From e4fef27aac2a99b7e04d69865ce645555927602b Mon Sep 17 00:00:00 2001
From: KooshaPari <koosha@example.com>
Date: Fri, 26 Jun 2026 05:00:03 -0700
Subject: [PATCH] fix(gemini): normalize audio input mime

---
 open-sse/translator/helpers/geminiHelper.ts    | 18 +++++++++---------
 .../unit/gemini-helper-audio-input-912.test.ts |  1 -
 2 files changed, 9 insertions(+), 10 deletions(-)

diff --git a/open-sse/translator/helpers/geminiHelper.ts b/open-sse/translator/helpers/geminiHelper.ts
index b5bf0780b8..6f054bcb04 100644
--- a/open-sse/translator/helpers/geminiHelper.ts
+++ b/open-sse/translator/helpers/geminiHelper.ts
@@ -95,6 +95,14 @@ export const DEFAULT_SAFETY_SETTINGS = [
   { category: "HARM_CATEGORY_CIVIC_INTEGRITY", threshold: "OFF" },
 ];
 
+function normalizeAudioMimeType(format: unknown): string {
+  const normalized = String(format || "wav").toLowerCase();
+  if (normalized === "mp3") {
+    return "audio/mpeg";
+  }
+  return `audio/${normalized}`;
+}
+
 // Convert OpenAI content to Gemini parts
 export function convertOpenAIContentToParts(content: unknown): JsonRecord[] {
   const parts: JsonRecord[] = [];
@@ -107,19 +115,11 @@ export function convertOpenAIContentToParts(content: unknown): JsonRecord[] {
       if (rec.type === "text") {
         parts.push({ text: rec.text });
       } else if (rec.type === "input_audio" || rec.type === "audio") {
-        // OpenAI Chat Completions audio input shape (ports decolua/9router#912 + #913):
-        // { type:"input_audio", input_audio:{data,format} } — some clients use the
-        // { type:"audio", audio:{data,format} } shape — -> Gemini
-        // `inlineData: { mimeType: "audio/<format>", data }`. mp3 normalizes to the
-        // canonical `audio/mpeg`; a leading `data:<mime>;base64,` prefix is stripped so
-        // Gemini receives raw base64.
         const audio = toRecord(rec.input_audio || rec.audio);
         if (typeof audio.data === "string" && audio.data) {
-          const format = typeof audio.format === "string" && audio.format ? audio.format : "wav";
-          const mimeType = format === "mp3" ? "audio/mpeg" : `audio/${format}`;
           parts.push({
             inlineData: {
-              mimeType,
+              mimeType: normalizeAudioMimeType(audio.format),
               data: audio.data.replace(/^data:[a-zA-Z0-9/+-]+;base64,/, ""),
             },
           });
diff --git a/tests/unit/gemini-helper-audio-input-912.test.ts b/tests/unit/gemini-helper-audio-input-912.test.ts
index 79473f3e48..6ca2960110 100644
--- a/tests/unit/gemini-helper-audio-input-912.test.ts
+++ b/tests/unit/gemini-helper-audio-input-912.test.ts
@@ -23,7 +23,6 @@ test("convertOpenAIContentToParts maps input_audio (mp3) and strips a data: pref
   const parts = gemini.convertOpenAIContentToParts([
     { type: "input_audio", input_audio: { data: "data:audio/mp3;base64,QUJDRA==", format: "mp3" } },
   ]);
-  // mp3 normalizes to the canonical audio/mpeg (matches the #913 translator test).
   assert.deepEqual(parts, [{ inlineData: { mimeType: "audio/mpeg", data: "QUJDRA==" } }]);
 });