@absolutejs/voice 0.0.22-beta.584 → 0.0.22-beta.585

package/dist/index.js

@@ -391,22 +391,146 @@ var resolveLogger = (logger) => ({
   ...logger
 });
 
+// src/core/turnDetection.ts
+var DEFAULT_SILENCE_MS = 700;
+var DEFAULT_SPEECH_THRESHOLD = 0.015;
+var DEFAULT_SEMANTIC_VETO_RECHECK_MS = 1200;
+var toUint8Array = (audio) => {
+  if (audio instanceof ArrayBuffer) {
+    return new Uint8Array(audio);
+  }
+  return new Uint8Array(audio.buffer, audio.byteOffset, audio.byteLength);
+};
+var measureAudioLevel = (audio) => {
+  const bytes = toUint8Array(audio);
+  if (bytes.byteLength < 2) {
+    return 0;
+  }
+  const samples = new Int16Array(bytes.buffer, bytes.byteOffset, Math.floor(bytes.byteLength / 2));
+  if (samples.length === 0) {
+    return 0;
+  }
+  let sumSquares = 0;
+  for (const sample of samples) {
+    const normalized = sample / 32768;
+    sumSquares += normalized * normalized;
+  }
+  return Math.sqrt(sumSquares / samples.length);
+};
+var normalizeText = (value) => value.trim().replace(/\s+/g, " ");
+var countWords = (value) => value.length > 0 ? value.split(" ").length : 0;
+var selectPreferredTranscriptText = (currentText, nextText) => {
+  const current = normalizeText(currentText);
+  const next = normalizeText(nextText);
+  if (!current) {
+    return next;
+  }
+  if (!next) {
+    return current;
+  }
+  if (current === next || current.includes(next)) {
+    return current;
+  }
+  if (next.includes(current)) {
+    return next;
+  }
+  if (countWords(next) > countWords(current)) {
+    return next;
+  }
+  if (countWords(next) === countWords(current) && next.length > current.length) {
+    return next;
+  }
+  return current;
+};
+var mergeSequentialTranscriptText = (currentText, nextText) => {
+  const current = normalizeText(currentText);
+  const next = normalizeText(nextText);
+  if (!current) {
+    return next;
+  }
+  if (!next) {
+    return current;
+  }
+  const currentWords = current.split(" ");
+  const nextWords = next.split(" ");
+  const maxOverlap = Math.min(currentWords.length, nextWords.length);
+  for (let overlap = maxOverlap;overlap > 0; overlap -= 1) {
+    const currentSuffix = currentWords.slice(-overlap).join(" ");
+    const nextPrefix = nextWords.slice(0, overlap).join(" ");
+    if (currentSuffix === nextPrefix) {
+      return [...currentWords, ...nextWords.slice(overlap)].join(" ");
+    }
+  }
+  return `${current} ${next}`.trim();
+};
+var countCommonPrefixWords = (currentText, nextText) => {
+  const currentWords = normalizeText(currentText).split(" ").filter(Boolean);
+  const nextWords = normalizeText(nextText).split(" ").filter(Boolean);
+  const maxWords = Math.min(currentWords.length, nextWords.length);
+  let count = 0;
+  for (let index = 0;index < maxWords; index += 1) {
+    if (currentWords[index] !== nextWords[index]) {
+      break;
+    }
+    count += 1;
+  }
+  return count;
+};
+var mergeTranscriptTexts = (transcripts) => {
+  const merged = [];
+  for (const transcript of transcripts) {
+    const nextText = normalizeText(transcript.text);
+    if (!nextText) {
+      continue;
+    }
+    const previous = merged.at(-1);
+    if (!previous) {
+      merged.push(nextText);
+      continue;
+    }
+    if (nextText === previous || previous.includes(nextText)) {
+      continue;
+    }
+    if (nextText.includes(previous)) {
+      merged[merged.length - 1] = nextText;
+      continue;
+    }
+    merged.push(nextText);
+  }
+  return merged.join(" ").trim();
+};
+var buildTurnText = (transcripts, partialText, options = {}) => {
