fix: pin numba for release CI wheel compatibility

fix: await for confirmation before deleting voices and channels

.bumpversion.cfg

@@ -1,5 +1,5 @@
 [bumpversion]
-current_version = 0.1.12
+current_version = 0.1.13
 commit = True
 tag = True
 tag_name = v{new_version}

.github/workflows/build-windows.yml

@@ -0,0 +1,63 @@
+name: Build Windows
+
+on:
+  workflow_dispatch:
+
+jobs:
+  build-windows:
+    permissions:
+      contents: write
+    runs-on: windows-latest
+
+    steps:
+      - uses: actions/checkout@v4
+
+      - name: Setup Python
+        uses: actions/setup-python@v5
+        with:
+          python-version: "3.12"
+          cache: "pip"
+
+      - name: Install Python dependencies
+        run: |
+          python -m pip install --upgrade pip
+          pip install pyinstaller
+          pip install -r backend/requirements.txt
+
+      - name: Build Python server
+        shell: bash
+        run: |
+          cd backend
+          python build_binary.py
+
+          PLATFORM=$(rustc --print host-tuple)
+          mkdir -p ../tauri/src-tauri/binaries
+          cp dist/voicebox-server.exe ../tauri/src-tauri/binaries/voicebox-server-${PLATFORM}.exe
+          echo "Built voicebox-server-${PLATFORM}.exe"
+
+      - name: Setup Bun
+        uses: oven-sh/setup-bun@v2
+
+      - name: Install Rust stable
+        uses: dtolnay/rust-toolchain@stable
+
+      - name: Rust cache
+        uses: swatinem/rust-cache@v2
+        with:
+          workspaces: "./tauri/src-tauri -> target"
+
+      - name: Install dependencies
+        run: bun install
+
+      - uses: tauri-apps/tauri-action@v0
+        env:
+          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+        with:
+          projectPath: tauri
+          tagName: v__VERSION__
+          releaseName: "voicebox v__VERSION__ (test build)"
+          releaseBody: "Test build for audio export fix"
+          releaseDraft: true
+          prerelease: true
+          args: ""
+          includeUpdaterJson: false

README.md

@@ -80,10 +80,10 @@ Voicebox is available now for macOS and Windows.
 
 | Platform | Download |
 |----------|----------|
-| macOS (Apple Silicon) | [voicebox_aarch64.app.tar.gz](https://github.com/jamiepine/voicebox/releases/download/v0.1.0/voicebox_aarch64.app.tar.gz) |
-| macOS (Intel) | [voicebox_x64.app.tar.gz](https://github.com/jamiepine/voicebox/releases/download/v0.1.0/voicebox_x64.app.tar.gz) |
-| Windows (MSI) | [voicebox_0.1.0_x64_en-US.msi](https://github.com/jamiepine/voicebox/releases/download/v0.1.0/voicebox_0.1.0_x64_en-US.msi) |
-| Windows (Setup) | [voicebox_0.1.0_x64-setup.exe](https://github.com/jamiepine/voicebox/releases/download/v0.1.0/voicebox_0.1.0_x64-setup.exe) |
+| macOS (Apple Silicon) | [Voicebox_aarch64.app.tar.gz](https://github.com/jamiepine/voicebox/releases/latest/download/Voicebox_aarch64.app.tar.gz) |
+| macOS (Intel) | [Voicebox_x64.app.tar.gz](https://github.com/jamiepine/voicebox/releases/latest/download/Voicebox_x64.app.tar.gz) |
+| Windows (MSI) | [Latest Windows MSI](https://github.com/jamiepine/voicebox/releases/latest) |
+| Windows (Setup) | [Latest Windows Setup](https://github.com/jamiepine/voicebox/releases/latest) |
 
 > **Linux builds coming soon** — Currently blocked by GitHub runner disk space limitations.
 
@@ -233,7 +233,7 @@ See [CONTRIBUTING.md](CONTRIBUTING.md) for detailed setup and contribution guide
 
 ```bash
 # Clone the repo
-git clone https://github.com/voicebox-sh/voicebox.git
+git clone https://github.com/jamiepine/voicebox.git
 cd voicebox
 
 # Setup everything
@@ -247,7 +247,7 @@ make dev
 
 ```bash
 # Clone the repo
-git clone https://github.com/voicebox-sh/voicebox.git
+git clone https://github.com/jamiepine/voicebox.git
 cd voicebox
 
 # Install dependencies
@@ -260,7 +260,7 @@ cd backend && pip install -r requirements.txt && cd ..
 bun run dev
 ```
 
-**Prerequisites:** [Bun](https://bun.sh), [Rust](https://rustup.rs), [Python 3.11+](https://python.org). 
+**Prerequisites:** [Bun](https://bun.sh), [Rust](https://rustup.rs), [Python 3.11+](https://python.org). [XCode on macOS](https://developer.apple.com/xcode/).
