@gmessier/nitro-speech 0.1.2 → 0.2.0

package/README.md

@@ -16,11 +16,13 @@ React Native Real-Time Speech Recognition Library, powered by [Nitro Modules](ht
 #### Key Features:
 
 - Built on Nitro Modules for low-overhead native bridging
+- Uses newest advanced Speech-to-Text API for iOS 26+ (with fallback to legacy SpeechRecognition for older versions)
 - Configurable Timer for silence (default: 8 sec)
   - Callback `onAutoFinishProgress` for progress bars, etc...
   - Method `addAutoFinishTime` for single timer update
   - Method `updateAutoFinishTime` for constant timer update
-- Optional Haptic Feedback on start and finish
+- Configurable Haptic Feedback on start and finish
+- Flexible `useVoiceInputVolume` hook to display user input volume in UI animations
 - Speech-quality configurations:
   - Result is grouped by speech segments into Batches.
   - Param `disableRepeatingFilter` for consecutive duplicate-word filtering.
@@ -38,6 +40,7 @@ React Native Real-Time Speech Recognition Library, powered by [Nitro Modules](ht
   - [Recommended: useRecognizer Hook](#recommended-userecognizer-hook)
   - [With React Navigation (important)](#with-react-navigation-important)
   - [Cross-component control: RecognizerRef](#cross-component-control-recognizerref)
+  - [Voice input volume](#voice-input-volume)
   - [Unsafe: RecognizerSession](#unsafe-recognizersession)
 - [API Reference](#api-reference)
 - [Requirements](#requirements)
@@ -107,6 +110,7 @@ Both permissions are required for speech recognition to work on iOS.
 | **Haptic feedback** | Optional haptics on recording start/stop | ✅ | ✅ |
 | **Background handling** | Auto-stop when app loses focus/goes to background | ✅ | Not Safe *(TODO)* |
 | **Permission handling** | Dedicated `onPermissionDenied` callback | ✅ | ✅ |
+| **Voice input volume** | Normalized voice input level for UI meters (`useVoiceInputVolume`) | ✅ | ✅ |
 | **Repeating word filter** | Removes consecutive duplicate words from artifacts | ✅ | ✅ |
 | **Locale support** | Configure speech recognizer for different languages | ✅ | ✅ |
 | **Contextual strings** | Domain-specific vocabulary for improved accuracy | ✅ | ✅ |
@@ -166,7 +170,7 @@ function MyComponent() {
         // iOS specific
         iosAddPunctuation: true,
         // Android specific
-        androidMaskOffensiveWords: false,
+        maskOffensiveWords: false,
         androidFormattingPreferQuality: false,
         androidUseWebSearchModel: false,
         androidDisableBatchHandling: false,
@@ -218,17 +222,58 @@ import { RecognizerRef } from '@gmessier/nitro-speech';
 RecognizerRef.startListening({ locale: 'en-US' });
 RecognizerRef.addAutoFinishTime(5000);
 RecognizerRef.updateAutoFinishTime(10000, true);
+RecognizerRef.getIsActive();
 RecognizerRef.stopListening();
 ```
 
 `RecognizerRef` exposes only method handlers and is safe for cross-component method access.
 
+### Voice input volume
+
+#### useVoiceInputVolume
+
+By default you have access to `useVoiceInputVolume` to read normalized voice input level (`0..1`) for UI meters.
+⚠️ **Technical limitation**: this approach re-renders component a lot.
+
+```typescript
+import { useVoiceInputVolume } from '@gmessier/nitro-speech';
+
+function VoiceMeter() {
+  const volume = useVoiceInputVolume();
+  return <Text>{volume.toFixed(2)}</Text>;
+}
+```
+
+#### Reanimated: useSharedValue, worklets, UI thread
+
+As a better alternative you can control volume via SharedValue and apply it only on UI thread with Reanimated
+This way you will avoid re-renders since the volume will be stored on UI thread
+
+```typescript
+function VoiceMeter() {
+  const sharedVolume = useSharedValue(0)
+  const {
+    // ...
+  } = useRecognizer(
+    {
+      // ...
+      onVolumeChange: (normVolume) => {
+        "worklet";
+        sharedVolume.value = normValue
+      },
+      // ...
+    }
+  );
+}
+```
+
+
 ### Unsafe: RecognizerSession
 
 `RecognizerSession` is the hybrid object. It gives direct access to callbacks and control methods, but it is unsafe to orchestrate the full session directly from it.
 
