fix: Improve Whisper transcription with WebM to WAV conversion

- Add ffmpeg conversion from WebM/Opus to WAV (16kHz mono PCM)
- Optimize transcription parameters (VAD, temperature, beam_size)
- Add Honduras Spanish context prompt with local expressions
- Fix chunk accumulation display in voice panel
- Add 1.5s recording buffer after releasing Ctrl+Space
- Skip small audio chunks (<5KB) that cause ffmpeg errors
- Use large-v3 model for better accuracy

frontend/src/components/FloatingVoice.vue

@@ -42,6 +42,7 @@ const connected = ref(false)
 let keyDownTime = 0
 let holdTimeout: number | null = null
 const isPushToTalk = ref(false)
+let pendingWhisperSend = false // Flag to send transcript when Whisper responds
 
 // ============ WHISPER MODE ============
 const useWhisper = ref(false)
@@ -194,19 +195,27 @@ function connectWhisperSocket() {
           const fullText = msg.text.trim()
 
           if (msg.partial) {
-            // For partial results, show as interim (will be replaced)
-            // Only show text that's new since last transcription
-            const newText = fullText.substring(lastTranscriptLength).trim()
-            if (newText) {
-              interimTranscript.value = newText
-              console.log(`[Voice] 🔄 WHISPER partial:`, newText)
-            }
-          } else {
-            // Final result - replace everything
+            // For partial results, show full accumulated transcription
+            transcript.value = fullText + ' '
+            interimTranscript.value = ''
+            console.log(`[Voice] 🔄 WHISPER partial:`, fullText)
+          } else {
+            // Final result
             transcript.value = fullText + ' '
             interimTranscript.value = ''
-            lastTranscriptLength = 0
             console.log(`[Voice] 🎯 WHISPER-GPU (${msg.model}/${msg.device}):`, fullText)
+
+            // Auto-send if push-to-talk was waiting for this
+            if (pendingWhisperSend) {
+              pendingWhisperSend = false
+              console.log('[Voice] Whisper response received, sending transcript')
+              if (transcript.value.trim()) {
+                sendTranscriptAndClose()
+              } else {
+                isPushToTalk.value = false
+                close()
+              }
+            }
           }
 
           // Update last transcript length for next partial
@@ -214,6 +223,11 @@ function connectWhisperSocket() {
         } else if (msg.error) {
           error.value = msg.error
           console.error('[Voice] Whisper error:', msg.error)
+          // Clear pending send on error
+          if (pendingWhisperSend) {
+            pendingWhisperSend = false
+            isPushToTalk.value = false
+          }
         }
       }
     } catch (e) {
@@ -263,7 +277,6 @@ async function startWhisperRecording() {
     // Start recording
     mediaRecorder.start(100) // Collect data every 100ms
     isRecording.value = true
-    interimTranscript.value = 'Listening (Whisper GPU)...'
 
     // Send chunks periodically for progressive transcription
     chunkInterval = window.setInterval(() => {
@@ -283,6 +296,16 @@ function sendAudioChunk(isFinal: boolean) {
 
   // Always send ALL accumulated audio (webm needs header from first chunk)
   const audioBlob = new Blob(audioChunks, { type: 'audio/webm' })
+  const chunkCount = audioChunks.length
+
+  // Skip if audio is too small (< 5KB) - WebM header alone is ~1-2KB
+  if (audioBlob.size < 5000) {
+    console.log(`[Voice] Skipping small chunk (${audioBlob.size} bytes)`)
+    if (isFinal) {
+      audioChunks = []
+    }
+    return
+  }
 
   // Clear chunks only if final
   if (isFinal) {
@@ -295,16 +318,13 @@ function sendAudioChunk(isFinal: boolean) {
     const base64 = (reader.result as string).split(',')[1]
 
     if (whisperSocket?.readyState === WebSocket.OPEN) {
-      if (!isFinal) {
-        interimTranscript.value = 'Processing...'
-      }
       whisperSocket.send(JSON.stringify({
         type: 'transcribe',
         audio: base64,
         language: 'es',
         partial: !isFinal
       }))
-      console.log(`[Voice] Sent ${isFinal ? 'FINAL' : 'partial'} audio (${audioChunks.length} chunks, ${audioBlob.size} bytes)`)
+      console.log(`[Voice] Sent ${isFinal ? 'FINAL' : 'partial'} audio (${chunkCount} chunks, ${audioBlob.size} bytes)`)
     }
   }
   reader.readAsDataURL(audioBlob)
@@ -524,21 +544,31 @@ function handleKeyUp(e: KeyboardEvent) {
       holdTimeout = null
     }
 
-    // If was push-to-talk recording, stop and send after 1200ms
+    // If was push-to-talk recording, continue recording for 1.5s buffer then stop
     if (isPushToTalk.value && isRecording.value) {
-      console.log('[Voice] Stopping recording, will send in 1200ms')
-      stopRecording()
+      console.log('[Voice] Key released, continuing recording for 1.5s buffer...')
+
+      // Keep recording for 1.5s more (UX buffer for trailing words)
       setTimeout(() => {
-        console.log('[Voice] Sending transcript:', transcript.value.trim())
-        console.log('[Voice] Socket state:', socket?.readyState)
-        if (transcript.value.trim()) {
-          sendTranscriptAndClose()
+        console.log('[Voice] Buffer complete, stopping recording')
+        stopRecording()
+
+        if (useWhisper.value) {
+          // For Whisper: wait for server response (handled in onmessage)
+          console.log('[Voice] Waiting for Whisper transcription...')
+          pendingWhisperSend = true
         } else {
-          // No transcript, just close
-          isPushToTalk.value = false
-          close()
+          // For Web Speech API: send after short delay for final results
+          setTimeout(() => {
+            if (transcript.value.trim()) {
+              sendTranscriptAndClose()
+            } else {
+              isPushToTalk.value = false
+              close()
+            }
+          }, 300)
         }
-      }, 1200)
+      }, 1500)
     }
 
     keyDownTime = 0