 
 **Performance:** 
 - **Apple Silicon (M1/M2/M3)**: Uses MLX backend with native Metal acceleration for 4-5x faster inference

app/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@voicebox/app",
-  "version": "0.1.12",
+  "version": "0.1.13",
   "private": true,
   "type": "module",
   "scripts": {

app/src/components/AudioTab/AudioTab.tsx

@@ -124,6 +124,13 @@ export function AudioTab() {
     );
   }
 
+  const handleChannelDelete = async (e, channelId) => {
+    e.stopPropagation();
+    if (await confirm('Delete this channel?')) {
+      deleteChannel.mutate(channelId);
+    }
+  }
+
   const allChannels = channels || [];
   const allDevices = devices || [];
   const selectedChannel = selectedChannelId
@@ -241,12 +248,7 @@ export function AudioTab() {
                             variant="ghost"
                             size="sm"
                             className="h-8 w-8 p-0"
-                            onClick={(e) => {
-                              e.stopPropagation();
-                              if (confirm('Delete this channel?')) {
-                                deleteChannel.mutate(channel.id);
-                              }
-                            }}
+                            onClick={(e) => handleChannelDelete(e, channel.id)}
                           >
                             <Trash2 className="h-4 w-4" />
                           </Button>

app/src/components/VoiceProfiles/AudioSampleRecording.tsx

@@ -58,6 +58,7 @@ export function AudioSampleRecording({
   // Request microphone access when component mounts
   useEffect(() => {
     if (!showWaveform) return;
+    if (!navigator.mediaDevices || !navigator.mediaDevices.getUserMedia) return;
 
     let stream: MediaStream | null = null;
 

app/src/components/VoiceProfiles/ProfileForm.tsx

@@ -43,7 +43,7 @@ import {
 } from '@/lib/hooks/useProfiles';
 import { useSystemAudioCapture } from '@/lib/hooks/useSystemAudioCapture';
 import { useTranscription } from '@/lib/hooks/useTranscription';
-import { formatAudioDuration, getAudioDuration } from '@/lib/utils/audio';
+import { convertToWav, formatAudioDuration, getAudioDuration } from '@/lib/utils/audio';
 import { usePlatform } from '@/platform/PlatformContext';
 import { useServerStore } from '@/stores/serverStore';
 import { type ProfileFormDraft, useUIStore } from '@/stores/uiStore';
@@ -505,10 +505,23 @@ export function ProfileForm() {
           language: data.language,
         });
 
+        // Convert non-WAV uploads to WAV so the backend can always use soundfile.
+        // Recorded audio is already WAV (from useAudioRecording's convertToWav call).
+        let fileToUpload: File = sampleFile;
+        if (!sampleFile.type.includes('wav') && !sampleFile.name.toLowerCase().endsWith('.wav')) {
+          try {
+            const wavBlob = await convertToWav(sampleFile);
+            const wavName = sampleFile.name.replace(/\.[^.]+$/, '.wav');
+            fileToUpload = new File([wavBlob], wavName, { type: 'audio/wav' });
+          } catch {
+            // If browser can't decode the format, send the original and let the backend try.
+          }
+        }
+
         try {
           await addSample.mutateAsync({
             profileId: profile.id,
-            file: sampleFile,
+            file: fileToUpload,
             referenceText: referenceText,
           });
 

app/src/components/VoicesTab/VoicesTab.tsx

@@ -79,8 +79,8 @@ export function VoicesTab() {
     setDialogOpen(true);
   };
 
-  const handleDelete = (profileId: string) => {
-    if (confirm('Are you sure you want to delete this profile?')) {
+  const handleProfileDelete = async (profileId: string) => {
+    if (await confirm('Are you sure you want to delete this profile?')) {
       deleteProfile.mutate(profileId);
     }
   };
@@ -147,7 +147,7 @@ export function VoicesTab() {
                 channels={channels || []}
                 onChannelChange={(channelIds) => handleChannelChange(profile.id, channelIds)}
                 onEdit={() => handleEdit(profile.id)}
-                onDelete={() => handleDelete(profile.id)}
+                onDelete={() => handleProfileDelete(profile.id)}
               />
             ))}
           </TableBody>

app/src/lib/hooks/useAudioRecording.ts

@@ -20,11 +20,13 @@ export function useAudioRecording({
   const streamRef = useRef<MediaStream | null>(null);
   const timerRef = useRef<number | null>(null);
   const startTimeRef = useRef<number | null>(null);
+  const cancelledRef = useRef<boolean>(false);
 
   const startRecording = useCallback(async () => {
     try {
       setError(null);
       chunksRef.current = [];
+      cancelledRef.current = false;
       setDuration(0);
 
       // Check if getUserMedia is available
@@ -87,31 +89,34 @@ export function useAudioRecording({
       };
 
       mediaRecorder.onstop = async () => {
+        // Snapshot the cancellation flag and recorded duration immediately —
+        // cancelRecording() clears chunks and sets cancelledRef synchronously
+        // before this async handler runs, so we must check it first.