+  const finalText = mergeTranscriptTexts(transcripts);
+  const nextPartial = normalizeText(partialText);
+  const lastFinalEndedAtMs = [...transcripts].reverse().find((transcript) => typeof transcript.endedAtMs === "number")?.endedAtMs;
+  if (finalText && nextPartial && typeof lastFinalEndedAtMs === "number" && typeof options.partialStartedAtMs === "number" && options.partialStartedAtMs - lastFinalEndedAtMs >= 250 && countCommonPrefixWords(finalText, nextPartial) === 0) {
+    return mergeSequentialTranscriptText(finalText, nextPartial);
+  }
+  return selectPreferredTranscriptText(finalText, nextPartial);
+};
+
 // src/core/turnProfiles.ts
 var TURN_PROFILE_DEFAULTS = {
   balanced: {
     qualityProfile: "general",
+    semanticVetoMaxMs: 0,
+    semanticVetoRecheckMs: DEFAULT_SEMANTIC_VETO_RECHECK_MS,
     silenceMs: 1400,
     speechThreshold: 0.012,
     transcriptStabilityMs: 1000
   },
   fast: {
     qualityProfile: "general",
+    semanticVetoMaxMs: 0,
+    semanticVetoRecheckMs: DEFAULT_SEMANTIC_VETO_RECHECK_MS,
     silenceMs: 700,
     speechThreshold: 0.015,
     transcriptStabilityMs: 450
   },
   "long-form": {
     qualityProfile: "general",
+    semanticVetoMaxMs: 0,
+    semanticVetoRecheckMs: DEFAULT_SEMANTIC_VETO_RECHECK_MS,
     silenceMs: 2200,
     speechThreshold: 0.01,
     transcriptStabilityMs: 1500
@@ -440,6 +564,8 @@ var resolveTurnDetectionConfig = (config) => {
   return {
     profile,
     qualityProfile,
+    semanticVetoMaxMs: config?.semanticVetoMaxMs ?? preset.semanticVetoMaxMs,
+    semanticVetoRecheckMs: config?.semanticVetoRecheckMs ?? preset.semanticVetoRecheckMs,
     silenceMs: config?.silenceMs ?? quality.silenceMs ?? preset.silenceMs,
     speechThreshold: config?.speechThreshold ?? quality.speechThreshold ?? preset.speechThreshold,
     transcriptStabilityMs: config?.transcriptStabilityMs ?? quality.transcriptStabilityMs ?? preset.transcriptStabilityMs
@@ -3454,123 +3580,6 @@ var createVoiceTwilioRedirectHandoffAdapter = (options) => ({
   }
 });
 
-// src/core/turnDetection.ts
-var DEFAULT_SILENCE_MS = 700;
-var DEFAULT_SPEECH_THRESHOLD = 0.015;
-var toUint8Array = (audio) => {
-  if (audio instanceof ArrayBuffer) {
-    return new Uint8Array(audio);
-  }
-  return new Uint8Array(audio.buffer, audio.byteOffset, audio.byteLength);
-};
-var measureAudioLevel = (audio) => {
-  const bytes = toUint8Array(audio);
-  if (bytes.byteLength < 2) {
-    return 0;
-  }
-  const samples = new Int16Array(bytes.buffer, bytes.byteOffset, Math.floor(bytes.byteLength / 2));
-  if (samples.length === 0) {
-    return 0;
-  }
-  let sumSquares = 0;
-  for (const sample of samples) {
-    const normalized = sample / 32768;
-    sumSquares += normalized * normalized;
-  }
-  return Math.sqrt(sumSquares / samples.length);
-};
-var normalizeText = (value) => value.trim().replace(/\s+/g, " ");
-var countWords = (value) => value.length > 0 ? value.split(" ").length : 0;
-var selectPreferredTranscriptText = (currentText, nextText) => {
-  const current = normalizeText(currentText);
-  const next = normalizeText(nextText);
-  if (!current) {
-    return next;
-  }
-  if (!next) {
-    return current;
-  }
-  if (current === next || current.includes(next)) {
-    return current;
-  }
-  if (next.includes(current)) {
-    return next;
-  }
-  if (countWords(next) > countWords(current)) {
-    return next;
-  }
-  if (countWords(next) === countWords(current) && next.length > current.length) {
-    return next;
-  }
-  return current;
-};
-var mergeSequentialTranscriptText = (currentText, nextText) => {
