fix(voice): resolve native crash and Whisper encoding failure during recording

- Win32AudioCaptureService: fix race condition between StopAsync resource cleanup and WaveInCallback on the multimedia thread. Move all pointer access inside lock and set _isRecording=false before waveInReset to prevent callbacks from touching freed GCHandles/buffers.
- WhisperSpeechToTextEngine: reject audio shorter than 1s (16k samples) to avoid WhisperProcessingException from the native encoder.

Made-with: Cursor

Author: Hongwei-MB

src/LiveLingo.Core/Speech/WhisperSpeechToTextEngine.cs

@@ -6,6 +6,9 @@ namespace LiveLingo.Core.Speech;
 
 public sealed class WhisperSpeechToTextEngine : ISpeechToTextEngine
 {
+    private const int MinSampleRate = 16000;
+    private const int MinDurationSamples = MinSampleRate; // 1 second at 16kHz
+
     private readonly IModelManager _modelManager;
     private readonly ILogger<WhisperSpeechToTextEngine>? _logger;
     private WhisperFactory? _factory;
@@ -32,6 +35,14 @@ public async Task<SpeechTranscriptionResult> TranscribeAsync(
         var processor = await GetOrLoadProcessorAsync(language, ct);
         var samples = ConvertPcmToFloat(audio);
 
+        if (samples.Length < MinDurationSamples)
+        {
+            _logger?.LogDebug(
+                "Audio too short for Whisper encoding ({Samples} samples, minimum {Min})",
+                samples.Length, MinDurationSamples);
+            return new SpeechTranscriptionResult(string.Empty, language ?? "en", 0f);
+        }
+
         var segments = new List<string>();
         string detectedLanguage = language ?? "en";
 

src/LiveLingo.Desktop/Platform/Windows/Win32AudioCaptureService.cs

@@ -18,7 +18,6 @@ internal sealed class Win32AudioCaptureService : IAudioCaptureService
     private readonly MemoryStream _capturedData = new();
     private readonly object _gate = new();
     private bool _isRecording;
-    private TaskCompletionSource? _stopTcs;
 
     public bool IsRecording
     {
@@ -94,10 +93,9 @@ public Task<AudioCaptureResult> StopAsync(CancellationToken ct = default)
         {
             if (!_isRecording)
                 throw new InvalidOperationException("Not recording.");
+            _isRecording = false;
         }
 
-        _stopTcs = new TaskCompletionSource();
-
         WaveIn.waveInStop(_hWaveIn);
         WaveIn.waveInReset(_hWaveIn);
 
@@ -116,13 +114,9 @@ public Task<AudioCaptureResult> StopAsync(CancellationToken ct = default)
         WaveIn.waveInClose(_hWaveIn);
         _hWaveIn = IntPtr.Zero;
 
-        lock (_gate) _isRecording = false;
-
         var pcm = _capturedData.ToArray();
         var duration = TimeSpan.FromSeconds((double)pcm.Length / (SampleRate * Channels * BitsPerSample / 8));
-        var result = new AudioCaptureResult(pcm, SampleRate, Channels, duration);
-
-        return Task.FromResult(result);
+        return Task.FromResult(new AudioCaptureResult(pcm, SampleRate, Channels, duration));
     }
 
     private readonly WaveIn.WaveInProc _waveInCallback = (hWaveIn, msg, _, headerPtr, _) =>
@@ -141,24 +135,21 @@ private void WaveInCallback(IntPtr hWaveIn, uint msg, IntPtr instance, IntPtr he
     {
         if (msg != WaveIn.WIM_DATA) return;
 
-        var header = Marshal.PtrToStructure<WaveIn.WAVEHDR>(headerPtr);
-        if (header.dwBytesRecorded > 0)
+        lock (_gate)
         {
-            var data = new byte[header.dwBytesRecorded];
-            Marshal.Copy(header.lpData, data, 0, (int)header.dwBytesRecorded);
-            lock (_gate)
+            if (!_isRecording) return;
+
+            var header = Marshal.PtrToStructure<WaveIn.WAVEHDR>(headerPtr);
+            if (header.dwBytesRecorded > 0)
             {
-                if (_isRecording && _capturedData.Length < MaxRecordingSeconds * SampleRate * Channels * BitsPerSample / 8)
+                var data = new byte[header.dwBytesRecorded];
+                Marshal.Copy(header.lpData, data, 0, (int)header.dwBytesRecorded);
+                if (_capturedData.Length < MaxRecordingSeconds * SampleRate * Channels * BitsPerSample / 8)
                     _capturedData.Write(data, 0, data.Length);
             }
-        }
 
-        lock (_gate)
-        {
-            if (_isRecording && _hWaveIn != IntPtr.Zero)
-            {
+            if (_hWaveIn != IntPtr.Zero)
                 WaveIn.waveInAddBuffer(_hWaveIn, headerPtr, (uint)Marshal.SizeOf<WaveIn.WAVEHDR>());
-            }
         }
     }