => clearInterval(id);
   });
 
+  // Stop any in-flight clip when the component unmounts (e.g. master mode
+  // toggled off mid-clip) so audio doesn't outlive the panel.
+  $effect(() => () => cancelPlayback());
+
   function handleNewGame() {
     if (state && !confirm("Bạn có muốn tạo ván mới không?")) return;
+    cancelPlayback();
     autoRunning = false;
     state = createFreshState();
     lastCalled = null;
@@ -124,6 +130,7 @@
       remaining: state.remaining.slice(1),
     };
     lastCalled = next;
+    if (settings.voiceEnabledMaster) playNumber(next);
   }
 
   function toggleAuto() {
diff --git a/web/src/lib/PlayerBoard.svelte b/web/src/lib/PlayerBoard.svelte
index 042d545..428af2b 100644
--- a/web/src/lib/PlayerBoard.svelte
+++ b/web/src/lib/PlayerBoard.svelte
@@ -8,6 +8,8 @@
     saveCrossedState,
     saveGrid,
   } from "$lib/game-logic.js";
+  import { settings } from "$lib/settings-store.svelte.js";
+  import { cancelPlayback, playBingo, playWaiting } from "$lib/voice.js";
 
   /**
    * @typedef {Object} Props
@@ -90,6 +92,7 @@
         notifiedWaitingRows.add(i);
         congratsRow = i + 1;
         showCongrats = true;
+        if (settings.voiceEnabledPlayer) playBingo();
         break;
       }
     }
@@ -101,14 +104,19 @@
       if (waitNum !== null && !notifiedWaitingRows.has(i)) {
         notifiedWaitingRows.add(i);
         showToast(`Chờ ${waitNum}`);
+        if (settings.voiceEnabledPlayer) playWaiting(waitNum);
       } else if (waitNum === null && notifiedWaitingRows.has(i)) {
         notifiedWaitingRows.delete(i);
       }
     }
   });
 
+  // Stop any in-flight clip on unmount so audio doesn't outlive the board.
+  $effect(() => () => cancelPlayback());
+
   function handleGenerate() {
     if (grid && !confirm("Bạn có muốn tạo lại bảng không?")) return;
+    cancelPlayback();
     const newGrid = generateGrid();
     const newCrossed = newGrid.map((row) => row.map(() => false));
     grid = newGrid;
@@ -124,6 +132,7 @@
     if (!grid) return;
     const hasMarks = crossed.some((row) => row.some(Boolean));
     if (hasMarks && !confirm("Bạn có muốn xoá tất cả đánh dấu không?")) return;
+    cancelPlayback();
     crossed = grid.map((row) => row.map(() => false));
     celebratedRows.clear();
     notifiedWaitingRows.clear();
diff --git a/web/src/lib/SettingsButton.svelte b/web/src/lib/SettingsButton.svelte
index 45b7860..673e88d 100644
--- a/web/src/lib/SettingsButton.svelte
+++ b/web/src/lib/SettingsButton.svelte
@@ -1,4 +1,5 @@
 <script>
+  import { VOICES } from "$lib/audio-manifest.js";
   import {
     resetSettings,
     saveSettings,
@@ -66,6 +67,22 @@
     }
   }
 
+  function toggleVoiceMaster() {
+    settings.voiceEnabledMaster = !settings.voiceEnabledMaster;
+    saveSettings();
+  }
+
+  function toggleVoicePlayer() {
+    settings.voiceEnabledPlayer = !settings.voiceEnabledPlayer;
+    saveSettings();
+  }
+
+  /** @param {string} id */
+  function pickVoice(id) {
+    settings.voice = id;
+    saveSettings();
+  }
+
   function close() {
     open = false;
   }
@@ -232,6 +249,64 @@
       </fieldset>
       {/if}
 
