@absolutejs/voice 0.0.22-beta.607 → 0.0.22-beta.608

package/dist/core/backchannel.d.ts

@@ -21,4 +21,5 @@ export type VoiceBackchannelDriver = {
     noteSilence: (timestampMs?: number) => void;
     reset: () => void;
 };
+export declare const isBackchannelUtterance: (text: string, maxWords?: number) => boolean;
 export declare const createVoiceBackchannelDriver: (options: VoiceBackchannelDriverOptions) => VoiceBackchannelDriver;

package/dist/core/types.d.ts

@@ -802,6 +802,14 @@ export type VoicePluginConfig<TContext = unknown, TSession extends VoiceSessionR
     turnDetection?: VoiceTurnDetectionConfig;
     semanticTurnDetector?: import("./semanticTurn").VoiceSemanticTurnDetector;
     bargeInMinPartialWords?: number;
+    /**
+     * When true, a pure listening cue ("mm-hm", "yeah", "right", "got it") spoken
+     * WHILE the assistant is talking does NOT barge-in — the assistant keeps going
+     * and the cue is dropped so it never becomes the caller's next turn. A bare
+     * "yeah" said AFTER the assistant finishes is a normal answer, unaffected.
+     * Default false (any in-speech words interrupt, the prior behavior).
+     */
+    backchannelBargeInGuard?: boolean;
     fillerPhrases?: ReadonlyArray<string>;
     fillerDelayMs?: number;
     fillerFor?: (input: {
@@ -975,6 +983,14 @@ export type CreateVoiceSessionOptions<TContext = unknown, TSession extends Voice
      * Word splitting is whitespace-based. Punctuation is left attached.
      */
     bargeInMinPartialWords?: number;
+    /**
+     * When true, a pure listening cue ("mm-hm", "yeah", "right", "got it") spoken
+     * WHILE the assistant is talking does NOT barge-in — the assistant keeps going
+     * and the cue is dropped so it never becomes the caller's next turn. A bare
+     * "yeah" said AFTER the assistant finishes is a normal answer, unaffected.
+     * Default false (any in-speech words interrupt, the prior behavior).
+     */
+    backchannelBargeInGuard?: boolean;
     fillerPhrases?: ReadonlyArray<string>;
     /** Milliseconds after turn-commit before the filler fires. Default 250ms — short enough to feel instant, long enough to skip if the LLM is very fast. */
     fillerDelayMs?: number;