-  const current = normalizeText(currentText);
-  const next = normalizeText(nextText);
-  if (!current) {
-    return next;
-  }
-  if (!next) {
-    return current;
-  }
-  const currentWords = current.split(" ");
-  const nextWords = next.split(" ");
-  const maxOverlap = Math.min(currentWords.length, nextWords.length);
-  for (let overlap = maxOverlap;overlap > 0; overlap -= 1) {
-    const currentSuffix = currentWords.slice(-overlap).join(" ");
-    const nextPrefix = nextWords.slice(0, overlap).join(" ");
-    if (currentSuffix === nextPrefix) {
-      return [...currentWords, ...nextWords.slice(overlap)].join(" ");
-    }
-  }
-  return `${current} ${next}`.trim();
-};
-var countCommonPrefixWords = (currentText, nextText) => {
-  const currentWords = normalizeText(currentText).split(" ").filter(Boolean);
-  const nextWords = normalizeText(nextText).split(" ").filter(Boolean);
-  const maxWords = Math.min(currentWords.length, nextWords.length);
-  let count = 0;
-  for (let index = 0;index < maxWords; index += 1) {
-    if (currentWords[index] !== nextWords[index]) {
-      break;
-    }
-    count += 1;
-  }
-  return count;
-};
-var mergeTranscriptTexts = (transcripts) => {
-  const merged = [];
-  for (const transcript of transcripts) {
-    const nextText = normalizeText(transcript.text);
-    if (!nextText) {
-      continue;
-    }
-    const previous = merged.at(-1);
-    if (!previous) {
-      merged.push(nextText);
-      continue;
-    }
-    if (nextText === previous || previous.includes(nextText)) {
-      continue;
-    }
-    if (nextText.includes(previous)) {
-      merged[merged.length - 1] = nextText;
-      continue;
-    }
-    merged.push(nextText);
-  }
-  return merged.join(" ").trim();
-};
-var buildTurnText = (transcripts, partialText, options = {}) => {
-  const finalText = mergeTranscriptTexts(transcripts);
-  const nextPartial = normalizeText(partialText);
-  const lastFinalEndedAtMs = [...transcripts].reverse().find((transcript) => typeof transcript.endedAtMs === "number")?.endedAtMs;
-  if (finalText && nextPartial && typeof lastFinalEndedAtMs === "number" && typeof options.partialStartedAtMs === "number" && options.partialStartedAtMs - lastFinalEndedAtMs >= 250 && countCommonPrefixWords(finalText, nextPartial) === 0) {
-    return mergeSequentialTranscriptText(finalText, nextPartial);
-  }
-  return selectPreferredTranscriptText(finalText, nextPartial);
-};
-
 // src/core/types.ts
 var ttsAdapterSessionCanCancel = (session) => typeof session.cancel === "function";
 
@@ -3907,8 +3916,11 @@ var createVoiceSession = (options) => {
   const turnDetection = {
     silenceMs: options.turnDetection.silenceMs ?? DEFAULT_SILENCE_MS,
     speechThreshold: options.turnDetection.speechThreshold ?? DEFAULT_SPEECH_THRESHOLD,
-    transcriptStabilityMs: options.turnDetection.transcriptStabilityMs ?? DEFAULT_TRANSCRIPT_STABILITY_MS
+    transcriptStabilityMs: options.turnDetection.transcriptStabilityMs ?? DEFAULT_TRANSCRIPT_STABILITY_MS,
+    semanticVetoMaxMs: options.turnDetection.semanticVetoMaxMs ?? 0,
+    semanticVetoRecheckMs: options.turnDetection.semanticVetoRecheckMs ?? DEFAULT_SEMANTIC_VETO_RECHECK_MS
   };
+  let semanticVetoElapsedMs = 0;
   const sttFallback = options.sttFallback ? {
     adapter: options.sttFallback.adapter,
     completionTimeoutMs: options.sttFallback.completionTimeoutMs ?? DEFAULT_FALLBACK_COMPLETION_TIMEOUT_MS,
@@ -4423,10 +4435,51 @@ var createVoiceSession = (options) => {
     silenceTimer = setTimeout(() => {
       silenceTimer = null;
       pendingCommitReason = null;