+      <!-- Voice / sound -->
+      <fieldset class="mb-5">
+        <legend
+          class="text-sm font-semibold text-slate-700 dark:text-slate-200 mb-2"
+        >
+          Âm thanh
+        </legend>
+        <p class="text-xs text-slate-400 dark:text-slate-500 mb-2">
+          Đọc số bằng tiếng Việt
+        </p>
+
+        <button
+          type="button"
+          aria-pressed={settings.voiceEnabledMaster}
+          onclick={toggleVoiceMaster}
+          class="w-full px-3 py-2 rounded-lg border-2 text-sm font-medium transition-all mb-2
+                 {settings.voiceEnabledMaster
+                   ? 'border-emerald-500 bg-emerald-50 dark:bg-emerald-950/30 text-emerald-700 dark:text-emerald-300'
+                   : 'border-slate-300 dark:border-slate-600 text-slate-500 dark:text-slate-400'}"
+        >
+          Quản trò: {settings.voiceEnabledMaster ? "Bật" : "Tắt"}
+        </button>
+
+        <button
+          type="button"
+          aria-pressed={settings.voiceEnabledPlayer}
+          onclick={toggleVoicePlayer}
+          class="w-full px-3 py-2 rounded-lg border-2 text-sm font-medium transition-all
+                 {settings.voiceEnabledPlayer
+                   ? 'border-emerald-500 bg-emerald-50 dark:bg-emerald-950/30 text-emerald-700 dark:text-emerald-300'
+                   : 'border-slate-300 dark:border-slate-600 text-slate-500 dark:text-slate-400'}"
+        >
+          Người chơi (Chờ / Kinh): {settings.voiceEnabledPlayer ? "Bật" : "Tắt"}
+        </button>
+
+        {#if VOICES.length > 1}
+          <p class="text-xs text-slate-500 dark:text-slate-400 mt-3 mb-1">
+            Giọng đọc
+          </p>
+          <div class="grid grid-cols-2 gap-2">
+            {#each VOICES as v (v.id)}
+              {@const selected = settings.voice === v.id}
+              <button
+                type="button"
+                aria-pressed={selected}
+                onclick={() => pickVoice(v.id)}
+                class="px-3 py-2 rounded-lg border-2 text-sm transition-all
+                       {selected
+                         ? 'border-indigo-500 dark:border-indigo-400 bg-indigo-50 dark:bg-indigo-950/40 text-indigo-700 dark:text-indigo-300'
+                         : 'border-slate-200 dark:border-slate-600 text-slate-600 dark:text-slate-300 hover:border-slate-300 dark:hover:border-slate-500'}"
+              >
+                {v.label}
+              </button>
+            {/each}
+          </div>
+        {/if}
+      </fieldset>
+
       <!-- Empty cell color -->
       <fieldset class="mb-5">
         <legend
diff --git a/web/src/lib/audio-manifest.js b/web/src/lib/audio-manifest.js
new file mode 100644
index 0000000..7a270d1
--- /dev/null
+++ b/web/src/lib/audio-manifest.js
@@ -0,0 +1,23 @@
+/**
+ * Voice manifest produced by `scripts/generate-audio.py`. Imported as a
+ * static asset so Vite bakes it into the bundle — no runtime fetch.
+ * @module lib/audio-manifest
+ */
+import manifest from "../../static/audio/manifest.json";
+
+/**
+ * @typedef {Object} VoiceEntry
+ * @property {string} id          — folder slug (e.g. "hoai-my")
+ * @property {string} edgeName    — original edge-tts ShortName
+ * @property {string} label       — display label
+ * @property {string} gender
+ */
+
+/** @type {VoiceEntry[]} */
+export const VOICES = manifest.voices;
+
+/** Set of valid voice ids for fast membership checks. */
+export const VOICE_IDS = new Set(VOICES.map((v) => v.id));
+
+/** First entry is the default — the script writes them in edge-tts order. */
+export const DEFAULT_VOICE = VOICES[0]?.id ?? "hoai-my";
diff --git a/web/src/lib/settings-store.svelte.js b/web/src/lib/settings-store.svelte.js
index cad4d2a..2118551 100644
--- a/web/src/lib/settings-store.svelte.js
+++ b/web/src/lib/settings-store.svelte.js
@@ -6,6 +6,8 @@
  * @module lib/settings-store
  */
 
