feat: upgrade local speech recognition engine to SenseVoice-Small with hardware audio enhancements

kouyu_english/lib/core/sherpa_stt_service.dart

@@ -15,7 +15,7 @@ class SherpaSttService {
 
   bool get isReady => _isInitialized && _recognizer != null;
 
-  /// Initializes Sherpa-ONNX bindings and unpacks bundled model assets to local disk if needed.
+  /// Initializes SenseVoice-Small ONNX bindings and unpacks bundled model assets to local disk if needed.
   Future<bool> initialize({String? nativeLibDir}) async {
     if (_isInitialized) return true;
     if (_isInitializing) return false;
@@ -29,32 +29,30 @@ class SherpaSttService {
       }
 
       final docDir = await getApplicationDocumentsDirectory();
-      final modelDir = Directory('${docDir.path}/sherpa_models');
+      final modelDir = Directory('${docDir.path}/sense_voice_models');
       if (!await modelDir.exists()) {
         await modelDir.create(recursive: true);
       }
 
       final modelFiles = [
-        'encoder-epoch-99-avg-1.int8.onnx',
-        'decoder-epoch-99-avg-1.int8.onnx',
-        'joiner-epoch-99-avg-1.int8.onnx',
+        'model.int8.onnx',
         'tokens.txt',
       ];
 
       for (final filename in modelFiles) {
         final targetFile = File('${modelDir.path}/$filename');
         if (!await targetFile.exists() || (await targetFile.length()) == 0) {
-          final ByteData data = await rootBundle.load('assets/models/sherpa/$filename');
+          final ByteData data = await rootBundle.load('assets/models/sense_voice/$filename');
           final Uint8List bytes = data.buffer.asUint8List(data.offsetInBytes, data.lengthInBytes);
           await targetFile.writeAsBytes(bytes, flush: true);
         }
       }
 
       final modelConfig = sherpa_onnx.OfflineModelConfig(
-        transducer: sherpa_onnx.OfflineTransducerModelConfig(
-          encoder: '${modelDir.path}/encoder-epoch-99-avg-1.int8.onnx',
-          decoder: '${modelDir.path}/decoder-epoch-99-avg-1.int8.onnx',
-          joiner: '${modelDir.path}/joiner-epoch-99-avg-1.int8.onnx',
+        senseVoice: sherpa_onnx.OfflineSenseVoiceModelConfig(
+          model: '${modelDir.path}/model.int8.onnx',
+          language: 'auto',
+          useInverseTextNormalization: true,
         ),
         tokens: '${modelDir.path}/tokens.txt',
         numThreads: 2,
@@ -69,17 +67,17 @@ class SherpaSttService {
       _recognizer = sherpa_onnx.OfflineRecognizer(recognizerConfig);
       _isInitialized = true;
       _isInitializing = false;
-      debugPrint('[SherpaSttService] Local ONNX ASR engine initialized successfully.');
+      debugPrint('[SherpaSttService] SenseVoice-Small ONNX ASR engine initialized successfully.');
       return true;
     } catch (e, stack) {
-      debugPrint('[SherpaSttService] Failed to initialize local ASR engine: $e\n$stack');
+      debugPrint('[SherpaSttService] Failed to initialize SenseVoice ASR engine: $e\n$stack');
       _isInitializing = false;
       _isInitialized = false;
       return false;
     }
   }
 
-  /// Transcribes a local 16kHz mono WAV audio file.
+  /// Transcribes a local 16kHz mono WAV audio file using SenseVoice-Small.
   Future<String?> transcribeWav(String wavPath) async {
     try {
       if (!_isInitialized) {
@@ -115,14 +113,19 @@ class SherpaSttService {
     }
   }
 
-  /// Cleans and formats raw recognized text into natural English casing.
+  /// Cleans and formats raw recognized SenseVoice text (strips emotion/event tags, normalizes casing).
   String _cleanText(String text) {
     if (text.isEmpty) return text;
-    // Lowercase first to normalize uppercase model output
-    final lower = text.toLowerCase().trim();
-    if (lower.isEmpty) return lower;
-    // Capitalize the first letter
-    return lower[0].toUpperCase() + lower.substring(1);
+    // Strip SenseVoice special tags like <|zh|>, <|en|>, <|NEUTRAL|>, <|HAPPY|>, <|Speech|>, <|withitn|>, <|woitn|>, etc.
+    var cleaned = text.replaceAll(RegExp(r'<\|[a-zA-Z0-9_\-\s]+\|>'), '').trim();
+    if (cleaned.isEmpty) return cleaned;
+    // Normalize consecutive spaces
+    cleaned = cleaned.replaceAll(RegExp(r'\s+'), ' ').trim();
+    // Capitalize first character if it's a letter
+    if (cleaned.isNotEmpty && cleaned[0].toLowerCase() != cleaned[0].toUpperCase()) {
+      cleaned = cleaned[0].toUpperCase() + cleaned.substring(1);
+    }
+    return cleaned;
   }
 
   void dispose() {

kouyu_english/lib/core/voice_service.dart

@@ -81,6 +81,9 @@ class VoiceService {
         encoder: AudioEncoder.wav,
         sampleRate: 16000,
         numChannels: 1,
+        noiseSuppress: true,
+        echoCancel: true,
+        autoGain: true,
       ),
       path: '${recordings.path}/practice_$timestamp.wav',
     );