fix(ai): preserve Gemini 3.8 TTS WAV output (#51300)

Co-authored-by: nexxeln <95541290+nexxeln@users.noreply.github.com>

packages/ai/src/protocols/google-speech.ts

@@ -56,10 +56,15 @@ interface State extends SpeechStream.Audio, GeminiGenerateContent.Metadata {
 // ---------------------------------------------------------------------------
 
 const fromRequest = Effect.fn("GoogleSpeech.fromRequest")(function* (request: MediaProtocol.Addressed<Request>) {
+  if (request.format === "pcm" && request.mode === "generate" && /^gemini-3\.8-.*-tts(?:-|$)/.test(request.model.id))
+    return yield* route.unsupported(
+      "media.format",
+      `${route.name} returns WAV by default for Gemini 3.8 TTS unary requests; omit the format to accept it`,
+    )
   if (request.format !== undefined && request.format !== "pcm")
     return yield* route.unsupported(
       "media.format",
-      `${route.name} only returns raw PCM; request format "pcm" or omit it, then wrap the samples yourself`,
+      `${route.name} only accepts raw PCM as an explicit format; omit it to accept the provider's default output`,
     )
   const voiceName = SpeechStream.voiceID(request.voice)
   return MediaProtocol.json(
@@ -97,10 +102,18 @@ const step = Effect.fn("GoogleSpeech.step")(function* (state: State, frame: stri
   return [next, audio.flatMap((part) => SpeechStream.delta(next, part.data)[1])] as const
 })
 
-const finish = (state: State) => {
+const finish = (state: State, context: MediaProtocol.ResponseContext<Request>) => {
   const sampleRate = SpeechStream.sampleRate(state.mimeType) ?? DEFAULT_SAMPLE_RATE
+  const output =
+    state.mimeType?.split(";")[0]?.toLowerCase() === "audio/wav"
+      ? SpeechStream.container("wav", sampleRate)
+      : SpeechStream.pcm("pcm_s16le", sampleRate, state.mimeType ?? `audio/L16;codec=pcm;rate=${sampleRate}`)
+  if (context.request.format === "pcm" && output.info.format !== "pcm")
+    return Effect.fail(
+      route.frameError(`Google Speech returned ${output.info.format} instead of the requested raw PCM`),
+    )
   return SpeechStream.finish(route, state, {
-    ...SpeechStream.pcm("pcm_s16le", sampleRate, state.mimeType ?? `audio/L16;codec=pcm;rate=${sampleRate}`),
+    ...output,
     usage: GeminiGenerateContent.usage(state.usage),
     providerMetadata: GeminiGenerateContent.providerMetadata(state),
     detail: state.finishReason === undefined ? undefined : `finish reason: ${state.finishReason}`,

packages/ai/test/speech.test.ts

@@ -28,10 +28,49 @@ const cartesia = Cartesia.configure({ apiKey: "test", baseURL: "https://cartesia
 const google = Google.configure({ apiKey: "test", baseURL: "https://google.test/v1beta" }).speech(
   "gemini-2.5-flash-preview-tts",
 )
+const google38 = Google.configure({ apiKey: "test", baseURL: "https://google.test/v1beta" }).speech(
+  "gemini-3.8-flash-tts",
+)
+const google38Lite = Google.configure({ apiKey: "test", baseURL: "https://google.test/v1beta" }).speech(
+  "gemini-3.8-flash-lite-tts",
+)
 const deepgram = Deepgram.configure({ apiKey: "test", baseURL: "https://deepgram.test" }).speech("aura-2-thalia-en")
 const voice = "JBFqnCBsd6RMkjVDRZzb"
 
 describe("Speech", () => {
+  it.effect("preserves Google's WAV output instead of describing it as raw PCM", () =>
+    Effect.gen(function* () {
+      const bytes = new TextEncoder().encode("RIFF....WAVEfmt ")
+      const response = yield* Speech.generate({ model: google38, text: "Hi" }).pipe(
+        Effect.provide(
+          respond(
+            JSON.stringify({
+              candidates: [
+                { content: { parts: [{ inlineData: { mimeType: "audio/wav", data: Encoding.encodeBase64(bytes) } }] } },
+              ],
+            }),
+            "application/json",
+          ),
+        ),
+      )
+      expect(response.audio.mediaType).toBe("audio/wav")
+      expect(response.audio.info?.format).toBe("wav")
+      expect(response.audio.info?.encoding).toBeUndefined()
+      expect(yield* response.audio.bytes()).toEqual(bytes)
+    }),
+  )
+
+  it.effect("rejects raw PCM for Gemini 3.8 unary requests before sending", () =>
+    Effect.gen(function* () {
+      const errors = yield* Effect.all(
+        [google38, google38Lite].map((model) =>
+          Speech.generate({ model, text: "Hi", format: "pcm" }).pipe(Effect.flip),
+        ),
+      ).pipe(Effect.provide(layer(() => Effect.die("An unsupported request reached the network"))))
+      expect(errors.map((error) => error.reason._tag)).toEqual(["UnsupportedOperation", "UnsupportedOperation"])
+    }),
+  )
+
   it.effect("rejects what a provider cannot produce before sending anything", () =>
     Effect.gen(function* () {
       const errors = yield* Effect.all(