+import { DEFAULT_VOICE, VOICE_IDS } from "$lib/audio-manifest.js";
+
 const STORAGE_KEY = "loto_settings";
 const HEX6 = /^#[0-9a-fA-F]{6}$/;
 const VALID_THEMES = /** @type {const} */ (["auto", "light", "dark"]);
@@ -21,6 +23,12 @@ export const DEFAULT_SETTINGS = Object.freeze({
   autoCallEnabled: false,
   /** Auto-call interval, seconds per number. Integer 1..10. */
   autoCallSpeed: 5,
+  /** Speak the called number aloud when master draws. */
+  voiceEnabledMaster: true,
+  /** Speak "Chờ N" / "Kinh" on player events. */
+  voiceEnabledPlayer: true,
+  /** Active voice id; matches an entry in audio manifest. */
+  voice: DEFAULT_VOICE,
 });
 
 export const settings = $state({ ...DEFAULT_SETTINGS });
@@ -45,6 +53,10 @@ function validSpeed(v) {
     ? v
     : null;
 }
+/** @param {unknown} v */
+function validVoiceId(v) {
+  return typeof v === "string" && VOICE_IDS.has(v) ? v : null;
+}
 
 /* ------------------------------------------------------------------ */
 /* DOM apply                                                          */
@@ -111,6 +123,14 @@ export function loadSettings() {
         validBool(parsed.autoCallEnabled) ?? DEFAULT_SETTINGS.autoCallEnabled;
       settings.autoCallSpeed =
         validSpeed(parsed.autoCallSpeed) ?? DEFAULT_SETTINGS.autoCallSpeed;
+      settings.voiceEnabledMaster =
+        validBool(parsed.voiceEnabledMaster) ??
+        DEFAULT_SETTINGS.voiceEnabledMaster;
+      settings.voiceEnabledPlayer =
+        validBool(parsed.voiceEnabledPlayer) ??
+        DEFAULT_SETTINGS.voiceEnabledPlayer;
+      settings.voice =
+        validVoiceId(parsed.voice) ?? DEFAULT_SETTINGS.voice;
     }
   } catch {
     /* private mode / quota / corrupt JSON — fall back to defaults */
@@ -133,5 +153,8 @@ export function resetSettings() {
   settings.masterMode = DEFAULT_SETTINGS.masterMode;
   settings.autoCallEnabled = DEFAULT_SETTINGS.autoCallEnabled;
   settings.autoCallSpeed = DEFAULT_SETTINGS.autoCallSpeed;
+  settings.voiceEnabledMaster = DEFAULT_SETTINGS.voiceEnabledMaster;
+  settings.voiceEnabledPlayer = DEFAULT_SETTINGS.voiceEnabledPlayer;
+  settings.voice = DEFAULT_SETTINGS.voice;
   saveSettings();
 }
