TTSをAzure SpeechからOpenAI gpt-4o-mini-ttsの女性声へ全面移行

.secrets.env.example

@@ -7,7 +7,6 @@
 #   ./scripts/put-secrets.sh --env production # prod（.secrets.prod.env を使うなら SECRETS_FILE=... を併用）
 
 SESSION_JWT_SECRET=
-AZURE_SPEECH_KEY=
 # GOOGLE_PLAY_SA_KEY はここに 1 行 JSON を入れてもよいが、複数行 JSON は
 # GOOGLE_PLAY_SA_KEY_FILE=./sa.json ./scripts/put-secrets.sh の方が扱いやすい。
 GOOGLE_PLAY_SA_KEY=

README.md

@@ -6,7 +6,7 @@ single Worker.
 
 ## Features
 
-- **TTS synthesis** (`POST /tts`): synthesizes SSML into audio via Azure Speech and caches it in KV/R2.
+- **TTS synthesis** (`POST /tts`): synthesizes plain text into audio via OpenAI `gpt-4o-mini-tts` and caches it in KV/R2.
 - **Session issuance** (`POST /auth/token`): issues a short-lived session JWT from an install ID (the replacement for Firebase anonymous auth).
 - **Feedback intake** (`POST /postFeedback`): enqueues feedback onto the triage queue.
 - **Image upload** (`POST /feedback/upload-image`): stores feedback images in R2 and returns a public URL.
@@ -22,7 +22,7 @@ single Worker.
 - **R2** — audio binaries and feedback images
 - **Cloudflare Queues** — `feedback-triage`
 - **Workers AI** — feedback triage
-- **Azure Speech** — TTS synthesis (SSML)
+- **OpenAI** — TTS synthesis (`gpt-4o-mini-tts`) and the conversational agent
 - **Google Android Publisher API** — Google Play review retrieval (service-account JWT)
 - **TypeScript / Biome / Jest / Wrangler**
 
@@ -60,8 +60,8 @@ wrangler queues create feedback-triage-dev
 
 ```bash
 wrangler secret put SESSION_JWT_SECRET          # signing key for session JWTs (any long random string)
-wrangler secret put AZURE_SPEECH_KEY            # Azure Speech subscription key
 wrangler secret put GOOGLE_PLAY_SA_KEY         # Android Publisher SA key JSON (single-line string)
+wrangler secret put OPENAI_API_KEY              # TTS synthesis and the conversational agent
 wrangler secret put OCTOKIT_PAT
 wrangler secret put DISCORD_CS_WEBHOOK_URL
 wrangler secret put DISCORD_CRASH_WEBHOOK_URL
@@ -76,7 +76,7 @@ You can also bulk-load secrets with the helper scripts: copy
 
 ### Non-secret configuration (vars)
 
-See `vars` in `wrangler.jsonc`. Configure the Azure region, voice names, AI model
+See `vars` in `wrangler.jsonc`. Configure the TTS model, voice names, AI model
 name, package name, public upload URL (the R2 public domain), and so on per
 environment.
 
@@ -103,10 +103,51 @@ format.
 
 A session JWT is obtained from `POST /auth/token` (body `{ "installId": "<uuid>" }`).
 
+### `POST /tts`
+
+Synthesis runs on OpenAI `gpt-4o-mini-tts`, which does **not** interpret SSML —
+the client sends plain text and steers delivery with `instructions`.
+
+```json
+{
+  "data": {
+    "textJa": "次は、オオサキです",
+    "textEn": "The next station is Osaki, J-Y 24.",
+    "model": "gpt-4o-mini-tts",
+    "jaVoiceName": "nova",
+    "enVoiceName": "nova",
+    "instructionsJa": "…",
+    "instructionsEn": "…"
+  }
+}
+```
+
+Every field is optional except that **at least one of `textJa` / `textEn` must
+be present**. Synthesis is billed per character, so the app omits a language the
+user has switched off; only the languages it asks for are synthesized, cached,
+and returned. `model` and the voice names are validated against an allowlist —
+anything unrecognized falls back to the KV config (`config:tts`) and then to the
+`TTS_*` vars, so a client cannot name an arbitrary (expensive) model.
+
+The response carries only the requested languages:
+
+```json
+{
+  "result": {
+    "id": "<sha256 of the request, used as the cache key>",
+    "jaAudioContent": "<base64>",
+    "jaAudioMimeType": "audio/mpeg",
+    "enAudioContent": "<base64>",
+    "enAudioMimeType": "audio/mpeg"
+  }
+}
+```
+
 ## Testing strategy
 
-Unit tests cover pure functions (SSML formatting, voice-name resolution, triage
-JSON normalization, review parsing) with Jest. Runtime integration for HTTP /
+Unit tests cover pure functions (TTS request building, voice/model resolution,
+text validation, cache writes, triage JSON normalization, review parsing) with
+Jest. Runtime integration for HTTP /
 queue / Cron is verified with `wrangler dev` / `wrangler dev --test-scheduled`.
 
 ## few-shot data
@@ -142,13 +183,13 @@ the KV namespace and R2 bucket from `wrangler.jsonc`).
 
 ### `find-tts-cache`
 
-Searches the TTS cache by SSML body and optionally deletes the matching KV
+Searches the TTS cache by spoken text and optionally deletes the matching KV
 document and R2 audio. KV is read via `wrangler kv key list` / `wrangler kv bulk
 get` and deleted via `wrangler kv key delete`; R2 audio is removed via `wrangler
 r2 object delete`.
 