+        const wasCancelled = cancelledRef.current;
+        const recordedDuration = startTimeRef.current
+          ? (Date.now() - startTimeRef.current) / 1000
+          : undefined;
+
         const webmBlob = new Blob(chunksRef.current, { type: 'audio/webm' });
 
-        // Convert to WAV format to avoid needing ffmpeg on backend
-        try {
-          const wavBlob = await convertToWav(webmBlob);
-
-          // Pass the actual recorded duration
-          const recordedDuration = startTimeRef.current
-            ? (Date.now() - startTimeRef.current) / 1000
-            : undefined;
-          onRecordingComplete?.(wavBlob, recordedDuration);
-        } catch (err) {
-          console.error('Error converting audio to WAV:', err);
-          // Fallback to original blob if conversion fails
-          const recordedDuration = startTimeRef.current
-            ? (Date.now() - startTimeRef.current) / 1000
-            : undefined;
-          onRecordingComplete?.(webmBlob, recordedDuration);
-        }
-
-        // Stop all tracks
+        // Stop all tracks now that we have the data
         streamRef.current?.getTracks().forEach((track) => {
           track.stop();
         });
         streamRef.current = null;
+
+        // Don't fire completion callback if the recording was cancelled
+        if (wasCancelled) return;
+
+        // Convert to WAV format to avoid needing ffmpeg on backend
+        try {
+          const wavBlob = await convertToWav(webmBlob);
+          onRecordingComplete?.(wavBlob, recordedDuration);
+        } catch (err) {
+          console.error('Error converting audio to WAV:', err);
+          // Fallback to original blob if conversion fails
+          onRecordingComplete?.(webmBlob, recordedDuration);
+        }
       };
 
       mediaRecorder.onerror = (event) => {
@@ -167,9 +172,10 @@ export function useAudioRecording({
 
   const cancelRecording = useCallback(() => {
     if (mediaRecorderRef.current) {
+      cancelledRef.current = true; // Must be set before stop() triggers onstop
+      chunksRef.current = [];
       mediaRecorderRef.current.stop();
       setIsRecording(false);
-      chunksRef.current = [];
       setDuration(0);
     }
 

app/src/lib/utils/audio.ts

@@ -22,6 +22,11 @@ export function formatAudioDuration(seconds: number): string {
  * If the file has a recordedDuration property (from recording hooks),
  * use that instead of trying to read metadata. This fixes issues on Windows
  * where WebM files from MediaRecorder don't have proper duration metadata.
+ *
+ * For uploaded files we use AudioContext.decodeAudioData which fully decodes
+ * the audio and returns the exact duration. This is more reliable than
+ * HTMLMediaElement.duration which can return incorrect large values for VBR
+ * MP3 files that lack a proper XING/VBRI header.
  */
 export async function getAudioDuration(
   file: File & { recordedDuration?: number },
@@ -30,26 +35,39 @@ export async function getAudioDuration(
     return file.recordedDuration;
   }
 
-  return new Promise((resolve, reject) => {
-    const audio = new Audio();
-    const url = URL.createObjectURL(file);
+  // Use Web Audio API for accurate duration — avoids VBR MP3 metadata issues.
+  try {
+    const audioContext = new AudioContext();
+    try {
+      const arrayBuffer = await file.arrayBuffer();
+      const audioBuffer = await audioContext.decodeAudioData(arrayBuffer);
+      return audioBuffer.duration;
+    } finally {
+      await audioContext.close();
+    }
+  } catch {
+    // Fallback: read duration from the media element (less accurate but works for WAV).
+    return new Promise((resolve, reject) => {
+      const audio = new Audio();
+      const url = URL.createObjectURL(file);
 
-    audio.addEventListener('loadedmetadata', () => {
-      URL.revokeObjectURL(url);
-      if (Number.isFinite(audio.duration) && audio.duration > 0) {
-        resolve(audio.duration);
-      } else {
-        reject(new Error('Audio file has invalid duration metadata'));
-      }
+      audio.addEventListener('loadedmetadata', () => {
+        URL.revokeObjectURL(url);
+        if (Number.isFinite(audio.duration) && audio.duration > 0) {
+          resolve(audio.duration);
+        } else {
+          reject(new Error('Audio file has invalid duration metadata'));
+        }
+      });
+
+      audio.addEventListener('error', () => {
+        URL.revokeObjectURL(url);
+        reject(new Error('Failed to load audio file'));
+      });
+
+      audio.src = url;
     });
-
-    audio.addEventListener('error', () => {
-      URL.revokeObjectURL(url);
-      reject(new Error('Failed to load audio file'));
-    });
-
-    audio.src = url;
-  });
+  }
 }
 
 /**

backend/__init__.py

@@ -1,3 +1,3 @@
 # Backend package
 
-__version__ = "0.1.12"
+__version__ = "0.1.13"

backend/backends/pytorch_backend.py

@@ -29,9 +29,23 @@ class PyTorchTTSBackend:
         """Get the best available device."""
