fix(tts): derive Azure SSML locale from selected voice (#1566)

* fix(tts): derive Azure SSML locale from voice

* fix(tts): preserve Azure voice script locales

---------

Co-authored-by: wyuc <wang-yc24@mails.tsinghua.edu.cn>

lib/audio/tts-providers.ts

@@ -790,12 +790,13 @@ async function generateAzureTTS(
   signal: AbortSignal,
 ): Promise<TTSGenerationResult> {
   const baseUrl = config.baseUrl || TTS_PROVIDERS['azure-tts'].defaultBaseUrl;
+  const voiceLocale = resolveAzureVoiceLocale(config.voice);
 
   // Build SSML
   const rate = config.speed ? `${((config.speed - 1) * 100).toFixed(0)}%` : '0%';
   const ssml = `
-    <speak version='1.0' xml:lang='zh-CN'>
-      <voice xml:lang='zh-CN' name='${config.voice}'>
+    <speak version='1.0' xml:lang='${voiceLocale}'>
+      <voice xml:lang='${voiceLocale}' name='${config.voice}'>
         <prosody rate='${rate}'>${escapeXml(text)}</prosody>
       </voice>
     </speak>
@@ -820,6 +821,22 @@ async function generateAzureTTS(
   return await validateTTSAudioResponse(response, 'Azure', 'mp3');
 }
 
+/** Resolve the BCP-47 locale encoded by an Azure voice identifier. */
+function resolveAzureVoiceLocale(voice: string): string {
+  const configuredVoice = TTS_PROVIDERS['azure-tts'].voices.find(({ id }) => id === voice);
+  if (configuredVoice?.language) return configuredVoice.language;
+
+  // Azure voice IDs conventionally start with a BCP-47 locale (for example,
+  // `en-US-JennyNeural` or `sr-Latn-RS-SophieNeural`). Preserve optional
+  // script and variant subtags for voices outside the small configured list
+  // while retaining the existing Chinese default for an unrecognised ID.
+  return (
+    voice.match(
+      /^[a-z]{2,3}(?:-[A-Z][a-z]{3})?-(?:[A-Z]{2}|\d{3})(?:-(?:[a-z0-9]{5,8}|\d[a-z0-9]{3}))*(?=-[A-Z]|$)/,
+    )?.[0] ?? 'zh-CN'
+  );
+}
+
 /**
  * GLM TTS implementation (GLM API)
  */

tests/audio/azure-tts.test.ts

@@ -0,0 +1,92 @@
+import { beforeEach, describe, expect, it, vi, type Mock } from 'vitest';
+import { generateTTS } from '@/lib/audio/tts-providers';
+
+const mockFetch = vi.hoisted(() => vi.fn() as Mock);
+
+vi.mock('undici', async (importOriginal) => {
+  const actual = await importOriginal<typeof import('undici')>();
+  return { ...actual, fetch: mockFetch };
+});
+
+function audioResponse() {
+  return {
+    ok: true,
+    status: 200,
+    headers: { get: () => 'audio/mpeg' },
+    arrayBuffer: async () => new Uint8Array([0xff, 0xfb, 0x90, 0x64]).buffer,
+  };
+}
+
+describe('Azure TTS SSML locale', () => {
+  beforeEach(() => {
+    mockFetch.mockReset();
+    mockFetch.mockResolvedValue(audioResponse());
+  });
+
+  it('uses the selected non-Chinese voice locale in both SSML language attributes', async () => {
+    await generateTTS(
+      {
+        providerId: 'azure-tts',
+        apiKey: 'azure-key',
+        baseUrl: 'https://eastus.tts.speech.microsoft.com',
+        voice: 'en-US-JennyNeural',
+      },
+      'Hello',
+    );
+
+    const ssml = mockFetch.mock.calls[0][1].body as string;
+    expect(ssml).toContain("<speak version='1.0' xml:lang='en-US'>");
+    expect(ssml).toContain("<voice xml:lang='en-US' name='en-US-JennyNeural'>");
+    expect(ssml).not.toContain("xml:lang='zh-CN'");
+  });
+
+  it('keeps zh-CN for the default Chinese voice', async () => {
+    await generateTTS(
+      {
+        providerId: 'azure-tts',
+        apiKey: 'azure-key',
+        baseUrl: 'https://eastus.tts.speech.microsoft.com',
+        voice: 'zh-CN-XiaoxiaoNeural',
+      },
+      '你好',
+    );
+
+    const ssml = mockFetch.mock.calls[0][1].body as string;
+    expect(ssml).toContain("<speak version='1.0' xml:lang='zh-CN'>");
+    expect(ssml).toContain("<voice xml:lang='zh-CN' name='zh-CN-XiaoxiaoNeural'>");
+  });
+
+  it('derives the locale from an otherwise unlisted Azure voice ID', async () => {
+    await generateTTS(
+      {
+        providerId: 'azure-tts',
+        apiKey: 'azure-key',
+        baseUrl: 'https://eastus.tts.speech.microsoft.com',
+        voice: 'fr-FR-DeniseNeural',
+      },
+      'Bonjour',
+    );
+
+    const ssml = mockFetch.mock.calls[0][1].body as string;
+    expect(ssml).toContain("xml:lang='fr-FR'");
+  });
+
+  it.each([
+    ['sr-Latn-RS-SophieNeural', 'sr-Latn-RS'],
+    ['iu-Cans-CA-SiqiniqNeural', 'iu-Cans-CA'],
+  ])('preserves the script subtag for %s', async (voice, locale) => {
+    await generateTTS(
+      {
+        providerId: 'azure-tts',
+        apiKey: 'azure-key',
+        baseUrl: 'https://eastus.tts.speech.microsoft.com',
+        voice,
+      },
+      'Hello',
+    );
+
+    const ssml = mockFetch.mock.calls[0][1].body as string;
+    expect(ssml).toContain(`<speak version='1.0' xml:lang='${locale}'>`);
+    expect(ssml).toContain(`<voice xml:lang='${locale}' name='${voice}'>`);
+  });
+});