 ```bash
-npm run find-tts-cache -- "東京" --field ssmlJa
+npm run find-tts-cache -- "東京" --field textJa
 npm run find-tts-cache -- "東京" --delete
 npm run find-tts-cache -- "東京" --env production --delete
 ```

scripts/build-secrets-json.mjs

@@ -8,7 +8,6 @@ import { readFileSync, writeFileSync } from 'node:fs';
 
 const SECRET_NAMES = [
   'SESSION_JWT_SECRET',
-  'AZURE_SPEECH_KEY',
   'GOOGLE_PLAY_SA_KEY',
   'APPSTORE_CONNECT_KEY',
   'OCTOKIT_PAT',

src/agent/prompt.test.ts

@@ -72,9 +72,7 @@ describe('buildSystemPrompt', () => {
     expect(prompt).toContain(
       'search_stations_by_name の結果は現在駅からの直通到達性しか保証しない'
     );
-    expect(prompt).toContain(
-      'それだけを根拠に駅を乗換地点として扱わない'
-    );
+    expect(prompt).toContain('それだけを根拠に駅を乗換地点として扱わない');
     expect(prompt).toContain(
       '最終目的地への接続を確認できない場合は suggestions を空配列'
     );

src/cli/find-tts-cache.ts

@@ -1,13 +1,13 @@
 /**
- * KV(TTS_KV) の voice:* メタを SSML 本文で検索し、必要なら KV ドキュメントと
+ * KV(TTS_KV) の voice:* メタを読み上げ本文で検索し、必要なら KV ドキュメントと
  * R2 上の音声ファイルを削除する。旧 Firestore+GCS 版の Cloudflare 移植。
  *
  * KV の一覧・値取得・削除、R2 の削除、バケット名解決はすべて wrangler
  * （wrangler.jsonc + ログイン済みアカウント）経由で行うため、API トークンや
  * ネームスペース ID、R2 認証情報を環境変数で渡す必要はない（要 `wrangler login`）。
  *
  * 例:
- *   npm run find-tts-cache -- "東京" --field ssmlJa
+ *   npm run find-tts-cache -- "東京" --field textJa
  *   npm run find-tts-cache -- "東京" --delete
  *   npm run find-tts-cache -- "東京" --env production --delete
  */
@@ -26,20 +26,20 @@ const R2_BINDING = 'TTS_BUCKET';
 
 interface CliArgs {
   searchTerm: string;
-  field?: 'ssmlJa' | 'ssmlEn';
+  field?: 'textJa' | 'textEn';
   exact: boolean;
   delete: boolean;
   env?: string;
 }
 
 function printUsage(): void {
   console.error(
-    'Usage: npm run find-tts-cache -- <search-term> [--field ssmlJa|ssmlEn] [--exact] [--delete] [--env <name>]'
+    'Usage: npm run find-tts-cache -- <search-term> [--field textJa|textEn] [--exact] [--delete] [--env <name>]'
   );
   console.error('');
   console.error('Options:');
   console.error(
-    '  --field <ssmlJa|ssmlEn>  検索対象フィールド（省略時は両方）'
+    '  --field <textJa|textEn>  検索対象フィールド（省略時は両方）'
   );
   console.error('  --exact                  部分一致ではなく完全一致で検索');
   console.error('  --delete                 KV ドキュメントと R2 音声を削除');
@@ -57,7 +57,7 @@ function parseArgs(argv: string[]): CliArgs | null {
