Add support for additional AM/PM formats (a.m./p.m., a.m/p.m, a m/p m)

Author: tanmaypawar-noise

examples/whisper.android/app/src/main/java/com/whispercppdemo/intent/SlotExtractor.kt

@@ -642,13 +642,17 @@ class SlotExtractor {
      * Examples:
      * - "730" -> "07:30"
      * - "730 pm" -> "19:30"
+     * - "730 p.m." -> "19:30"
+     * - "730 p m" -> "19:30"
      * - "1030 pm" -> "22:30"
      * - "2230" -> "22:30"
      * - "7:30 am" -> "07:30"
+     * - "7:30 a.m." -> "07:30"
      * - "5 30 pm" -> "17:30"
-     * - "5.30 am" -> "05:30"
+     * - "5.30 a.m" -> "05:30"
      * - "12:00 pm" -> "12:00"
      * - "12:00 am" -> "00:00"
+     * Supports AM/PM formats: am, pm, a.m., p.m., a.m, p.m, a m, p m
      */
     private fun normalizeTimeFormat(timeString: String): String {
         val cleanTime = timeString.trim().lowercase()
@@ -675,10 +679,12 @@ class SlotExtractor {
         }
 
         // Pattern for times like "730 pm", "1030 am" (digits + space + am/pm)
-        val amPmPattern = "^(\\d{1,4})\\s*(am|pm)$".toRegex()
+        // Supports: am, pm, a.m., p.m., a.m, p.m, a m, p m
+        val amPmPattern = "^(\\d{1,4})\\s*(?:(?:a\\.?\\s*m\\.?)|(?:p\\.?\\s*m\\.?))$".toRegex()
         amPmPattern.find(cleanTime)?.let { match ->
             val timeDigits = match.groupValues[1]
-            val amPm = match.groupValues[2]
+            val amPmText = match.value.substring(timeDigits.length).trim()
+            val amPm = if (amPmText.startsWith("a")) "am" else "pm"
 
             var hour = 0
             var minute = 0
@@ -711,11 +717,13 @@ class SlotExtractor {
         }
 
         // Pattern for "5 30 pm" or "5.30 am" format (hour [space|dot] minute am/pm)
-        val hourMinuteAmPmPattern = "^(\\d{1,2})[\\s.]+?(\\d{1,2})\\s*(am|pm)$".toRegex()
+        // Supports: am, pm, a.m., p.m., a.m, p.m, a m, p m
+        val hourMinuteAmPmPattern = "^(\\d{1,2})[\\s.]+?(\\d{1,2})\\s*(?:(?:a\\.?\\s*m\\.?)|(?:p\\.?\\s*m\\.?))$".toRegex()
         hourMinuteAmPmPattern.find(cleanTime)?.let { match ->
             var hour = match.groupValues[1].toInt()
             val minute = match.groupValues[2].toInt()
-            val amPm = match.groupValues[3]
+            val amPmText = match.value.substring(match.value.lastIndexOf(minute.toString()) + minute.toString().length).trim()
+            val amPm = if (amPmText.startsWith("a")) "am" else "pm"
 
             // Validate hour and minute ranges
             if (hour > 12 || minute >= 60) {
@@ -732,11 +740,13 @@ class SlotExtractor {
         }
 
         // Pattern for times with colon like "7:30 pm", "10:30 am"
-        val colonAmPmPattern = "^(\\d{1,2}):(\\d{2})\\s*(am|pm)$".toRegex()
+        // Supports: am, pm, a.m., p.m., a.m, p.m, a m, p m
+        val colonAmPmPattern = "^(\\d{1,2}):(\\d{2})\\s*(?:(?:a\\.?\\s*m\\.?)|(?:p\\.?\\s*m\\.?))$".toRegex()
         colonAmPmPattern.find(cleanTime)?.let { match ->
             var hour = match.groupValues[1].toInt()
             val minute = match.groupValues[2].toInt()
-            val amPm = match.groupValues[3]
+            val amPmText = match.value.substring(match.value.indexOf(":") + 3).trim()
+            val amPm = if (amPmText.startsWith("a")) "am" else "pm"
 