package/dist/index.js

@@ -3118,6 +3118,100 @@ var DEFAULT_CUES = [
   { text: "right" },
   { text: "go on" }
 ];
+var BACKCHANNEL_TOKENS = new Set([
+  "mm",
+  "mmm",
+  "mhm",
+  "mmhm",
+  "mmhmm",
+  "hm",
+  "hmm",
+  "uh-huh",
+  "uhhuh",
+  "uh",
+  "huh",
+  "ah",
+  "oh",
+  "yeah",
+  "yep",
+  "yup",
+  "yes",
+  "ya",
+  "yah",
+  "ok",
+  "okay",
+  "k",
+  "kay",
+  "right",
+  "sure",
+  "totally",
+  "exactly",
+  "absolutely",
+  "definitely",
+  "gotcha",
+  "cool",
+  "nice",
+  "wow",
+  "true",
+  "fair",
+  "aha",
+  "perfect",
+  "awesome",
+  "great",
+  "good",
+  "wonderful",
+  "amazing",
+  "interesting",
+  "understood",
+  "agreed"
+]);
+var BACKCHANNEL_PHRASES = new Set([
+  "i see",
+  "got it",
+  "makes sense",
+  "of course",
+  "for sure",
+  "fair enough",
+  "sounds good",
+  "i know",
+  "oh ok",
+  "oh okay",
+  "that's right",
+  "thats right",
+  "oh wow",
+  "oh nice",
+  "oh cool",
+  "uh huh",
+  "mm hm",
+  "mm hmm",
+  "i hear you",
+  "for real",
+  "no way",
+  "makes total sense",
+  "got you",
+  "i get it",
+  "right right",
+  "yeah yeah",
+  "ok ok",
+  "oh i see",
+  "oh got it",
+  "yeah totally",
+  "yeah exactly"
+]);
+var isBackchannelUtterance = (text, maxWords = 3) => {
+  const normalized = text.toLowerCase().replace(/[^a-z']/g, " ").replace(/\s+/g, " ").trim();
+  if (!normalized) {
+    return false;
+  }
+  if (BACKCHANNEL_PHRASES.has(normalized)) {
+    return true;
+  }
+  const words = normalized.split(" ");
+  if (words.length > maxWords) {
+    return false;
+  }
+  return words.every((word) => BACKCHANNEL_TOKENS.has(word));
+};
 var createVoiceBackchannelDriver = (options) => {
   const cues = options.cues ?? DEFAULT_CUES;
   const minSpeechMs = options.minSpeechMs ?? 2500;
@@ -3757,6 +3851,7 @@ var MAX_TTS_CHUNK_CHARS = 320;
 var STREAM_SENTENCE_END = /[.!?\u2026]['")\]]*$/;
 var STREAM_IDLE_FLUSH_MS = 350;
 var SPECULATIVE_DELAY_MS = 500;
+var BACKCHANNEL_DROP_WINDOW_MS = 2000;
 var nextSpeakableBoundary = (buffer) => {
   const match = STREAM_SENTENCE_BOUNDARY.exec(buffer);
   return match ? match.index + match[0].length : -1;
@@ -3996,6 +4091,8 @@ var createVoiceSession = (options) => {
   const fillerPhrases = (options.fillerPhrases ?? []).filter((p) => typeof p === "string" && p.trim().length > 0);
   const fillerDelayMs = options.fillerDelayMs ?? 250;
   const bargeInMinPartialWords = Math.max(1, options.bargeInMinPartialWords ?? 1);
+  const backchannelBargeInGuard = options.backchannelBargeInGuard ?? false;
+  let backchannelSuppressedAt = null;
   const fillerFor = options.fillerFor;
   const fillerForTimeoutMs = options.fillerForTimeoutMs ?? 600;
   const currentTurnAudio = [];
@@ -5169,7 +5266,19 @@ var createVoiceSession = (options) => {
       const triggeringText = transcript.text.trim();
       if (triggeringText) {
         const wordCount = triggeringText.split(/\s+/).length;
-        if (wordCount >= bargeInMinPartialWords) {
+        if (wordCount >= bargeInMinPartialWords && backchannelBargeInGuard && isBackchannelUtterance(triggeringText)) {
+          backchannelSuppressedAt = Date.now();
+          appendTurnLatencyStage({
+            metadata: {
+              partial: triggeringText.slice(0, 200),
+              reason: "backchannel",
+              wordCount
+            },
+            stage: "barge_in_suppressed",
+            turnId: activeTTSTurnId
+          }).catch(() => {});
+        } else if (wordCount >= bargeInMinPartialWords) {
+          backchannelSuppressedAt = null;
           appendTurnLatencyStage({
             metadata: {
               partial: triggeringText.slice(0, 200),
@@ -5231,6 +5340,11 @@ var createVoiceSession = (options) => {
   };
   const handleFinal = async (transcript) => {
     sttReconnectCount = 0;
+    if (backchannelBargeInGuard && backchannelSuppressedAt !== null && Date.now() - backchannelSuppressedAt < BACKCHANNEL_DROP_WINDOW_MS && isBackchannelUtterance(transcript.text.trim())) {
+      backchannelSuppressedAt = null;
+      return;
+    }
+    backchannelSuppressedAt = null;
     const session = await writeSession((session2) => {
       const alreadyPresent = session2.currentTurn.transcripts.some((existing) => existing.id === transcript.id);
       if (!alreadyPresent) {
@@ -39581,6 +39695,7 @@ var voice = (config) => {
       sttLifecycle: sessionOptions.sttLifecycle,
       ...config.semanticTurnDetector ? { semanticTurnDetector: config.semanticTurnDetector } : {},
       ...config.bargeInMinPartialWords !== undefined ? { bargeInMinPartialWords: config.bargeInMinPartialWords } : {},
+      ...config.backchannelBargeInGuard !== undefined ? { backchannelBargeInGuard: config.backchannelBargeInGuard } : {},
       ...config.fillerPhrases ? { fillerPhrases: config.fillerPhrases } : {},
       ...config.fillerDelayMs !== undefined ? { fillerDelayMs: config.fillerDelayMs } : {},
       ...config.fillerFor ? { fillerFor: config.fillerFor } : {},