         if torch.cuda.is_available():
             return "cuda"
-        elif hasattr(torch.backends, 'mps') and torch.backends.mps.is_available():
-            # MPS can have issues, use CPU for stability
-            return "cpu"
+        # Intel Arc / Intel Xe GPU via intel-extension-for-pytorch (IPEX)
+        try:
+            import intel_extension_for_pytorch  # noqa: F401
+            if hasattr(torch, 'xpu') and torch.xpu.is_available():
+                return "xpu"
+        except ImportError:
+            pass
+        # Any GPU on Windows via DirectML (torch-directml)
+        try:
+            import torch_directml
+            if torch_directml.device_count() > 0:
+                return torch_directml.device(0)
+        except ImportError:
+            pass
+        # MPS (Apple Silicon) — kept for completeness but MLX backend is preferred
+        if hasattr(torch.backends, 'mps') and torch.backends.mps.is_available():
+            return "cpu"  # MPS disabled for stability; MLX backend handles Apple Silicon
         return "cpu"
     
     def is_loaded(self) -> bool:
@@ -166,11 +180,21 @@ class PyTorchTTSBackend:
 
             # Load the model (tqdm is patched, but filters out non-download progress)
             try:
-                self.model = Qwen3TTSModel.from_pretrained(
-                    model_path,
-                    device_map=self.device,
-                    torch_dtype=torch.float32 if self.device == "cpu" else torch.bfloat16,
-                )
+                # Don't pass device_map on CPU: accelerate's meta-tensor mechanism
+                # causes "Cannot copy out of meta tensor" when moving to CPU.
+                # Instead load directly then call .to(device) if needed.
+                if self.device == "cpu":
+                    self.model = Qwen3TTSModel.from_pretrained(
+                        model_path,
+                        torch_dtype=torch.float32,
+                        low_cpu_mem_usage=False,
+                    )
+                else:
+                    self.model = Qwen3TTSModel.from_pretrained(
+                        model_path,
+                        device_map=self.device,
+                        torch_dtype=torch.bfloat16,
+                    )
             finally:
                 # Exit the patch context
                 tracker_context.__exit__(None, None, None)
@@ -358,9 +382,22 @@ class PyTorchSTTBackend:
         """Get the best available device."""
         if torch.cuda.is_available():
             return "cuda"
-        elif hasattr(torch.backends, 'mps') and torch.backends.mps.is_available():
-            # MPS support for Whisper
-            return "cpu"  # Use CPU for stability
+        # Intel Arc / Intel Xe GPU via intel-extension-for-pytorch (IPEX)
+        try:
+            import intel_extension_for_pytorch  # noqa: F401
+            if hasattr(torch, 'xpu') and torch.xpu.is_available():
+                return "xpu"
+        except ImportError:
+            pass
+        # Any GPU on Windows via DirectML (torch-directml)
+        try:
+            import torch_directml
+            if torch_directml.device_count() > 0:
+                return torch_directml.device(0)
+        except ImportError:
+            pass
+        if hasattr(torch.backends, 'mps') and torch.backends.mps.is_available():
+            return "cpu"  # MPS disabled for stability
         return "cpu"
     
     def is_loaded(self) -> bool:

backend/build_binary.py

@@ -83,9 +83,13 @@ def build_server():
             '--hidden-import', 'mlx_audio.stt',
             '--collect-submodules', 'mlx',
             '--collect-submodules', 'mlx_audio',
-            # Collect MLX data files including Metal shader libraries (.metallib)
-            '--collect-data', 'mlx',
-            '--collect-data', 'mlx_audio',
+            # Use --collect-all so PyInstaller bundles both data files AND
+            # native shared libraries (.dylib, .metallib) for MLX.
+            # Previously only --collect-data was used, which caused MLX to
+            # raise OSError at runtime inside the bundled binary because
+            # the Metal shader libraries were missing.
+            '--collect-all', 'mlx',
+            '--collect-all', 'mlx_audio',
         ])
     else:
         print("Building for non-Apple Silicon platform - PyTorch only")

backend/config.py

@@ -4,8 +4,17 @@ Configuration module for voicebox backend.
 Handles data directory configuration for production bundling.
 """
 
+import os
 from pathlib import Path
 
+# Allow users to override the HuggingFace model download directory.
+# Set VOICEBOX_MODELS_DIR to an absolute path before starting the server.
+# This sets HF_HUB_CACHE so all huggingface_hub downloads go to that path.
+_custom_models_dir = os.environ.get("VOICEBOX_MODELS_DIR")
+if _custom_models_dir:
+    os.environ["HF_HUB_CACHE"] = _custom_models_dir
+    print(f"[config] Model download path set to: {_custom_models_dir}")
+
 # Default data directory (used in development)
 _data_dir = Path("data")
 

backend/main.py

@@ -22,6 +22,24 @@ import uuid
 import asyncio
 import signal
 import os
+from urllib.parse import quote
+
+
+def _safe_content_disposition(disposition_type: str, filename: str) -> str:
+    """Build a Content-Disposition header that is safe for non-ASCII filenames.
+
+    Uses RFC 5987 ``filename*`` parameter so that browsers can decode
+    UTF-8 filenames while the ``filename`` fallback stays ASCII-only.
+    """
+    ascii_name = "".join(
+        c for c in filename if c.isascii() and (c.isalnum() or c in " -_.")