-      api.commitTurn(reason);
+      runScheduledCommit(reason);
     }, delayMs);
   };
   const scheduleSilenceCommit = (delayMs = turnDetection.silenceMs, reset = true) => scheduleTurnCommit(delayMs, "silence", reset);
+  const shouldDeferSilenceCommit = async (reason) => {
+    if (reason !== "silence" || turnDetection.semanticVetoMaxMs <= 0 || !options.semanticTurnDetector || semanticVetoElapsedMs >= turnDetection.semanticVetoMaxMs) {
+      return false;
+    }
+    const session = await readSession();
+    const { partialText, transcripts } = session.currentTurn;
+    const userText = buildTurnText(transcripts, partialText, {
+      partialEndedAtMs: session.currentTurn.partialEndedAt,
+      partialStartedAtMs: session.currentTurn.partialStartedAt
+    });
+    if (!userText) {
+      return false;
+    }
+    const silenceMs = session.currentTurn.silenceStartedAt !== undefined ? Date.now() - session.currentTurn.silenceStartedAt : turnDetection.silenceMs;
+    let endOfTurn = true;
+    try {
+      const verdict = await Promise.resolve(options.semanticTurnDetector.evaluate({
+        lastFinalTranscript: transcripts.at(-1),
+        partialText,
+        silenceMs,
+        transcripts
+      }));
+      endOfTurn = verdict.endOfTurn;
+    } catch {
+      return false;
+    }
+    if (endOfTurn !== false) {
+      return false;
+    }
+    const remaining = turnDetection.semanticVetoMaxMs - semanticVetoElapsedMs;
+    const extendMs = Math.max(1, Math.min(turnDetection.semanticVetoRecheckMs, remaining));
+    semanticVetoElapsedMs += extendMs;
+    scheduleTurnCommit(extendMs, reason);
+    return true;
+  };
+  const runScheduledCommit = async (reason) => {
+    if (await shouldDeferSilenceCommit(reason)) {
+      return;
+    }
+    await api.commitTurn(reason);
+  };
   const requestTurnCommit = async (reason) => {
     const session = await readSession();
     const text = buildTurnText(session.currentTurn.transcripts, session.currentTurn.partialText, {
@@ -5138,6 +5191,7 @@ var createVoiceSession = (options) => {
       session2.lastActivityAt = Date.now();
       session2.status = "active";
     });
+    semanticVetoElapsedMs = 0;
     if (silenceTimer && pendingCommitReason === "vendor") {
       scheduleTurnCommit(getVendorCommitDelayMs(), "vendor");
     }
@@ -5841,6 +5895,7 @@ var createVoiceSession = (options) => {
   };
   const commitTurnInternal = async (reason = "manual") => {
     clearSilenceTimer();
+    semanticVetoElapsedMs = 0;
     backchannelDriver?.reset();
     amdLastTurnCommitAt = Date.now();
     const session = await readSession();
@@ -42388,9 +42443,12 @@ var createVoiceConfiguration = (configuration) => configuration;
 var DEFAULT_SPEECH_THRESHOLD2 = 0.015;
 var DEFAULT_SILENCE_MS2 = 700;
 var DEFAULT_TRANSCRIPT_STABILITY_MS2 = 200;
+var DEFAULT_SEMANTIC_VETO_RECHECK_MS2 = 1200;
 var resolveTurnDetection = (input) => ({
   profile: input?.profile ?? "balanced",
   qualityProfile: input?.qualityProfile ?? "general",
+  semanticVetoMaxMs: input?.semanticVetoMaxMs ?? 0,
+  semanticVetoRecheckMs: input?.semanticVetoRecheckMs ?? DEFAULT_SEMANTIC_VETO_RECHECK_MS2,
   silenceMs: input?.silenceMs ?? DEFAULT_SILENCE_MS2,
   speechThreshold: input?.speechThreshold ?? DEFAULT_SPEECH_THRESHOLD2,
   transcriptStabilityMs: input?.transcriptStabilityMs ?? DEFAULT_TRANSCRIPT_STABILITY_MS2

package/dist/react/index.js

@@ -12243,22 +12243,146 @@ var resolveAudioConditioningConfig = (config) => {
   };
 };
 
+// src/core/turnDetection.ts
+var DEFAULT_SILENCE_MS = 700;