diff --git a/web/src/lib/settings-store.test.js b/web/src/lib/settings-store.test.js
index 7079eae..7deea76 100644
--- a/web/src/lib/settings-store.test.js
+++ b/web/src/lib/settings-store.test.js
@@ -48,6 +48,9 @@ beforeEach(() => {
     "masterMode",
     "autoCallEnabled",
     "autoCallSpeed",
+    "voiceEnabledMaster",
+    "voiceEnabledPlayer",
+    "voice",
   ])) {
     /** @type {any} */ (settings)[k] = /** @type {any} */ (DEFAULT_SETTINGS)[k];
   }
@@ -74,6 +77,56 @@ describe("settings-store — defaults", () => {
     expect(DEFAULT_SETTINGS.autoCallEnabled).toBe(false);
     expect(DEFAULT_SETTINGS.autoCallSpeed).toBe(5);
   });
+
+  it("voice toggles default ON; voice id picks first manifest entry", () => {
+    expect(DEFAULT_SETTINGS.voiceEnabledMaster).toBe(true);
+    expect(DEFAULT_SETTINGS.voiceEnabledPlayer).toBe(true);
+    expect(typeof DEFAULT_SETTINGS.voice).toBe("string");
+    expect(DEFAULT_SETTINGS.voice.length).toBeGreaterThan(0);
+  });
+
+  it("voice round-trips through localStorage and survives reload", () => {
+    settings.voice = "nam-minh";
+    settings.voiceEnabledMaster = false;
+    saveSettings();
+    settings.voice = "hoai-my";
+    settings.voiceEnabledMaster = true;
+    loadSettings();
+    expect(settings.voice).toBe("nam-minh");
+    expect(settings.voiceEnabledMaster).toBe(false);
+  });
+
+  it("invalid voice id falls back to default; other settings survive", () => {
+    localStorage.setItem(
+      KEY,
+      JSON.stringify({
+        ...DEFAULT_SETTINGS,
+        voice: "made-up-voice-id",
+        autoCallSpeed: 7,
+      }),
+    );
+    loadSettings();
+    expect(settings.voice).toBe(DEFAULT_SETTINGS.voice);
+    expect(settings.autoCallSpeed).toBe(7);
+  });
+
+  it("non-boolean voiceEnabled* falls back to default", () => {
+    localStorage.setItem(
+      KEY,
+      JSON.stringify({
+        ...DEFAULT_SETTINGS,
+        voiceEnabledMaster: "yes",
+        voiceEnabledPlayer: 1,
+      }),
+    );
+    loadSettings();
+    expect(settings.voiceEnabledMaster).toBe(
+      DEFAULT_SETTINGS.voiceEnabledMaster,
+    );
+    expect(settings.voiceEnabledPlayer).toBe(
+      DEFAULT_SETTINGS.voiceEnabledPlayer,
+    );
+  });
 });
 