package/dist/testing/index.js

@@ -5450,6 +5450,100 @@ var DEFAULT_CUES = [
   { text: "right" },
   { text: "go on" }
 ];
+var BACKCHANNEL_TOKENS = new Set([
+  "mm",
+  "mmm",
+  "mhm",
+  "mmhm",
+  "mmhmm",
+  "hm",
+  "hmm",
+  "uh-huh",
+  "uhhuh",
+  "uh",
+  "huh",
+  "ah",
+  "oh",
+  "yeah",
+  "yep",
+  "yup",
+  "yes",
+  "ya",
+  "yah",
+  "ok",
+  "okay",
+  "k",
+  "kay",
+  "right",
+  "sure",
+  "totally",
+  "exactly",
+  "absolutely",
+  "definitely",
+  "gotcha",
+  "cool",
+  "nice",
+  "wow",
+  "true",
+  "fair",
+  "aha",
+  "perfect",
+  "awesome",
+  "great",
+  "good",
+  "wonderful",
+  "amazing",
+  "interesting",
+  "understood",
+  "agreed"
+]);
+var BACKCHANNEL_PHRASES = new Set([
+  "i see",
+  "got it",
+  "makes sense",
+  "of course",
+  "for sure",
+  "fair enough",
+  "sounds good",
+  "i know",
+  "oh ok",
+  "oh okay",
+  "that's right",
+  "thats right",
+  "oh wow",
+  "oh nice",
+  "oh cool",
+  "uh huh",
+  "mm hm",
+  "mm hmm",
+  "i hear you",
+  "for real",
+  "no way",
+  "makes total sense",
+  "got you",
+  "i get it",
+  "right right",
+  "yeah yeah",
+  "ok ok",
+  "oh i see",
+  "oh got it",
+  "yeah totally",
+  "yeah exactly"
+]);
+var isBackchannelUtterance = (text, maxWords = 3) => {
+  const normalized = text.toLowerCase().replace(/[^a-z']/g, " ").replace(/\s+/g, " ").trim();
+  if (!normalized) {
+    return false;
+  }
+  if (BACKCHANNEL_PHRASES.has(normalized)) {
+    return true;
+  }
+  const words = normalized.split(" ");
+  if (words.length > maxWords) {
+    return false;
+  }
+  return words.every((word) => BACKCHANNEL_TOKENS.has(word));
+};
 var createVoiceBackchannelDriver = (options) => {
   const cues = options.cues ?? DEFAULT_CUES;
   const minSpeechMs = options.minSpeechMs ?? 2500;
@@ -5984,6 +6078,7 @@ var MAX_TTS_CHUNK_CHARS = 320;
 var STREAM_SENTENCE_END = /[.!?\u2026]['")\]]*$/;
 var STREAM_IDLE_FLUSH_MS = 350;
 var SPECULATIVE_DELAY_MS = 500;
+var BACKCHANNEL_DROP_WINDOW_MS = 2000;
 var nextSpeakableBoundary = (buffer) => {
   const match = STREAM_SENTENCE_BOUNDARY.exec(buffer);
   return match ? match.index + match[0].length : -1;
@@ -6223,6 +6318,8 @@ var createVoiceSession = (options) => {
   const fillerPhrases = (options.fillerPhrases ?? []).filter((p) => typeof p === "string" && p.trim().length > 0);
   const fillerDelayMs = options.fillerDelayMs ?? 250;
   const bargeInMinPartialWords = Math.max(1, options.bargeInMinPartialWords ?? 1);
+  const backchannelBargeInGuard = options.backchannelBargeInGuard ?? false;
+  let backchannelSuppressedAt = null;
   const fillerFor = options.fillerFor;
   const fillerForTimeoutMs = options.fillerForTimeoutMs ?? 600;
   const currentTurnAudio = [];
@@ -7396,7 +7493,19 @@ var createVoiceSession = (options) => {
       const triggeringText = transcript.text.trim();
       if (triggeringText) {
         const wordCount = triggeringText.split(/\s+/).length;
-        if (wordCount >= bargeInMinPartialWords) {
+        if (wordCount >= bargeInMinPartialWords && backchannelBargeInGuard && isBackchannelUtterance(triggeringText)) {
+          backchannelSuppressedAt = Date.now();
+          appendTurnLatencyStage({
+            metadata: {
+              partial: triggeringText.slice(0, 200),
+              reason: "backchannel",
+              wordCount
+            },
+            stage: "barge_in_suppressed",
+            turnId: activeTTSTurnId
+          }).catch(() => {});
+        } else if (wordCount >= bargeInMinPartialWords) {
+          backchannelSuppressedAt = null;
           appendTurnLatencyStage({
             metadata: {
               partial: triggeringText.slice(0, 200),
@@ -7458,6 +7567,11 @@ var createVoiceSession = (options) => {
   };
   const handleFinal = async (transcript) => {
     sttReconnectCount = 0;
+    if (backchannelBargeInGuard && backchannelSuppressedAt !== null && Date.now() - backchannelSuppressedAt < BACKCHANNEL_DROP_WINDOW_MS && isBackchannelUtterance(transcript.text.trim())) {
+      backchannelSuppressedAt = null;
+      return;
+    }
+    backchannelSuppressedAt = null;
     const session = await writeSession((session2) => {
       const alreadyPresent = session2.currentTurn.transcripts.some((existing) => existing.id === transcript.id);
       if (!alreadyPresent) {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@absolutejs/voice",
-  "version": "0.0.22-beta.607",
+  "version": "0.0.22-beta.608",
   "description": "Voice primitives and Elysia plugin for AbsoluteJS",
   "repository": {
     "type": "git",