 ```typescript
-import { RecognizerSession } from '@gmessier/nitro-speech';
+import { RecognizerSession, unsafe_onVolumeChange } from '@gmessier/nitro-speech';
 
 // Set up callbacks
 RecognizerSession.onReadyForSpeech = () => {
@@ -255,6 +300,13 @@ RecognizerSession.onPermissionDenied = () => {
   console.log('Permission denied');
 };
 
+RecognizerSession.onVolumeChange = (volume) => {
+  console.log('new volume: ', volume);
+};
+// OR use unsafe_onVolumeChange to enable useVoiceInputVolume hook manually
+RecognizerSession.onVolumeChange = unsafe_onVolumeChange
+
+
 // Start listening
 RecognizerSession.startListening({
   locale: 'en-US',
@@ -305,6 +357,7 @@ The `RecognizerSession.dispose()` method is **NOT SAFE** and should rarely be us
 - `stopListening()` - Stop speech recognition
 - `addAutoFinishTime(additionalTimeMs?: number)` - Add time to the auto-finish timer (or reset to original if no parameter)
 - `updateAutoFinishTime(newTimeMs: number, withRefresh?: boolean)` - Update the auto-finish timer
+- `getIsActive()` - Returns true if the speech recognition is active
 
 ### `RecognizerRef`
 
@@ -312,6 +365,11 @@ The `RecognizerSession.dispose()` method is **NOT SAFE** and should rarely be us
 - `stopListening()`
 - `addAutoFinishTime(additionalTimeMs?: number)`
 - `updateAutoFinishTime(newTimeMs: number, withRefresh?: boolean)`
+- `getIsActive()`
+
+### `useVoiceInputVolume`
+
+- `useVoiceInputVolume(): number`
 
 ### `RecognizerSession`
 
@@ -328,8 +386,9 @@ Configuration object for speech recognition.
 - `autoFinishRecognitionMs?: number` - Auto-stop timeout in milliseconds (default: `8000`)
 - `contextualStrings?: string[]` - Array of domain-specific words for better recognition
 - `disableRepeatingFilter?: boolean` - Disable filter that removes consecutive duplicate words (default: `false`)
-- `startHapticFeedbackStyle?: 'light' | 'medium' | 'heavy'` - Haptic feedback style when microphone starts recording (default: `null` / disabled)
-- `stopHapticFeedbackStyle?: 'light' | 'medium' | 'heavy'` - Haptic feedback style when microphone stops recording (default: `null` / disabled)
+- `startHapticFeedbackStyle?: 'light' | 'medium' | 'heavy' | 'none'` - Haptic feedback style when microphone starts recording (default: `"medium"`)
+- `stopHapticFeedbackStyle?: 'light' | 'medium' | 'heavy' | 'none'` - Haptic feedback style when microphone stops recording (default: `"medium"`)
+- `maskOffensiveWords?: boolean` - Mask offensive words with asterisks. (Android 13+, iOS 26+, default: `false`. iOS <26: always `false`)
 
 #### iOS-Specific Parameters
 
@@ -337,7 +396,6 @@ Configuration object for speech recognition.
 
 #### Android-Specific Parameters
 
-- `androidMaskOffensiveWords?: boolean` - Mask offensive words (Android 13+, default: `false`)
 - `androidFormattingPreferQuality?: boolean` - Prefer quality over latency (Android 13+, default: `false`)
 - `androidUseWebSearchModel?: boolean` - Use web search language model instead of free-form (default: `false`)
 - `androidDisableBatchHandling?: boolean` - Disable default batch handling (may add many empty batches, default: `false`)
@@ -361,8 +419,3 @@ cd android && ./gradlew :react-native-nitro-modules:preBuild
 ## License
 
 MIT
-
-## TODO
-
-- [ ] (Android) Timer till the auto finish is called
-- [ ] (Android) Cleanup when app loses the focus

package/android/build.gradle

@@ -62,6 +62,8 @@ android {
         }
       }
     }
+
+    consumerProguardFiles 'proguard-rules.pro'
   }
 
   externalNativeBuild {

package/android/src/main/java/com/margelo/nitro/nitrospeech/recognizer/HapticImpact.kt

@@ -8,7 +8,7 @@ import android.os.VibratorManager
 import com.margelo.nitro.nitrospeech.HapticFeedbackStyle
 
 class HapticImpact(
-  private val style: HapticFeedbackStyle = HapticFeedbackStyle.MEDIUM,
+  private val style: HapticFeedbackStyle?
 ) {
   private data class LegacyOneShot(
     val durationMs: Long,
@@ -16,6 +16,10 @@ class HapticImpact(
   )
 
   fun trigger(context: Context) {
+    if (style == HapticFeedbackStyle.NONE) {
+      return
+    }
+
     val vibrator = getVibrator(context) ?: return
     if (!vibrator.hasVibrator()) return
 