+var DEFAULT_SPEECH_THRESHOLD = 0.015;
+var DEFAULT_SEMANTIC_VETO_RECHECK_MS = 1200;
+var toUint8Array = (audio) => {
+  if (audio instanceof ArrayBuffer) {
+    return new Uint8Array(audio);
+  }
+  return new Uint8Array(audio.buffer, audio.byteOffset, audio.byteLength);
+};
+var measureAudioLevel = (audio) => {
+  const bytes = toUint8Array(audio);
+  if (bytes.byteLength < 2) {
+    return 0;
+  }
+  const samples = new Int16Array(bytes.buffer, bytes.byteOffset, Math.floor(bytes.byteLength / 2));
+  if (samples.length === 0) {
+    return 0;
+  }
+  let sumSquares = 0;
+  for (const sample of samples) {
+    const normalized = sample / 32768;
+    sumSquares += normalized * normalized;
+  }
+  return Math.sqrt(sumSquares / samples.length);
+};
+var normalizeText = (value) => value.trim().replace(/\s+/g, " ");
+var countWords = (value) => value.length > 0 ? value.split(" ").length : 0;
+var selectPreferredTranscriptText = (currentText, nextText) => {
+  const current = normalizeText(currentText);
+  const next = normalizeText(nextText);
+  if (!current) {
+    return next;
+  }
+  if (!next) {
+    return current;
+  }
+  if (current === next || current.includes(next)) {
+    return current;
+  }
+  if (next.includes(current)) {
+    return next;
+  }
+  if (countWords(next) > countWords(current)) {
+    return next;
+  }
+  if (countWords(next) === countWords(current) && next.length > current.length) {
+    return next;
+  }
+  return current;
+};
+var mergeSequentialTranscriptText = (currentText, nextText) => {
+  const current = normalizeText(currentText);
+  const next = normalizeText(nextText);
+  if (!current) {
+    return next;
+  }
+  if (!next) {
+    return current;
+  }
+  const currentWords = current.split(" ");
+  const nextWords = next.split(" ");
+  const maxOverlap = Math.min(currentWords.length, nextWords.length);
+  for (let overlap = maxOverlap;overlap > 0; overlap -= 1) {
+    const currentSuffix = currentWords.slice(-overlap).join(" ");
+    const nextPrefix = nextWords.slice(0, overlap).join(" ");
+    if (currentSuffix === nextPrefix) {
+      return [...currentWords, ...nextWords.slice(overlap)].join(" ");
+    }
+  }
+  return `${current} ${next}`.trim();
+};
+var countCommonPrefixWords = (currentText, nextText) => {
+  const currentWords = normalizeText(currentText).split(" ").filter(Boolean);
+  const nextWords = normalizeText(nextText).split(" ").filter(Boolean);
+  const maxWords = Math.min(currentWords.length, nextWords.length);
+  let count = 0;
+  for (let index = 0;index < maxWords; index += 1) {
+    if (currentWords[index] !== nextWords[index]) {
+      break;
+    }
+    count += 1;
+  }
+  return count;
+};
+var mergeTranscriptTexts = (transcripts) => {
+  const merged = [];
+  for (const transcript of transcripts) {
+    const nextText = normalizeText(transcript.text);
+    if (!nextText) {
+      continue;
+    }
+    const previous = merged.at(-1);
+    if (!previous) {
+      merged.push(nextText);
+      continue;
+    }
+    if (nextText === previous || previous.includes(nextText)) {
+      continue;
+    }
+    if (nextText.includes(previous)) {
+      merged[merged.length - 1] = nextText;
+      continue;
+    }
+    merged.push(nextText);
+  }
+  return merged.join(" ").trim();
+};
+var buildTurnText = (transcripts, partialText, options = {}) => {
+  const finalText = mergeTranscriptTexts(transcripts);
+  const nextPartial = normalizeText(partialText);
+  const lastFinalEndedAtMs = [...transcripts].reverse().find((transcript) => typeof transcript.endedAtMs === "number")?.endedAtMs;