 describe("settings-store — loadSettings (color)", () => {
@@ -188,6 +241,9 @@ describe("settings-store — saveSettings", () => {
     settings.masterMode = true;
     settings.autoCallEnabled = true;
     settings.autoCallSpeed = 7;
+    settings.voiceEnabledMaster = false;
+    settings.voiceEnabledPlayer = false;
+    settings.voice = "nam-minh";
     saveSettings();
     const raw = localStorage.getItem(KEY);
     expect(raw).not.toBeNull();
@@ -197,6 +253,9 @@ describe("settings-store — saveSettings", () => {
       masterMode: true,
       autoCallEnabled: true,
       autoCallSpeed: 7,
+      voiceEnabledMaster: false,
+      voiceEnabledPlayer: false,
+      voice: "nam-minh",
     });
   });
 
diff --git a/web/src/lib/vietnamese-number.js b/web/src/lib/vietnamese-number.js
new file mode 100644
index 0000000..1eb465f
--- /dev/null
+++ b/web/src/lib/vietnamese-number.js
@@ -0,0 +1,42 @@
+/**
+ * Vietnamese number-words for lô tô calls. Pure, no DOM, no I/O.
+ * @module lib/vietnamese-number
+ */
+
+const ONES = [
+  "không",
+  "một",
+  "hai",
+  "ba",
+  "bốn",
+  "năm",
+  "sáu",
+  "bảy",
+  "tám",
+  "chín",
+];
+
+/**
+ * Convert an integer 0..90 to its spoken Vietnamese form, honoring the
+ * tonal exceptions that show up in lô tô (15 = "mười lăm", 21 = "hai mươi
+ * mốt", 25 = "hai mươi lăm", etc.). Out-of-range values fall back to a
+ * plain `String(n)` so the caller never has to guard against undefined.
+ * @param {number} n
+ * @returns {string}
+ */
+export function numberToVietnamese(n) {
+  if (!Number.isInteger(n) || n < 0 || n > 90) return String(n);
+  if (n < 10) return ONES[n];
+  if (n === 10) return "mười";
+  if (n < 20) {
+    const u = n - 10;
+    return u === 5 ? "mười lăm" : `mười ${ONES[u]}`;
+  }
+  const t = Math.floor(n / 10);
+  const u = n % 10;
+  const tens = `${ONES[t]} mươi`;
+  if (u === 0) return tens;
+  if (u === 1) return `${tens} mốt`;
+  if (u === 5) return `${tens} lăm`;
+  return `${tens} ${ONES[u]}`;
+}
diff --git a/web/src/lib/vietnamese-number.test.js b/web/src/lib/vietnamese-number.test.js
new file mode 100644
index 0000000..f4df274
--- /dev/null
+++ b/web/src/lib/vietnamese-number.test.js
@@ -0,0 +1,70 @@
+import { describe, expect, it } from "vitest";
+import { numberToVietnamese } from "./vietnamese-number.js";
+
+describe("numberToVietnamese — units 0..9", () => {
+  it.each([
+    [0, "không"],
+    [1, "một"],
+    [2, "hai"],
+    [3, "ba"],
+    [4, "bốn"],
+    [5, "năm"],
+    [6, "sáu"],
+    [7, "bảy"],
+    [8, "tám"],
+    [9, "chín"],
+  ])("%i → %s", (n, expected) => {
+    expect(numberToVietnamese(n)).toBe(expected);
+  });
+});
+
+describe("numberToVietnamese — teens 10..19 (mười …, with mười lăm exception)", () => {
+  it.each([
+    [10, "mười"],
+    [11, "mười một"],
+    [12, "mười hai"],
+    [13, "mười ba"],
+    [14, "mười bốn"],
+    [15, "mười lăm"],
+    [16, "mười sáu"],
+    [17, "mười bảy"],
+    [18, "mười tám"],
+    [19, "mười chín"],
+  ])("%i → %s", (n, expected) => {
+    expect(numberToVietnamese(n)).toBe(expected);
+  });
+});
+
+describe("numberToVietnamese — 20..90 (mốt and lăm exceptions, ten-only forms)", () => {
+  it.each([
+    [20, "hai mươi"],
+    [21, "hai mươi mốt"],
+    [22, "hai mươi hai"],
+    [25, "hai mươi lăm"],
+    [29, "hai mươi chín"],
+    [30, "ba mươi"],
+    [31, "ba mươi mốt"],
+    [40, "bốn mươi"],
+    [45, "bốn mươi lăm"],
+    [55, "năm mươi lăm"],
+    [61, "sáu mươi mốt"],
+    [70, "bảy mươi"],
+    [81, "tám mươi mốt"],
+    [85, "tám mươi lăm"],
+    [90, "chín mươi"],
+  ])("%i → %s", (n, expected) => {
+    expect(numberToVietnamese(n)).toBe(expected);
+  });
+});
+
+describe("numberToVietnamese — out of range falls back to String(n)", () => {
+  it.each([
+    [-1, "-1"],
+    [91, "91"],
+    [100, "100"],