@@ -25,7 +29,10 @@ class HapticImpact(
           HapticFeedbackStyle.LIGHT -> VibrationEffect.EFFECT_TICK
           HapticFeedbackStyle.MEDIUM -> VibrationEffect.EFFECT_CLICK
           HapticFeedbackStyle.HEAVY -> VibrationEffect.EFFECT_HEAVY_CLICK
+          null -> VibrationEffect.EFFECT_CLICK
+          else -> null
         }
+        if (effect == null) { return }
         vibrator.vibrate(VibrationEffect.createPredefined(effect))
         return
       }
@@ -34,7 +41,10 @@ class HapticImpact(
         HapticFeedbackStyle.LIGHT -> LegacyOneShot(durationMs = 12L, amplitude = 50)
         HapticFeedbackStyle.MEDIUM -> LegacyOneShot(durationMs = 18L, amplitude = 100)
         HapticFeedbackStyle.HEAVY -> LegacyOneShot(durationMs = 28L, amplitude = 180)
+        null -> LegacyOneShot(durationMs = 18L, amplitude = 100)
+        else -> null
       }
+      if (legacyOneShot == null) { return }
       if (Build.VERSION.SDK_INT >= Build.VERSION_CODES.O) {
         vibrator.vibrate(
           VibrationEffect.createOneShot(

package/android/src/main/java/com/margelo/nitro/nitrospeech/recognizer/HybridRecognizer.kt

@@ -33,6 +33,11 @@ class HybridRecognizer: HybridRecognizerSpec() {
   override var onAutoFinishProgress: ((timeLeftMs: Double) -> Unit)? = null
   override var onError: ((error: String) -> Unit)? = null
   override var onPermissionDenied: (() -> Unit)? = null
+  override var onVolumeChange: ((normVolume: Double) -> Unit)? = null
+
+  override fun getIsActive(): Boolean {
+    return isActive
+  }
 
   @DoNotStrip
   @Keep
@@ -86,7 +91,7 @@ class HybridRecognizer: HybridRecognizerSpec() {
     mainHandler.postDelayed({
       val context = NitroModules.applicationContext
       val hapticImpact = config?.stopHapticFeedbackStyle
-      if (hapticImpact != null && context != null) {
+      if (context != null) {
         HapticImpact(hapticImpact).trigger(context)
       }
       cleanup()
@@ -129,6 +134,7 @@ class HybridRecognizer: HybridRecognizerSpec() {
         val recognitionListenerSession = RecognitionListenerSession(
           autoStopper,
           config,
+          onVolumeChange
         ) { result: ArrayList<String>?, errorMessage: String?, recordingStopped: Boolean ->
           onFinishRecognition(result, errorMessage, recordingStopped)
         }
@@ -140,10 +146,10 @@ class HybridRecognizer: HybridRecognizerSpec() {
         intent.putExtra(RecognizerIntent.EXTRA_LANGUAGE_MODEL, languageModel)
         intent.putExtra(RecognizerIntent.EXTRA_LANGUAGE, config?.locale ?: "en-US")
         intent.putExtra(RecognizerIntent.EXTRA_PARTIAL_RESULTS, true)
-        // set many secs to avoid cutting early
+        // Set a lot of time to avoid cutting early
         intent.putExtra(RecognizerIntent.EXTRA_SPEECH_INPUT_COMPLETE_SILENCE_LENGTH_MILLIS, 300000)
 
-        if (config?.androidMaskOffensiveWords != true && Build.VERSION.SDK_INT >= Build.VERSION_CODES.TIRAMISU) {
+        if (config?.maskOffensiveWords != true && Build.VERSION.SDK_INT >= Build.VERSION_CODES.TIRAMISU) {
           intent.putExtra(RecognizerIntent.EXTRA_MASK_OFFENSIVE_WORDS, false)
         }
 
@@ -163,10 +169,8 @@ class HybridRecognizer: HybridRecognizerSpec() {
         isActive = true
         
         val hapticImpact = config?.startHapticFeedbackStyle
-        if (hapticImpact != null) {
-          HapticImpact(hapticImpact).trigger(context)
-        }
 
+        HapticImpact(hapticImpact).trigger(context)
         mainHandler.postDelayed({
           if (isActive) {
             onReadyForSpeech?.invoke()
@@ -192,6 +196,8 @@ class HybridRecognizer: HybridRecognizerSpec() {
       speechRecognizer?.destroy()
       speechRecognizer = null
       isActive = false
+      // Reset voice meter in JS consumers after stop/error cleanup.
+      onVolumeChange?.invoke(0.0)
     } catch (e: Exception) {
       onFinishRecognition(
         null,

package/android/src/main/java/com/margelo/nitro/nitrospeech/recognizer/RecognitionListenerSession.kt