+  if (finalText && nextPartial && typeof lastFinalEndedAtMs === "number" && typeof options.partialStartedAtMs === "number" && options.partialStartedAtMs - lastFinalEndedAtMs >= 250 && countCommonPrefixWords(finalText, nextPartial) === 0) {
+    return mergeSequentialTranscriptText(finalText, nextPartial);
+  }
+  return selectPreferredTranscriptText(finalText, nextPartial);
+};
+
 // src/core/turnProfiles.ts
 var TURN_PROFILE_DEFAULTS = {
   balanced: {
     qualityProfile: "general",
+    semanticVetoMaxMs: 0,
+    semanticVetoRecheckMs: DEFAULT_SEMANTIC_VETO_RECHECK_MS,
     silenceMs: 1400,
     speechThreshold: 0.012,
     transcriptStabilityMs: 1000
   },
   fast: {
     qualityProfile: "general",
+    semanticVetoMaxMs: 0,
+    semanticVetoRecheckMs: DEFAULT_SEMANTIC_VETO_RECHECK_MS,
     silenceMs: 700,
     speechThreshold: 0.015,
     transcriptStabilityMs: 450
   },
   "long-form": {
     qualityProfile: "general",
+    semanticVetoMaxMs: 0,
+    semanticVetoRecheckMs: DEFAULT_SEMANTIC_VETO_RECHECK_MS,
     silenceMs: 2200,
     speechThreshold: 0.01,
     transcriptStabilityMs: 1500
@@ -12292,6 +12416,8 @@ var resolveTurnDetectionConfig = (config) => {
   return {
     profile,
     qualityProfile,
+    semanticVetoMaxMs: config?.semanticVetoMaxMs ?? preset.semanticVetoMaxMs,
+    semanticVetoRecheckMs: config?.semanticVetoRecheckMs ?? preset.semanticVetoRecheckMs,
     silenceMs: config?.silenceMs ?? quality.silenceMs ?? preset.silenceMs,
     speechThreshold: config?.speechThreshold ?? quality.speechThreshold ?? preset.speechThreshold,
     transcriptStabilityMs: config?.transcriptStabilityMs ?? quality.transcriptStabilityMs ?? preset.transcriptStabilityMs

package/dist/svelte/index.js

@@ -1380,22 +1380,146 @@ var resolveAudioConditioningConfig = (config) => {
   };
 };
 
+// src/core/turnDetection.ts
+var DEFAULT_SILENCE_MS = 700;
+var DEFAULT_SPEECH_THRESHOLD = 0.015;
+var DEFAULT_SEMANTIC_VETO_RECHECK_MS = 1200;
+var toUint8Array = (audio) => {
+  if (audio instanceof ArrayBuffer) {
+    return new Uint8Array(audio);
+  }
+  return new Uint8Array(audio.buffer, audio.byteOffset, audio.byteLength);
+};
+var measureAudioLevel = (audio) => {
+  const bytes = toUint8Array(audio);
+  if (bytes.byteLength < 2) {
+    return 0;
+  }
+  const samples = new Int16Array(bytes.buffer, bytes.byteOffset, Math.floor(bytes.byteLength / 2));
+  if (samples.length === 0) {
+    return 0;
+  }
+  let sumSquares = 0;
+  for (const sample of samples) {
+    const normalized = sample / 32768;
+    sumSquares += normalized * normalized;
+  }
+  return Math.sqrt(sumSquares / samples.length);
+};
+var normalizeText = (value) => value.trim().replace(/\s+/g, " ");
+var countWords = (value) => value.length > 0 ? value.split(" ").length : 0;
+var selectPreferredTranscriptText = (currentText, nextText) => {
+  const current = normalizeText(currentText);
+  const next = normalizeText(nextText);
+  if (!current) {
+    return next;
+  }
+  if (!next) {
+    return current;
+  }
+  if (current === next || current.includes(next)) {
+    return current;
+  }
+  if (next.includes(current)) {
+    return next;
+  }
+  if (countWords(next) > countWords(current)) {
+    return next;
+  }
+  if (countWords(next) === countWords(current) && next.length > current.length) {
+    return next;
+  }
+  return current;
+};
+var mergeSequentialTranscriptText = (currentText, nextText) => {
+  const current = normalizeText(currentText);
+  const next = normalizeText(nextText);
+  if (!current) {
+    return next;
+  }
+  if (!next) {
+    return current;