+    ).strip() or "download"
+    utf8_name = quote(filename, safe="")
+    return (
+        f'{disposition_type}; filename="{ascii_name}"; '
+        f"filename*=UTF-8''{utf8_name}"
+    )
+
 
 from . import database, models, profiles, history, tts, transcribe, config, export_import, channels, stories, __version__
 from .database import get_db, Generation as DBGeneration, VoiceProfile as DBVoiceProfile
@@ -77,10 +95,39 @@ async def health():
     tts_model = tts.get_tts_model()
     backend_type = get_backend_type()
 
-    # Check for GPU availability (CUDA or MPS)
+    # Check for GPU availability (CUDA, MPS, Intel Arc XPU, or DirectML)
     has_cuda = torch.cuda.is_available()
     has_mps = hasattr(torch.backends, 'mps') and torch.backends.mps.is_available()
-    gpu_available = has_cuda or has_mps
+
+    # Intel Arc / Intel Xe via intel-extension-for-pytorch (IPEX)
+    has_xpu = False
+    xpu_name = None
+    try:
+        import intel_extension_for_pytorch as ipex  # noqa: F401
+        if hasattr(torch, 'xpu') and torch.xpu.is_available():
+            has_xpu = True
+            try:
+                xpu_name = torch.xpu.get_device_name(0)
+            except Exception:
+                xpu_name = "Intel GPU"
+    except ImportError:
+        pass
+
+    # DirectML backend (torch-directml) for any Windows GPU
+    has_directml = False
+    directml_name = None
+    try:
+        import torch_directml
+        if torch_directml.device_count() > 0:
+            has_directml = True
+            try:
+                directml_name = torch_directml.device_name(0)
+            except Exception:
+                directml_name = "DirectML GPU"
+    except ImportError:
+        pass
+
+    gpu_available = has_cuda or has_mps or has_xpu or has_directml or backend_type == "mlx"
 
     gpu_type = None
     if has_cuda:
@@ -89,6 +136,10 @@ async def health():
         gpu_type = "MPS (Apple Silicon)"
     elif backend_type == "mlx":
         gpu_type = "Metal (Apple Silicon via MLX)"
+    elif has_xpu:
+        gpu_type = f"XPU ({xpu_name})"
+    elif has_directml:
+        gpu_type = f"DirectML ({directml_name})"
 
     vram_used = None
     if has_cuda:
@@ -252,12 +303,17 @@ async def add_profile_sample(
     db: Session = Depends(get_db),
 ):
     """Add a sample to a voice profile."""
-    # Save uploaded file to temporary location
-    with tempfile.NamedTemporaryFile(suffix=".wav", delete=False) as tmp:
+    # Preserve the uploaded file's extension so librosa can detect format correctly.
+    # Defaulting to .wav was causing soundfile to reject MP3/WebM content as invalid WAV.
+    _allowed_audio_exts = {'.wav', '.mp3', '.m4a', '.ogg', '.flac', '.aac', '.webm', '.opus'}
+    _uploaded_ext = Path(file.filename or '').suffix.lower()
+    file_suffix = _uploaded_ext if _uploaded_ext in _allowed_audio_exts else '.wav'
+
+    with tempfile.NamedTemporaryFile(suffix=file_suffix, delete=False) as tmp:
         content = await file.read()
         tmp.write(content)
         tmp_path = tmp.name
-    
+
     try:
         sample = await profiles.add_profile_sample(
             profile_id,
@@ -268,6 +324,8 @@ async def add_profile_sample(
         return sample
     except ValueError as e:
         raise HTTPException(status_code=400, detail=str(e))
+    except Exception as e:
+        raise HTTPException(status_code=500, detail=f"Failed to process audio file: {str(e)}")
     finally:
         # Clean up temp file
         Path(tmp_path).unlink(missing_ok=True)
@@ -388,7 +446,7 @@ async def export_profile(
             io.BytesIO(zip_bytes),
             media_type="application/zip",
             headers={
-                "Content-Disposition": f'attachment; filename="{filename}"'
+                "Content-Disposition": _safe_content_disposition("attachment", filename)
             }
         )
     except ValueError as e:
@@ -542,47 +600,50 @@ async def generate_speech(
         if not profile:
             raise HTTPException(status_code=404, detail="Profile not found")
         
-        # Create voice prompt from profile
-        voice_prompt = await profiles.create_voice_prompt_for_profile(
-            data.profile_id,
-            db,
-        )
-        
         # Generate audio
+
+        # Resolve model size and load the correct model FIRST.
+        # This must happen before create_voice_prompt_for_profile because that
+        # function calls load_model_async(None), which falls back to self.model_size.
+        # If the model is already loaded with the right size at that point, it
+        # returns immediately and the voice prompt is created by the correct model.
         tts_model = tts.get_tts_model()
-        # Load the requested model size if different from current (async to not block)
         model_size = data.model_size or "1.7B"
 
         # Check if model needs to be downloaded first
         model_path = tts_model._get_model_path(model_size)
-        if model_path.startswith("Qwen/"):
-            # Model not cached - check if it exists remotely or needs download
-            from huggingface_hub import constants as hf_constants
-            repo_cache = Path(hf_constants.HF_HUB_CACHE) / ("models--" + model_path.replace("/", "--"))
-            if not repo_cache.exists():
-                # Start download in background
-                model_name = f"qwen-tts-{model_size}"
+        if not tts_model._is_model_cached(model_size):
+            # Model is not fully cached — kick off a background download and tell
+            # the client to retry once it's ready.
