@lumiastream/wakeword 1.1.6 → 1.1.7

package/Readme.md

@@ -96,6 +96,7 @@ Starts the wake word detection process.
 - `sampleRate` (number): Sample rate, default 16000
 - `grammar` (string[]): Array of wake words to detect
 - `confidence` (number): Confidence threshold (0-1), default 0.7
+- `modelPath` (string): Full path to a Vosk model directory (optional)
 
 **Returns:** ChildProcess instance
 
@@ -153,4 +154,4 @@ Source: https://github.com/zackees/static-sox/tree/main/bin
 
 ## License
 
-See LICENSE file for details.
+See LICENSE file for details.

package/lib/index.js

@@ -19,6 +19,7 @@ export { listAudioDevices };
  * @param {number} [options.sampleRate] - Sample rate (default: 16000)
  * @param {Array<string>} [options.grammar] - Wake words to detect
  * @param {number} [options.confidence] - Confidence threshold (0-1)
+ * @param {string} [options.modelPath] - Full path to a Vosk model directory
  * @returns {ChildProcess} The spawned voice detection process
  */
 export function startWakeWord(options = {}) {
@@ -28,6 +29,7 @@ export function startWakeWord(options = {}) {
 		sampleRate = 16000,
 		grammar = [],
 		confidence = 0.7,
+		modelPath = null,
 	} = options;
 
 	const voiceScript = join(here, "voice.js");
@@ -50,6 +52,9 @@ export function startWakeWord(options = {}) {
 	if (device && !args[2]) {
 		env.AUDIO_DEVICE = device;
 	}
+	if (modelPath) {
+		env.LUMIA_VOICE_MODEL_PATH = modelPath;
+	}
 
 	const proc = spawn("node", args, {
 		env,
@@ -72,4 +77,4 @@ export function startWakeWord(options = {}) {
 export default {
 	listAudioDevices,
 	startWakeWord,
-};
+};

package/lib/voice.js

@@ -62,6 +62,7 @@ const defaultExeName = {
 }[process.platform];
 const exeName = defaultExeName;
 const MATCH_SENTENCE = toBool(process.env.LUMIA_VOICE_MATCH_SENTENCE);
+const DISABLE_GRAMMAR = toBool(process.env.LUMIA_VOICE_DISABLE_GRAMMAR);
 
 /* Priority for sox path: argv[2] → fallback to sibling binaries/<exe> */
 /* Priority for device: argv[3] → env var → default */
@@ -91,7 +92,8 @@ try {
 /* ------------------------------------------------------------------ */
 /* 2. Resolve Vosk model                                              */
 /* ------------------------------------------------------------------ */
-let modelPath = join(here, "..", "models", "vosk-model-small-en-us-0.15");
+const envModelPath = (process.env.LUMIA_VOICE_MODEL_PATH || "").trim();
+let modelPath = envModelPath || join(here, "..", "models", "vosk-model-small-en-us-0.15");
 modelPath = unpacked(modelPath);
 
 if (!existsSync(modelPath))
@@ -108,7 +110,7 @@ let COMMANDS = [];
 
 const model = new Model(modelPath);
 const buildRecognizer = () => {
-	const recognizer = MATCH_SENTENCE
+	const recognizer = MATCH_SENTENCE || DISABLE_GRAMMAR
 		? new Recognizer({ model, sampleRate: SAMPLE_RATE })
 		: new Recognizer({ model, sampleRate: SAMPLE_RATE, grammar: GRAMMAR });
 	recognizer.setWords(true);
@@ -188,8 +190,14 @@ mic.on("data", (buf) => {
 		if (fullResult && fullResult.result && Array.isArray(fullResult.result)) {
 			let recognizedWords = [];
 			let totalConfidence = 0;
+			let totalConfidenceAll = 0;
+			let totalConfidenceCount = 0;
 
 			for (const wordDetail of fullResult.result) {
+				if (typeof wordDetail.conf === "number") {
+					totalConfidenceAll += wordDetail.conf;
+					totalConfidenceCount += 1;
+				}
 				// Each word has its own confidence ('conf')
 				if (wordDetail.conf >= WORD_CONFIDENCE_THRESHOLD) {
 					recognizedWords.push(wordDetail.word);
@@ -204,10 +212,12 @@ mic.on("data", (buf) => {
 			}
 
 			const finalRecognizedText = recognizedWords.join(" ").trim();
+			const averageConfidenceAll =
+				totalConfidenceCount > 0 ? totalConfidenceAll / totalConfidenceCount : 0;
 			const averageConfidence =
 				recognizedWords.length > 0
 					? totalConfidence / recognizedWords.length
-					: 0;
+					: averageConfidenceAll;
 
 			handle(finalRecognizedText, averageConfidence, fullResult.text); // Pass both the filtered text and an average confidence
 		} else if (fullResult && fullResult.text) {
@@ -262,6 +272,12 @@ function handle(processedWord, averageConfidence, originalText) {
 	findMatches(normalizedProcessed);
 	findMatches(normalizedOriginal, [...confidentCommands]);
 
+	// If word-level confidence filtering removed all words, fall back to the
+	// original text when overall confidence is still acceptable.
+	if (!matches.size && normalizedOriginal && averageConfidence >= WORD_CONFIDENCE_THRESHOLD) {
+		findMatches(normalizedOriginal);
+	}
+
 	if (!matches.size) return;
 
 	matches.forEach((match) => {

package/package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@lumiastream/wakeword",
-	"version": "1.1.6",
+	"version": "1.1.7",
 	"type": "module",
 	"main": "lib/index.js",
 	"files": [