+  }
+  const currentWords = current.split(" ");
+  const nextWords = next.split(" ");
+  const maxOverlap = Math.min(currentWords.length, nextWords.length);
+  for (let overlap = maxOverlap;overlap > 0; overlap -= 1) {
+    const currentSuffix = currentWords.slice(-overlap).join(" ");
+    const nextPrefix = nextWords.slice(0, overlap).join(" ");
+    if (currentSuffix === nextPrefix) {
+      return [...currentWords, ...nextWords.slice(overlap)].join(" ");
+    }
+  }
+  return `${current} ${next}`.trim();
+};
+var countCommonPrefixWords = (currentText, nextText) => {
+  const currentWords = normalizeText(currentText).split(" ").filter(Boolean);
+  const nextWords = normalizeText(nextText).split(" ").filter(Boolean);
+  const maxWords = Math.min(currentWords.length, nextWords.length);
+  let count = 0;
+  for (let index = 0;index < maxWords; index += 1) {
+    if (currentWords[index] !== nextWords[index]) {
+      break;
+    }
+    count += 1;
+  }
+  return count;
+};
+var mergeTranscriptTexts = (transcripts) => {
+  const merged = [];
+  for (const transcript of transcripts) {
+    const nextText = normalizeText(transcript.text);
+    if (!nextText) {
+      continue;
+    }
+    const previous = merged.at(-1);
+    if (!previous) {
+      merged.push(nextText);
+      continue;
+    }
+    if (nextText === previous || previous.includes(nextText)) {
+      continue;
+    }
+    if (nextText.includes(previous)) {
+      merged[merged.length - 1] = nextText;
+      continue;
+    }
+    merged.push(nextText);
+  }
+  return merged.join(" ").trim();
+};
+var buildTurnText = (transcripts, partialText, options = {}) => {
+  const finalText = mergeTranscriptTexts(transcripts);
+  const nextPartial = normalizeText(partialText);
+  const lastFinalEndedAtMs = [...transcripts].reverse().find((transcript) => typeof transcript.endedAtMs === "number")?.endedAtMs;
+  if (finalText && nextPartial && typeof lastFinalEndedAtMs === "number" && typeof options.partialStartedAtMs === "number" && options.partialStartedAtMs - lastFinalEndedAtMs >= 250 && countCommonPrefixWords(finalText, nextPartial) === 0) {
+    return mergeSequentialTranscriptText(finalText, nextPartial);
+  }
+  return selectPreferredTranscriptText(finalText, nextPartial);
+};
+
 // src/core/turnProfiles.ts
 var TURN_PROFILE_DEFAULTS = {
   balanced: {
     qualityProfile: "general",
+    semanticVetoMaxMs: 0,
+    semanticVetoRecheckMs: DEFAULT_SEMANTIC_VETO_RECHECK_MS,
     silenceMs: 1400,
     speechThreshold: 0.012,
     transcriptStabilityMs: 1000
   },
   fast: {
     qualityProfile: "general",
+    semanticVetoMaxMs: 0,
+    semanticVetoRecheckMs: DEFAULT_SEMANTIC_VETO_RECHECK_MS,
     silenceMs: 700,
     speechThreshold: 0.015,
     transcriptStabilityMs: 450
   },
   "long-form": {
     qualityProfile: "general",
+    semanticVetoMaxMs: 0,
+    semanticVetoRecheckMs: DEFAULT_SEMANTIC_VETO_RECHECK_MS,
     silenceMs: 2200,
     speechThreshold: 0.01,
     transcriptStabilityMs: 1500
@@ -1429,6 +1553,8 @@ var resolveTurnDetectionConfig = (config) => {
   return {
     profile,
     qualityProfile,
+    semanticVetoMaxMs: config?.semanticVetoMaxMs ?? preset.semanticVetoMaxMs,
+    semanticVetoRecheckMs: config?.semanticVetoRecheckMs ?? preset.semanticVetoRecheckMs,
     silenceMs: config?.silenceMs ?? quality.silenceMs ?? preset.silenceMs,
     speechThreshold: config?.speechThreshold ?? quality.speechThreshold ?? preset.speechThreshold,
     transcriptStabilityMs: config?.transcriptStabilityMs ?? quality.transcriptStabilityMs ?? preset.transcriptStabilityMs