fix: convert AAC audio with ffmpeg for transcription (#128)

OpenAI Whisper rejects raw AAC files even when renamed to .m4a - it
checks actual file format, not just extension. Signal voice messages
are often AAC.

Now uses ffmpeg to convert unsupported formats (aac, amr, caf, 3gp)
to MP3 before sending to Whisper API.

Requires ffmpeg installed on system.

Written by Cameron ◯ Letta Code

"When in doubt, transcode it out." - Audio engineering wisdom

src/transcription/openai.ts

@@ -4,6 +4,10 @@
 
 import OpenAI from 'openai';
 import { loadConfig } from '../config/index.js';
+import { execSync } from 'node:child_process';
+import { writeFileSync, readFileSync, unlinkSync, mkdirSync } from 'node:fs';
+import { join } from 'node:path';
+import { tmpdir } from 'node:os';
 
 let openaiClient: OpenAI | null = null;
 
@@ -34,15 +38,25 @@ function getModel(): string {
  */
 export async function transcribeAudio(audioBuffer: Buffer, filename: string = 'audio.ogg'): Promise<string> {
   const client = getClient();
+  const ext = filename.split('.').pop()?.toLowerCase() || '';
   
-  // Normalize filename for Whisper API (e.g., .aac -> .m4a)
-  const normalizedFilename = normalizeFilename(filename);
+  // Check if format needs conversion (not just renaming)
+  let finalBuffer = audioBuffer;
+  let finalFilename = filename;
+  
+  if (NEEDS_CONVERSION.includes(ext)) {
+    console.log(`[Transcription] Converting .${ext} to .mp3 with ffmpeg`);
+    const converted = convertAudioToMp3(audioBuffer, ext);
+    finalBuffer = converted;
+    finalFilename = filename.replace(/\.[^.]+$/, '.mp3');
+  } else {
+    // Just normalize the filename for formats that work with renaming
+    finalFilename = normalizeFilename(filename);
+  }
   
   // Create a File object from the buffer
-  // OpenAI SDK expects a File-like object
-  // Convert Buffer to Uint8Array to satisfy BlobPart type
-  const file = new File([new Uint8Array(audioBuffer)], normalizedFilename, { 
-    type: getMimeType(normalizedFilename) 
+  const file = new File([new Uint8Array(finalBuffer)], finalFilename, { 
+    type: getMimeType(finalFilename) 
   });
   
   const response = await client.audio.transcriptions.create({
@@ -53,6 +67,41 @@ export async function transcribeAudio(audioBuffer: Buffer, filename: string = 'a
   return response.text;
 }
 
+/**
+ * Formats that need actual conversion (not just renaming)
+ */
+const NEEDS_CONVERSION = ['aac', 'amr', 'caf', 'x-caf', '3gp', '3gpp'];
+
+/**
+ * Convert audio to MP3 using ffmpeg
+ */
+function convertAudioToMp3(audioBuffer: Buffer, inputExt: string): Buffer {
+  const tempDir = join(tmpdir(), 'lettabot-transcription');
+  mkdirSync(tempDir, { recursive: true });
+  
+  const inputPath = join(tempDir, `input-${Date.now()}.${inputExt}`);
+  const outputPath = join(tempDir, `output-${Date.now()}.mp3`);
+  
+  try {
+    // Write input file
+    writeFileSync(inputPath, audioBuffer);
+    
+    // Convert with ffmpeg
+    execSync(`ffmpeg -y -i "${inputPath}" -acodec libmp3lame -q:a 2 "${outputPath}" 2>/dev/null`, {
+      timeout: 30000,
+    });
+    
+    // Read output
+    const converted = readFileSync(outputPath);
+    console.log(`[Transcription] Converted ${audioBuffer.length} bytes → ${converted.length} bytes`);
+    return converted;
+  } finally {
+    // Cleanup temp files
+    try { unlinkSync(inputPath); } catch {}
+    try { unlinkSync(outputPath); } catch {}
+  }
+}
+
 /**
  * Supported formats for OpenAI Whisper API
  */