@@ -5,17 +5,32 @@ import android.speech.RecognitionListener
 import android.speech.SpeechRecognizer
 import android.util.Log
 import com.margelo.nitro.nitrospeech.SpeechToTextParams
+import kotlin.math.max
+import kotlin.math.roundToInt
 
 class RecognitionListenerSession (
     private val autoStopper: AutoStopper?,
     private val config: SpeechToTextParams?,
+    private val onVolumeChange: ((normVolume: Double) -> Unit)?,
     private val onFinishRecognition: (result: ArrayList<String>?, errorMessage: String?, recordingStopped: Boolean) -> Unit,
 ) {
     companion object {
         private const val TAG = "HybridRecognizer"
+        private const val SPEECH_LEVEL_THRESHOLD = 0.08f
+        private const val FLOOR_RISE_ALPHA = 0.01f
+        private const val FLOOR_FALL_ALPHA = 0.20f
+        private const val PEAK_ATTACK_ALPHA = 0.25f
+        private const val PEAK_DECAY_ALPHA = 0.01f
+        private const val METER_ATTACK = 0.35f
+        private const val METER_RELEASE = 0.08f
+        private const val MIN_SPAN_DB = 6f
+        private const val PRECISION_SCALE = 1_000_000f
     }
 
     private var resultBatches: ArrayList<String>? = null
+    private var noiseFloorDb = Float.NaN
+    private var peakDb = Float.NaN
+    private var levelSmoothed = 0f
 
     fun createRecognitionListener(): RecognitionListener {
         resultBatches = null
@@ -23,7 +38,11 @@ class RecognitionListenerSession (
             override fun onReadyForSpeech(params: Bundle?) {}
             override fun onBeginningOfSpeech() {}
             override fun onRmsChanged(rmsdB: Float) {
-                autoStopper?.indicateRecordingActivity()
+                val normLevel = normalizeRmsDb(rmsdB)
+                onVolumeChange?.invoke(normLevel.toDouble())
+                if (normLevel > SPEECH_LEVEL_THRESHOLD) {
+                    autoStopper?.indicateRecordingActivity()
+                }
             }
             override fun onBufferReceived(buffer: ByteArray?) {}
             override fun onEndOfSpeech() {}
@@ -92,15 +111,62 @@ class RecognitionListenerSession (
         }
     }
 
-    // Filters out 2 or more repeating words in a row, like "and and"
+    // Filters out 2 or more consecutive duplicate words, like "and and"
     private fun repeatingFilter(text: String): String {
-        val words = text.split(Regex("\\s+")).toMutableList()
-        var joiner = words[0]
+        var words = text.split(Regex("\\s+")).filter { it.isNotBlank() }
+        if (words.isEmpty()) {
+            return ""
+        }
+
+        val joiner = StringBuilder()
+
+        // 10 - arbitrary number of last substrings that is still unstable
+        // and needs to be filtered. Prev substrings were handled earlier.
+        if (words.size >= 10) {
+            joiner.append(words.take(words.size - 9).joinToString(" "))
+            words = words.takeLast(10)
+        } else {
+            joiner.append(words.first())
+        }
+
         for (i in words.indices) {
             if (i == 0) continue
-            if (words[i] == words[i-1]) continue
-            joiner += " ${words[i]}"
+            // Always add number-containing strings.
+            if (Regex("\\d+").containsMatchIn(words[i])) {
+                joiner.append(" ").append(words[i])
+                continue
+            }
+
+            // Skip consecutive duplicate strings.
+            if (words[i] == words[i - 1]) continue
+            joiner.append(" ").append(words[i])
         }
-        return joiner
+        return joiner.toString()
+    }
+
+    private fun normalizeRmsDb(rmsdB: Float): Double {
+        if (!rmsdB.isFinite()) {
+            return 0.0
+        }
+
+        if (noiseFloorDb.isNaN()) {
+            noiseFloorDb = rmsdB
+        }
+        if (peakDb.isNaN()) {
+            peakDb = rmsdB + MIN_SPAN_DB
+        }
+
+        val floorAlpha = if (rmsdB < noiseFloorDb) FLOOR_FALL_ALPHA else FLOOR_RISE_ALPHA
+        noiseFloorDb += floorAlpha * (rmsdB - noiseFloorDb)
+
+        val peakAlpha = if (rmsdB > peakDb) PEAK_ATTACK_ALPHA else PEAK_DECAY_ALPHA
+        peakDb += peakAlpha * (rmsdB - peakDb)
+
+        val span = max(peakDb - noiseFloorDb, MIN_SPAN_DB)
+        val raw = ((rmsdB - noiseFloorDb) / span).coerceIn(0f, 1f)
+        val smoothingCoeff = if (raw > levelSmoothed) METER_ATTACK else METER_RELEASE
+        levelSmoothed += smoothingCoeff * (raw - levelSmoothed)
+
+        return ((levelSmoothed * PRECISION_SCALE).roundToInt() / PRECISION_SCALE).toDouble()
     }
   }