             // Convert to 24-hour format
             when {

examples/whisper.swiftui/Package.swift

@@ -0,0 +1,30 @@
+// swift-tools-version: 5.9
+import PackageDescription
+
+let package = Package(
+    name: "WhisperSwiftUI",
+    platforms: [
+        .iOS(.v15),
+        .macOS(.v12)
+    ],
+    products: [
+        .library(
+            name: "WhisperSwiftUI",
+            targets: ["WhisperSwiftUI"]),
+    ],
+    dependencies: [
+        .package(url: "https://github.com/huggingface/swift-transformers", from: "0.1.0"),
+        .package(url: "https://github.com/tensorflow/tensorflow", from: "2.19.0")
+    ],
+    targets: [
+        .target(
+            name: "WhisperSwiftUI",
+            dependencies: [
+                .product(name: "Transformers", package: "swift-transformers"),
+                .product(name: "TensorFlowLiteC", package: "tensorflow")
+            ]),
+        .testTarget(
+            name: "WhisperSwiftUITests",
+            dependencies: ["WhisperSwiftUI"]),
+    ]
+)

examples/whisper.swiftui/SETUP_SWIFT_TRANSFORMERS.md

@@ -0,0 +1,66 @@
+# Swift Transformers Setup Guide
+
+## Installation Instructions
+
+### Option 1: Swift Package Manager (Recommended)
+
+1. **Open Xcode project**: Open `whisper.swiftui.xcodeproj` in Xcode
+2. **Add Package Dependencies**:
+   - Go to `File` → `Add Package Dependencies...`
+   - Add the following URLs:
+     - `https://github.com/huggingface/swift-transformers`
+     - `https://github.com/tensorflow/swift-apis` (for TensorFlow support)
+
+3. **Configure Package Dependencies**:
+   - Select `swift-transformers` for the `Transformers` framework
+   - Select appropriate version (latest stable)
+
+### Option 2: Manual Package.swift Integration
+
+If using the Package.swift approach:
+
+```bash
+# From the whisper.swiftui directory
+swift package resolve
+swift package update
+```
+
+## Required Files Structure
+
+Ensure your bundle includes these files:
+
+```
+Resources/
+├── intent_classifier.tflite    # TensorFlow Lite model
+├── label_encoder.json         # Intent label mappings
+└── tokenizer/                 # BERT tokenizer files
+    ├── tokenizer.json        # Main tokenizer configuration
+    ├── tokenizer_config.json # Tokenizer metadata
+    └── vocab.txt             # Vocabulary file
+```
+
+## Usage
+
+The updated `IntentClassifier` now uses proper BERT tokenization:
+
+1. **Proper WordPiece tokenization** with special tokens ([CLS], [SEP])
+2. **Attention masking** for variable-length sequences
+3. **Automatic padding/truncation** to 256 tokens
+4. **Full compatibility** with Hugging Face BERT models
+
+## Troubleshooting
+
+If you encounter initialization errors:
+
+1. **Check bundle resources**: Verify all files are included in the Xcode project target
+2. **Verify tokenizer files**: Ensure the tokenizer directory contains all required files
+3. **Check dependencies**: Ensure Swift Transformers is properly linked
+4. **Review logs**: Check the console for detailed error messages
+
+## Key Improvements
+
+- ✅ **Proper BERT tokenization** using Swift Transformers
+- ✅ **WordPiece tokenization** with correct special tokens
+- ✅ **Attention masking** for better model performance
+- ✅ **Padding and truncation** handling
+- ✅ **Error handling** and debugging support

examples/whisper.swiftui/SWIFT_TRANSFORMERS_FIX.md

@@ -0,0 +1,63 @@
+# Swift Transformers Integration Fix
+
+## Issues Fixed:
+
+### 1. **Tokenizer Type Declaration**
+```swift
+// Before (incorrect):
+private let tokenizer: AutoTokenizer
+
+// After (correct):
+private var tokenizer: any Tokenizer
+```
+
+### 2. **AutoTokenizer Initialization**
+```swift
+// Before (incorrect):
+let encoded = try await tokenizer(text, maxLength: maxLength, ...)
+
+// After (correct):
+self.tokenizer = try await AutoTokenizer.from(pretrained: tokenizerPath)
+```
+
+### 3. **Tokenization Method**
+```swift
+// Before (incorrect API):
+let encoded = try await tokenizer(text, maxLength: maxLength, padding: .maxLength, ...)
+
+// After (correct swift-transformers API):
+let tokens = try tokenizer.encode(text: text)
+```
+
+### 4. **BERT Token Handling**
+- Added proper [CLS] token (101) at beginning
+- Added proper [SEP] token (102) at end  
+- Proper padding with [PAD] tokens (0)
+- Correct attention masking (1 for real tokens, 0 for padding)
+
+### 5. **Missing Variable Fix**
+- Restored `allProbabilities` mapping for IntentResult
+
+## Key Changes:
+
+1. **Tokenizer Property**: Changed to `any Tokenizer` type as per swift-transformers API
+2. **Encoding**: Uses `tokenizer.encode(text: text)` method directly
+3. **BERT Structure**: Ensures proper BERT token sequence: [CLS] + tokens + [SEP] + padding
+4. **Length Handling**: Truncates to maxLength-1 and adds [SEP] at end if needed
+
+## Expected Behavior:
+
+For input: "what is my heart rate"
+
+Should produce:
+- Input IDs: [101, 2054, 2003, 2026, 2540, 3954, 102, 0, 0, ...]
+- Attention: [1, 1, 1, 1, 1, 1, 1, 0, 0, ...]
+- Length: Exactly 256 tokens
+
+## Compatibility:
+
+This implementation now matches:
+- ✅ Swift Transformers API (tokenizer.encode)
+- ✅ BERT tokenization format ([CLS] + text + [SEP])
+- ✅ Python test code behavior (proper padding/truncation)
+- ✅ TensorFlow Lite model expectations