+            model_name = f"qwen-tts-{model_size}"
 
-                async def download_model_background():
-                    try:
-                        await tts_model.load_model_async(model_size)
-                    except Exception as e:
-                        task_manager.error_download(model_name, str(e))
+            async def download_model_background():
+                try:
+                    await tts_model.load_model_async(model_size)
+                except Exception as e:
+                    task_manager.error_download(model_name, str(e))
 
-                task_manager.start_download(model_name)
-                asyncio.create_task(download_model_background())
+            task_manager.start_download(model_name)
+            asyncio.create_task(download_model_background())
 
-                # Return 202 Accepted with download info
-                raise HTTPException(
-                    status_code=202,
-                    detail={
-                        "message": f"Model {model_size} is being downloaded. Please wait and try again.",
-                        "model_name": model_name,
-                        "downloading": True
-                    }
-                )
+            raise HTTPException(
+                status_code=202,
+                detail={
+                    "message": f"Model {model_size} is being downloaded. Please wait and try again.",
+                    "model_name": model_name,
+                    "downloading": True,
+                },
+            )
 
+        # Load (or switch to) the requested model before building the voice prompt
         await tts_model.load_model_async(model_size)
+
+        # Create voice prompt from profile (model is already loaded with correct size)
+        voice_prompt = await profiles.create_voice_prompt_for_profile(
+            data.profile_id,
+            db,
+        )
+
         audio, sample_rate = await tts_model.generate(
             data.text,
             voice_prompt,
@@ -625,6 +686,59 @@ async def generate_speech(
         raise HTTPException(status_code=500, detail=str(e))
 
 
+@app.post("/generate/stream")
+async def stream_speech(
+    data: models.GenerationRequest,
+    db: Session = Depends(get_db),
+):
+    """
+    Generate speech and stream the WAV audio directly without saving to disk.
+
+    Returns raw WAV bytes via a StreamingResponse so the client can start
+    playing audio before the entire file has been received.  This endpoint
+    does NOT create a history entry — use /generate for that.
+    """
+    profile = await profiles.get_profile(data.profile_id, db)
+    if not profile:
+        raise HTTPException(status_code=404, detail="Profile not found")
+
+    tts_model = tts.get_tts_model()
+    model_size = data.model_size or "1.7B"
+
+    if not tts_model._is_model_cached(model_size):
+        raise HTTPException(
+            status_code=400,
+            detail=f"Model {model_size} is not downloaded yet. Use /generate to trigger a download.",
+        )
+
+    # Load the correct model before building the voice prompt (fixes issue #96)
+    await tts_model.load_model_async(model_size)
+
+    voice_prompt = await profiles.create_voice_prompt_for_profile(data.profile_id, db)
+
+    audio, sample_rate = await tts_model.generate(
+        data.text,
+        voice_prompt,
+        data.language,
+        data.seed,
+        data.instruct,
+    )
+
+    wav_bytes = tts.audio_to_wav_bytes(audio, sample_rate)
+
+    async def _wav_stream():
+        # Yield in chunks so large responses don't block the event loop
+        chunk_size = 64 * 1024  # 64 KB
+        for i in range(0, len(wav_bytes), chunk_size):
+            yield wav_bytes[i : i + chunk_size]
+
+    return StreamingResponse(
+        _wav_stream(),
+        media_type="audio/wav",
+        headers={"Content-Disposition": 'attachment; filename="speech.wav"'},
+    )
+
+
 # ============================================
 # HISTORY ENDPOINTS
 # ============================================
@@ -753,7 +867,7 @@ async def export_generation(
             io.BytesIO(zip_bytes),
             media_type="application/zip",
             headers={
-                "Content-Disposition": f'attachment; filename="{filename}"'
+                "Content-Disposition": _safe_content_disposition("attachment", filename)
             }
         )
     except ValueError as e:
@@ -786,7 +900,7 @@ async def export_generation_audio(
         audio_path,
         media_type="audio/wav",
         headers={
-            "Content-Disposition": f'attachment; filename="{filename}"'
+            "Content-Disposition": _safe_content_disposition("attachment", filename)
         }
     )
 
@@ -1054,7 +1168,7 @@ async def export_story_audio(
             io.BytesIO(audio_bytes),
             media_type="audio/wav",
             headers={
-                "Content-Disposition": f'attachment; filename="{filename}"'
+                "Content-Disposition": _safe_content_disposition("attachment", filename)
             }
         )
     except HTTPException:

backend/platform_detect.py

@@ -19,15 +19,17 @@ def is_apple_silicon() -> bool:
 def get_backend_type() -> Literal["mlx", "pytorch"]:
     """
     Detect the best backend for the current platform.
-    
+
     Returns:
-        "mlx" on Apple Silicon (if MLX is available), "pytorch" otherwise
+        "mlx" on Apple Silicon (if MLX is available and functional), "pytorch" otherwise
     """
     if is_apple_silicon():
         try:
-            import mlx
+            import mlx.core  # noqa: F401 — triggers native lib loading
             return "mlx"
-        except ImportError:
-            # MLX not installed, fallback to PyTorch
+        except (ImportError, OSError, RuntimeError):
+            # MLX not installed, or native libraries failed to load inside a
+            # PyInstaller bundle (OSError on missing .dylib / .metallib).
