fix(tts): apply upstream timeout and validate provider/model form in Speech SDK path

Review-driven: thread ttsUpstreamTimeoutMs into the speech-sdk branch as an
overall budget across SDK retries, and fail fast on model ids missing the
provider/model form instead of falling through to provider defaults.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

docs-site/docs/configure/tts-provider-guides/speech-sdk.md

@@ -24,7 +24,7 @@ See [TTS Providers](../tts-providers) for admin-shared vs per-user behavior.
 
 ## Built-in models
 
-- `openai/gpt-4o-mini-tts`
+- `openai/gpt-4o-mini-tts` (works with your existing OpenAI API key)
 - `elevenlabs/eleven_multilingual_v2`
 - `cartesia/sonic-3.5`
 - `deepgram/aura-2`

src/lib/server/tts/generate.ts

@@ -696,10 +696,15 @@ const SPEECH_SDK_PROVIDER_FACTORIES: Record<string, SpeechSdkModelFactory> = {
 async function runSpeechSdkRequest(
   request: ResolvedServerTTSRequest,
   signal: AbortSignal,
-  maxRetries: number,
+  upstreamSettings: ResolvedTtsUpstreamRuntimeSettings,
 ): Promise<Buffer> {
   const model = request.model as string;
   const prefix = speechSdkProviderPrefix(model);
+  const modelId = model.slice(prefix.length + 1);
+  if (!prefix || !modelId) {
+    throw new Error(`Invalid Speech SDK model "${model}". Expected "provider/model".`);
+  }
+
   const factory = SPEECH_SDK_PROVIDER_FACTORIES[prefix];
   if (!factory) {
     throw new Error(
@@ -707,19 +712,30 @@ async function runSpeechSdkRequest(
     );
   }
 
-  const modelId = model.slice(prefix.length + 1);
   // 'default' is the placeholder voice for providers without a static voice
   // list; omit it so the provider's own default applies.
   const voice = request.voice === 'default' ? undefined : request.voice;
-  const result = await generateSpeech({
+
-    model: factory({ apiKey: request.apiKey || undefined })(modelId || undefined),
+  // Overall budget across the SDK's internal retries, mirroring the timeout
-    text: request.text,
+  // the OpenAI-compatible path configures on its client.
-    voice: voice as string,
+  const timeoutController = new AbortController();
-    output: { format: 'mp3' },
+  const timeoutId = setTimeout(() => timeoutController.abort(), upstreamSettings.ttsUpstreamTimeoutMs);
-    maxRetries,
+  const onAbort = () => timeoutController.abort();
-    abortSignal: signal,
+  signal.addEventListener('abort', onAbort, { once: true });
-  });
+  try {
-  return Buffer.from(result.audio.uint8Array);
+    const result = await generateSpeech({
+      model: factory({ apiKey: request.apiKey || undefined })(modelId),
+      text: request.text,
+      voice: voice as string,
+      output: { format: 'mp3' },
+      maxRetries: upstreamSettings.ttsUpstreamMaxRetries,
+      abortSignal: timeoutController.signal,
+    });
+    return Buffer.from(result.audio.uint8Array);
+  } finally {
+    clearTimeout(timeoutId);
+    signal.removeEventListener('abort', onAbort);
+  }
 }
 
 async function runProviderRequest(
@@ -733,7 +749,7 @@ async function runProviderRequest(
   const raw = request.provider === 'replicate'
     ? await runReplicateRequest(request, signal, upstreamSettings.ttsUpstreamMaxRetries)
     : request.provider === 'speech-sdk'
-      ? await runSpeechSdkRequest(request, signal, upstreamSettings.ttsUpstreamMaxRetries)
+      ? await runSpeechSdkRequest(request, signal, upstreamSettings)
       : await runOpenAiCompatibleRequest(request, signal, upstreamSettings);
 
   // OpenAI-compatible servers (and some Replicate models) may emit wav/ogg/etc.;

tests/unit/tts-generate.vitest.spec.ts

@@ -149,4 +149,17 @@ describe('speech-sdk request resolution', () => {
       apiKey: 'unused',
     })).rejects.toThrow('Unknown Speech SDK provider prefix');
   });
+
+  test('rejects models missing the provider/model form', async () => {
+    for (const model of ['openai', 'openai/', '/gpt-4o-mini-tts']) {
+      await expect(generateTTSBuffer({
+        text: 'hello',
+        voice: 'alloy',
+        speed: 1,
+        model,
+        provider: 'speech-sdk',
+        apiKey: 'unused',
+      })).rejects.toThrow('Expected "provider/model"');
+    }
+  });
 });