   if (args.length === 0) return null;
 
   let searchTerm = '';
-  let field: 'ssmlJa' | 'ssmlEn' | undefined;
+  let field: 'textJa' | 'textEn' | undefined;
   let exact = false;
   let deleteMode = false;
   let env: string | undefined;
@@ -66,8 +66,8 @@ function parseArgs(argv: string[]): CliArgs | null {
     switch (args[i]) {
       case '--field': {
         const value = args[++i];
-        if (value !== 'ssmlJa' && value !== 'ssmlEn') {
-          console.error('Error: --field は "ssmlJa" か "ssmlEn" を指定');
+        if (value !== 'textJa' && value !== 'textEn') {
+          console.error('Error: --field は "textJa" か "textEn" を指定');
           process.exit(1);
         }
         field = value;
@@ -141,9 +141,14 @@ async function main(): Promise<void> {
     if (typeof rec.id !== 'string' || rec.id.length === 0) {
       continue;
     }
+    // ssmlJa/ssmlEn は Azure 時代のレコード。旧エントリも掃除できるよう併せて見る
     const hit = field
-      ? matchValue(rec[field])
-      : matchValue(rec.ssmlJa) || matchValue(rec.ssmlEn);
+      ? matchValue(rec[field]) ||
+        matchValue(field === 'textJa' ? rec.ssmlJa : rec.ssmlEn)
+      : matchValue(rec.textJa) ||
+        matchValue(rec.textEn) ||
+        matchValue(rec.ssmlJa) ||
+        matchValue(rec.ssmlEn);
     if (hit) matches.push(rec);
   }
 
@@ -155,8 +160,9 @@ async function main(): Promise<void> {
   console.log(`${matches.length}件のドキュメントが見つかりました:\n`);
   for (const rec of matches) {
     console.log(`ID:         ${rec.id}`);
-    console.log(`SSML (JA):  ${rec.ssmlJa ?? ''}`);
-    console.log(`SSML (EN):  ${rec.ssmlEn ?? ''}`);
+    console.log(`Text (JA):  ${rec.textJa ?? rec.ssmlJa ?? ''}`);
+    console.log(`Text (EN):  ${rec.textEn ?? rec.ssmlEn ?? ''}`);
+    console.log(`Model:      ${rec.model ?? ''}`);
     console.log(`Path (JA):  ${rec.pathJa ?? ''}`);
     console.log(`Path (EN):  ${rec.pathEn ?? ''}`);
     console.log(`Voice (JA): ${rec.voiceJa ?? ''}`);

src/cli/lib/wrangler.ts

@@ -219,6 +219,10 @@ export function confirm(prompt: string): Promise<boolean> {
 // --- 共有: voice メタの型 ---
 export interface VoiceCacheRecord {
   id: string;
+  model?: string;
+  textJa?: string;
+  textEn?: string;
+  /** Azure/SSML 時代のレコード。旧エントリを検索・削除できるよう残している */
   ssmlJa?: string;
   ssmlEn?: string;
   pathJa?: string;