+            # Fall through to PyTorch.
             return "pytorch"
     return "pytorch"

backend/requirements.txt

@@ -18,6 +18,7 @@ qwen-tts>=0.0.5
 librosa>=0.10.0
 soundfile>=0.12.0
 numpy>=1.24.0
+numba>=0.60.0,<0.61.0
 
 # Utilities
 python-multipart>=0.0.6

backend/tts.py

@@ -32,11 +32,3 @@ def audio_to_wav_bytes(audio: np.ndarray, sample_rate: int) -> bytes:
     sf.write(buffer, audio, sample_rate, format="WAV")
     buffer.seek(0)
     return buffer.read()
-
-
-def audio_to_wav_bytes(audio: np.ndarray, sample_rate: int) -> bytes:
-    """Convert audio array to WAV bytes."""
-    buffer = io.BytesIO()
-    sf.write(buffer, audio, sample_rate, format="WAV")
-    buffer.seek(0)
-    return buffer.read()

backend/voicebox-server.spec

@@ -6,8 +6,13 @@ from PyInstaller.utils.hooks import copy_metadata
 datas = []
 hiddenimports = ['backend', 'backend.main', 'backend.config', 'backend.database', 'backend.models', 'backend.profiles', 'backend.history', 'backend.tts', 'backend.transcribe', 'backend.platform_detect', 'backend.backends', 'backend.backends.pytorch_backend', 'backend.utils.audio', 'backend.utils.cache', 'backend.utils.progress', 'backend.utils.hf_progress', 'backend.utils.validation', 'torch', 'transformers', 'fastapi', 'uvicorn', 'sqlalchemy', 'librosa', 'soundfile', 'qwen_tts', 'qwen_tts.inference', 'qwen_tts.inference.qwen3_tts_model', 'qwen_tts.inference.qwen3_tts_tokenizer', 'qwen_tts.core', 'qwen_tts.cli', 'pkg_resources.extern', 'backend.backends.mlx_backend', 'mlx', 'mlx.core', 'mlx.nn', 'mlx_audio', 'mlx_audio.tts', 'mlx_audio.stt']
 datas += collect_data_files('qwen_tts')
-datas += collect_data_files('mlx')
-datas += collect_data_files('mlx_audio')
+# Use collect_all (not collect_data_files) so native .dylib and .metallib
+# files are bundled as binaries, not data. Without this, MLX raises OSError
+# when loading Metal shaders inside the PyInstaller bundle.
+from PyInstaller.utils.hooks import collect_all as _collect_all
+_mlx_datas, _mlx_bins, _mlx_hidden = _collect_all('mlx')
+_mlxa_datas, _mlxa_bins, _mlxa_hidden = _collect_all('mlx_audio')
+datas += _mlx_datas + _mlxa_datas
 datas += copy_metadata('qwen-tts')
 hiddenimports += collect_submodules('qwen_tts')
 hiddenimports += collect_submodules('jaraco')
@@ -18,7 +23,7 @@ hiddenimports += collect_submodules('mlx_audio')
 a = Analysis(
     ['server.py'],
     pathex=[],
-    binaries=[],
+    binaries=_mlx_bins + _mlxa_bins,
     datas=datas,
     hiddenimports=hiddenimports,
     hookspath=[],

landing/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@voicebox/landing",
-  "version": "0.1.12",
+  "version": "0.1.13",
   "description": "Landing page for voicebox.sh",
   "scripts": {
     "dev": "bun --bun next dev --turbo",

landing/src/app/layout.tsx

@@ -1,6 +1,7 @@
 import type { Metadata } from 'next';
 import { Inter } from 'next/font/google';
 import './globals.css';
+import { Banner } from '@/components/Banner';
 import { Footer } from '@/components/Footer';
 import { Header } from '@/components/Header';
 
@@ -31,6 +32,7 @@ export default function RootLayout({ children }: { children: React.ReactNode })
     <html lang="en" suppressHydrationWarning className="dark">
       <body className={inter.variable}>
         <div className="relative min-h-screen bg-background font-sans flex flex-col">
+          <Banner />
           <Header />
           <main className="container mx-auto px-4 sm:px-6 md:px-4 flex-1 py-4 sm:py-6 md:py-0">
             {children}

landing/src/components/Banner.tsx

@@ -0,0 +1,25 @@
+import { ArrowRight } from 'lucide-react';
+
+export function Banner() {
+  return (
+    <div className="bg-primary/[0.06] border-b border-border backdrop-blur-sm">
+      <div className="container mx-auto px-4">
+        <div className="flex items-center justify-center h-10 text-sm">
+          <a
+            href="https://spacebot.sh"
+            target="_blank"
+            rel="noopener noreferrer"
+            className="flex items-center gap-2 text-muted-foreground hover:text-foreground transition-colors group"
+          >
+            <span>
+              Also by the creator of Voicebox:{' '}
+              <strong className="text-foreground/90">Spacebot</strong>, an AI agent OS for teams.
+              Connect Discord, Slack, or Telegram in one click.