+    [0.5, "0.5"],
+    [Number.NaN, "NaN"],
+  ])("%p → %p", (n, expected) => {
+    expect(numberToVietnamese(/** @type {number} */ (n))).toBe(expected);
+  });
+});
diff --git a/web/src/lib/voice.js b/web/src/lib/voice.js
new file mode 100644
index 0000000..396491a
--- /dev/null
+++ b/web/src/lib/voice.js
@@ -0,0 +1,113 @@
+/**
+ * Bundled-MP3 playback for Vietnamese number calls. No runtime TTS, no
+ * API. Clips live under `static/audio/{voiceId}/{1..90,cho,kinh}.mp3`
+ * and are picked by `settings.voice`.
+ * @module lib/voice
+ */
+import { base } from "$app/paths";
+import { settings } from "$lib/settings-store.svelte.js";
+
+/** @type {Map<string, HTMLAudioElement>} */
+const cache = new Map();
+
+/** @type {HTMLAudioElement | null} */
+let activeClip = null;
+
+/** @type {symbol | null} */
+let activeToken = null;
+
+/** @type {(() => void) | null} — resolver of the in-flight playClip promise */
+let activeResolver = null;
+
+function isBrowser() {
+  return typeof window !== "undefined" && typeof Audio !== "undefined";
+}
+
+/** @param {string} url */
+function getAudio(url) {
+  let a = cache.get(url);
+  if (!a) {
+    a = new Audio(url);
+    a.preload = "auto";
+    cache.set(url, a);
+  }
+  return a;
+}
+
+/** @param {string} name — clip basename without extension */
+function clipUrl(name) {
+  return `${base}/audio/${settings.voice}/${name}.mp3`;
+}
+
+export function cancelPlayback() {
+  if (!isBrowser()) return;
+  if (activeClip) {
+    activeClip.onended = null;
+    activeClip.onerror = null;
+    activeClip.pause();
+    activeClip.currentTime = 0;
+  }
+  activeClip = null;
+  activeToken = null;
+  // Resolve any pending playClip promise so async chains exit cleanly
+  // (the next await will see the token mismatch and short-circuit).
+  if (activeResolver) {
+    const r = activeResolver;
+    activeResolver = null;
+    r();
+  }
+}
+
+/**
+ * Play one clip; resolve when it ends, errors, or is canceled.
+ * @param {string} url
+ * @param {symbol} token — caller's session marker; mismatch = canceled
+ */
+function playClip(url, token) {
+  return new Promise((resolve) => {
+    if (!isBrowser() || activeToken !== token) return resolve();
+    const a = getAudio(url);
+    a.currentTime = 0;
+    activeClip = a;
+    activeResolver = resolve;
+    const done = () => {
+      a.onended = null;
+      a.onerror = null;
+      if (activeClip === a) activeClip = null;
+      if (activeResolver === resolve) activeResolver = null;
+      resolve();
+    };
+    a.onended = done;
+    a.onerror = done;
+    a.play().catch(done);
+  });
+}
+
+/** @param {number} n */
+export function playNumber(n) {
+  if (!isBrowser()) return;
+  cancelPlayback();
+  const token = Symbol("playNumber");
+  activeToken = token;
+  void playClip(clipUrl(String(n)), token);
+}
+
+/** @param {number} n */
+export function playWaiting(n) {
+  if (!isBrowser()) return;
+  cancelPlayback();
+  const token = Symbol("playWaiting");
+  activeToken = token;
+  (async () => {
+    await playClip(clipUrl("cho"), token);
+    await playClip(clipUrl(String(n)), token);
+  })();
+}
+
+export function playBingo() {
+  if (!isBrowser()) return;
+  cancelPlayback();
+  const token = Symbol("playBingo");
+  activeToken = token;
+  void playClip(clipUrl("kinh"), token);
+}
diff --git a/web/static/audio/manifest.json b/web/static/audio/manifest.json
new file mode 100644
index 0000000..c4e94ee
--- /dev/null
+++ b/web/static/audio/manifest.json
@@ -0,0 +1,16 @@
+{
+  "voices": [
+    {
+      "id": "hoai-my",
+      "edgeName": "vi-VN-HoaiMyNeural",
+      "label": "Hoai My (nữ)",
+      "gender": "female"
+    },
+    {
+      "id": "nam-minh",
+      "edgeName": "vi-VN-NamMinhNeural",
+      "label": "Nam Minh (nam)",
+      "gender": "male"
+    }
+  ]
+}