increasing the speed of uploading to the server

backend/routers/stems.py

@@ -91,7 +91,7 @@ async def get_stem(
             detail=f"Stem '{stem_name}' not found"
         )
 
-    # Serve raw float32 PCM — skip WAV encode/decode entirely
+    # Serve raw int16 PCM — half the wire size of float32, no encode/decode libraries
     if format == "pcm":
         audio = stem.audio
         if audio.ndim == 2:
@@ -99,8 +99,9 @@ async def get_stem(
         else:
             audio = np.ascontiguousarray(audio, dtype=np.float32)
         num_frames = len(audio)
-        pcm_bytes = audio.tobytes()
-        print(f"[{stem_name}] Serving {len(pcm_bytes)/1024/1024:.1f}MB raw PCM ({num_frames} frames @ {stem.sample_rate}Hz)")
+        audio_int16 = (audio * 32767).astype(np.int16)
+        pcm_bytes = audio_int16.tobytes()
+        print(f"[{stem_name}] Serving {len(pcm_bytes)/1024/1024:.1f}MB int16 PCM ({num_frames} frames @ {stem.sample_rate}Hz)")
         return Response(
             content=pcm_bytes,
             media_type="application/octet-stream",
@@ -110,7 +111,8 @@ async def get_stem(
                 "X-Sample-Rate": str(stem.sample_rate),
                 "X-Channels": "1",
                 "X-Frames": str(num_frames),
-                "Access-Control-Expose-Headers": "X-Sample-Rate, X-Channels, X-Frames",
+                "X-Bit-Depth": "16",
+                "Access-Control-Expose-Headers": "X-Sample-Rate, X-Channels, X-Frames, X-Bit-Depth",
             }
         )
 

frontend/src/hooks/useAudioEngine.js

@@ -148,7 +148,11 @@ export function useAudioEngine() {
           console.log(`[${stem}] ArrayBuffer: ${sizeMB}MB in ${(bufferEnd - bufferStart).toFixed(0)}ms`)
 
           const constructStart = performance.now()
-          const float32 = new Float32Array(arrayBuffer)
+          const int16 = new Int16Array(arrayBuffer)
+          const float32 = new Float32Array(int16.length)
+          for (let i = 0; i < int16.length; i++) {
+            float32[i] = int16[i] / 32767
+          }
           const audioBuffer = ctx.createBuffer(numChannels, numFrames, sampleRate)
           audioBuffer.copyToChannel(float32, 0)
           const constructEnd = performance.now()