examples/whisper.swiftui/TEST_INTENT_CLASSIFIER.md

@@ -0,0 +1,73 @@
+#!/bin/bash
+# Test Script for Swift Intent Classifier
+
+echo "🧪 Testing Swift Intent Classifier Implementation"
+echo "================================================"
+
+echo ""
+echo "📋 Key Changes Made:"
+echo "✅ Updated tokenization to match Python: tokenizer(text, maxLength=256, padding='max_length', truncation=True)"
+echo "✅ Fixed label mapping to handle both label_to_intent and intent_to_label"
+echo "✅ Updated softmax to match Python: np.exp(logits) / np.sum(np.exp(logits))"
+echo "✅ Added proper logging to match Python test output format"
+echo "✅ Enhanced model loading logs with input/output shapes"
+
+echo ""
+echo "🔧 Expected Workflow:"
+echo "1. Load label_encoder.json (should show classes and mappings)"
+echo "2. Load AutoTokenizer from './tokenizer' directory"
+echo "3. Load intent_classifier.tflite model"
+echo "4. For each prediction:"
+echo "   - Tokenize text with max_length=256, padding='max_length'"
+echo "   - Run TFLite inference: input_ids + attention_mask → logits"
+echo "   - Apply softmax: probabilities = exp(logits) / sum(exp(logits))"
+echo "   - Get prediction: argmax(probabilities) and map to intent"
+
+echo ""
+echo "📁 Required Files Structure:"
+echo "Resources/"
+echo "├── intent_classifier.tflite"
+echo "├── label_encoder.json"
+echo "└── tokenizer/"
+echo "    ├── tokenizer.json"
+echo "    ├── tokenizer_config.json"
+echo "    └── vocab.txt"
+
+echo ""
+echo "🐛 Common Issues to Check:"
+echo "1. Bundle resource loading - verify files are included in Xcode target"
+echo "2. Swift Transformers import - ensure package is properly linked"
+echo "3. Tokenization parameters - must match Python exactly"
+echo "4. Input tensor shapes - verify [1, 256] for both input_ids and attention_mask"
+echo "5. Output tensor shape - should be [1, 13] for 13 intent classes"
+
+echo ""
+echo "✨ Test Commands in Xcode Console:"
+echo "po await intentClassifier.initialize()"
+echo "po await intentClassifier.classifyIntent(\"what is my heart rate\")"
+echo "po intentClassifier.getIntentList()"
+
+echo ""
+echo "🎯 Expected Output Format:"
+echo "✓ Label encoder loaded from JSON"
+echo "  Total classes: 13"
+echo "  Classes: QueryPoint, SetGoal, SetThreshold, ..."
+echo "✓ BERT tokenizer loaded from [path]"
+echo "  Max sequence length: 256"
+echo "✓ Complete TFLite model loaded successfully!"
+echo "  Input 0: Shape: [1, 256]"
+echo "  Input 1: Shape: [1, 256]"
+echo "  Output 0: Shape: [1, 13]"
+echo "This is a single end-to-end model: Text → Intent"
+
+echo ""
+echo "For text 'what is my heart rate':"
+echo "Tokenized 'what is my heart rate' -> X tokens (seq_len=256)"
+echo "Input IDs: [101, 2054, 2003, 2026, 2540, 3954, 102, ...]"
+echo "Attention: [1, 1, 1, 1, 1, 1, 1, ...]"
+echo "Model logits: [0.123, -1.456, 2.789, ...] (13 classes)"
+echo "Prediction: QueryPoint (label: 0)"
+echo "Confidence: 0.8945"
+
+echo ""
+echo "📊 If you see these logs, the implementation should work!"