+            </span>
+            <ArrowRight className="h-3.5 w-3.5 transition-transform group-hover:translate-x-0.5" />
+          </a>
+        </div>
+      </div>
+    </div>
+  );
+}

package.json

@@ -1,6 +1,6 @@
 {
   "name": "voicebox",
-  "version": "0.1.12",
+  "version": "0.1.13",
   "private": true,
   "workspaces": [
     "app",

requirements.txt

@@ -0,0 +1,9 @@
+uvicorn
+fastapi
+sqlalchemy
+torch
+torchvision
+soundfile
+librosa
+python-multipart
+huggingface_hub

tauri/package.json

@@ -1,7 +1,7 @@
 {
   "name": "@voicebox/tauri",
   "private": true,
-  "version": "0.1.12",
+  "version": "0.1.13",
   "type": "module",
   "scripts": {
     "dev": "vite",

tauri/src-tauri/Cargo.lock

@@ -5041,7 +5041,7 @@ checksum = "0b928f33d975fc6ad9f86c8f283853ad26bdd5b10b7f1542aa2fa15e2289105a"
 
 [[package]]
 name = "voicebox"
-version = "0.1.11"
+version = "0.1.12"
 dependencies = [
  "base64 0.22.1",
  "core-foundation-sys",

tauri/src-tauri/Cargo.toml

@@ -1,6 +1,6 @@
 [package]
 name = "voicebox"
-version = "0.1.12"
+version = "0.1.13"
 description = "A production-quality desktop app for Qwen3-TTS voice cloning and generation"
 authors = ["you"]
 license = ""

tauri/src-tauri/tauri.conf.json

@@ -1,7 +1,7 @@
 {
   "$schema": "https://schema.tauri.app/config/2",
   "productName": "Voicebox",
-  "version": "0.1.12",
+  "version": "0.1.13",
   "identifier": "sh.voicebox.app",
   "build": {
     "beforeDevCommand": "bun run dev",
@@ -12,7 +12,7 @@
   "bundle": {
     "active": true,
     "targets": "all",
-    "createUpdaterArtifacts": true,
+    "createUpdaterArtifacts": false,
     "externalBin": ["binaries/voicebox-server"],
     "icon": [
       "icons/32x32.png",

tauri/src/platform/filesystem.ts

@@ -2,40 +2,25 @@ import type { PlatformFilesystem, FileFilter } from '@/platform/types';
 
 export const tauriFilesystem: PlatformFilesystem = {
   async saveFile(filename: string, blob: Blob, filters?: FileFilter[]) {
-    try {
-      const { save } = await import('@tauri-apps/plugin-dialog');
-      const filePath = await save({
-        defaultPath: filename,
-        filters: filters || [],
-      });
+    const { save } = await import('@tauri-apps/plugin-dialog');
+    const { writeFile } = await import('@tauri-apps/plugin-fs');
 
-      if (filePath) {
-        let resolvedPath = '';
-        if (typeof filePath === 'string') {
-          resolvedPath = filePath;
-        } else if (filePath && typeof filePath === 'object' && 'path' in filePath) {
-          resolvedPath = (filePath as { path: string }).path;
-        }
+    const filePath = await save({
+      defaultPath: filename,
+      filters: filters || [],
+    });
 
-        if (!resolvedPath) {
-          throw new Error('Failed to resolve save path');
-        }
+    if (!filePath) return; // User cancelled the dialog
 
-        const { writeFile } = await import('@tauri-apps/plugin-fs');
-        const arrayBuffer = await blob.arrayBuffer();
-        await writeFile(resolvedPath, new Uint8Array(arrayBuffer));
-      }
-    } catch (error) {
-      console.error('Failed to use Tauri dialog, falling back to browser download:', error);
-      // Fall back to browser download if Tauri dialog fails
-      const url = window.URL.createObjectURL(blob);
-      const a = document.createElement('a');
-      a.href = url;
-      a.download = filename;
-      document.body.appendChild(a);
-      a.click();
-      window.URL.revokeObjectURL(url);
-      document.body.removeChild(a);
+    const resolvedPath = typeof filePath === 'string'
+      ? filePath
+      : (filePath as { path: string }).path;
+
+    if (!resolvedPath) {
+      throw new Error('Failed to resolve save path from dialog');
     }
+
+    const arrayBuffer = await blob.arrayBuffer();
+    await writeFile(resolvedPath, new Uint8Array(arrayBuffer));
   },
 };

web/package.json

@@ -1,7 +1,7 @@
 {
   "name": "@voicebox/web",
   "private": true,
-  "version": "0.1.12",
+  "version": "0.1.13",
   "type": "module",
   "scripts": {
     "dev": "vite",
@@ -21,6 +21,7 @@
     "@types/react-dom": "^18.3.0",
     "@typescript-eslint/eslint-plugin": "^7.0.0",
     "@typescript-eslint/parser": "^7.0.0",
+    "@tailwindcss/vite": "^4.0.0",
     "@vitejs/plugin-react": "^4.3.0",
     "eslint": "^8.57.0",
     "eslint-plugin-react-hooks": "^4.6.0",

web/vite.config.ts

@@ -1,9 +1,10 @@
 import path from 'node:path';
 import react from '@vitejs/plugin-react';
+import tailwindcss from '@tailwindcss/vite';
 import { defineConfig } from 'vite';
 
 export default defineConfig({
-  plugins: [react()],
+  plugins: [react(), tailwindcss()],
   resolve: {
     alias: {
       '@': path.resolve(__dirname, '../app/src'),