@absolutejs/voice 0.0.22-beta.598 → 0.0.22-beta.599

package/dist/svelte/index.js

@@ -1409,146 +1409,25 @@ 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,
+    minSilenceMs: 400,
     silenceMs: 1400,
     speechThreshold: 0.012,
     transcriptStabilityMs: 1000
   },
   fast: {
     qualityProfile: "general",
-    semanticVetoMaxMs: 0,
-    semanticVetoRecheckMs: DEFAULT_SEMANTIC_VETO_RECHECK_MS,
+    minSilenceMs: 300,
     silenceMs: 700,
     speechThreshold: 0.015,
     transcriptStabilityMs: 450
   },
   "long-form": {
     qualityProfile: "general",
-    semanticVetoMaxMs: 0,
-    semanticVetoRecheckMs: DEFAULT_SEMANTIC_VETO_RECHECK_MS,
+    minSilenceMs: 600,
     silenceMs: 2200,
     speechThreshold: 0.01,
     transcriptStabilityMs: 1500
@@ -1579,12 +1458,12 @@ var resolveTurnDetectionConfig = (config) => {
   const qualityProfile = config?.qualityProfile ?? DEFAULT_QUALITY_PROFILE;
   const preset = TURN_PROFILE_DEFAULTS[profile];
   const quality = QUALITY_PROFILE_DEFAULTS[qualityProfile];
+  const silenceMs = config?.silenceMs ?? quality.silenceMs ?? preset.silenceMs;
   return {
     profile,
     qualityProfile,
-    semanticVetoMaxMs: config?.semanticVetoMaxMs ?? preset.semanticVetoMaxMs,
-    semanticVetoRecheckMs: config?.semanticVetoRecheckMs ?? preset.semanticVetoRecheckMs,
-    silenceMs: config?.silenceMs ?? quality.silenceMs ?? preset.silenceMs,
+    minSilenceMs: Math.min(silenceMs, config?.minSilenceMs ?? quality.minSilenceMs ?? preset.minSilenceMs),
+    silenceMs,
     speechThreshold: config?.speechThreshold ?? quality.speechThreshold ?? preset.speechThreshold,
     transcriptStabilityMs: config?.transcriptStabilityMs ?? quality.transcriptStabilityMs ?? preset.transcriptStabilityMs
   };

package/dist/testing/index.js

@@ -86,7 +86,7 @@ var __require = import.meta.require;
 // src/core/turnDetection.ts
 var DEFAULT_SILENCE_MS = 700;
 var DEFAULT_SPEECH_THRESHOLD = 0.015;
-var DEFAULT_SEMANTIC_VETO_RECHECK_MS = 1200;
+var DEFAULT_MIN_SILENCE_MS = 400;
 var toUint8Array = (audio) => {
   if (audio instanceof ArrayBuffer) {
     return new Uint8Array(audio);
@@ -3163,24 +3163,21 @@ var resolveAudioConditioningConfig = (config) => {
 var TURN_PROFILE_DEFAULTS = {
   balanced: {
     qualityProfile: "general",
-    semanticVetoMaxMs: 0,
-    semanticVetoRecheckMs: DEFAULT_SEMANTIC_VETO_RECHECK_MS,
+    minSilenceMs: 400,
     silenceMs: 1400,
     speechThreshold: 0.012,
     transcriptStabilityMs: 1000
   },
   fast: {
     qualityProfile: "general",
-    semanticVetoMaxMs: 0,
-    semanticVetoRecheckMs: DEFAULT_SEMANTIC_VETO_RECHECK_MS,
+    minSilenceMs: 300,
     silenceMs: 700,
     speechThreshold: 0.015,
     transcriptStabilityMs: 450
   },
   "long-form": {
     qualityProfile: "general",
-    semanticVetoMaxMs: 0,
-    semanticVetoRecheckMs: DEFAULT_SEMANTIC_VETO_RECHECK_MS,
+    minSilenceMs: 600,
     silenceMs: 2200,
     speechThreshold: 0.01,
     transcriptStabilityMs: 1500
@@ -3211,12 +3208,12 @@ var resolveTurnDetectionConfig = (config) => {
   const qualityProfile = config?.qualityProfile ?? DEFAULT_QUALITY_PROFILE;
   const preset = TURN_PROFILE_DEFAULTS[profile];
   const quality = QUALITY_PROFILE_DEFAULTS[qualityProfile];
+  const silenceMs = config?.silenceMs ?? quality.silenceMs ?? preset.silenceMs;
   return {
     profile,
     qualityProfile,
-    semanticVetoMaxMs: config?.semanticVetoMaxMs ?? preset.semanticVetoMaxMs,
-    semanticVetoRecheckMs: config?.semanticVetoRecheckMs ?? preset.semanticVetoRecheckMs,
-    silenceMs: config?.silenceMs ?? quality.silenceMs ?? preset.silenceMs,
+    minSilenceMs: Math.min(silenceMs, config?.minSilenceMs ?? quality.minSilenceMs ?? preset.minSilenceMs),
+    silenceMs,
     speechThreshold: config?.speechThreshold ?? quality.speechThreshold ?? preset.speechThreshold,
     transcriptStabilityMs: config?.transcriptStabilityMs ?? quality.transcriptStabilityMs ?? preset.transcriptStabilityMs
   };
@@ -6153,14 +6150,22 @@ var createVoiceSession = (options) => {
     strategy: options.reconnect.strategy ?? "resume-last-turn",
     timeout: options.reconnect.timeout ?? DEFAULT_RECONNECT_TIMEOUT
   };
+  const resolvedSilenceMs = options.turnDetection.silenceMs ?? DEFAULT_SILENCE_MS;
   const turnDetection = {
-    silenceMs: options.turnDetection.silenceMs ?? DEFAULT_SILENCE_MS,
+    silenceMs: resolvedSilenceMs,
+    minSilenceMs: Math.min(resolvedSilenceMs, options.turnDetection.minSilenceMs ?? DEFAULT_MIN_SILENCE_MS),
     speechThreshold: options.turnDetection.speechThreshold ?? DEFAULT_SPEECH_THRESHOLD,
-    transcriptStabilityMs: options.turnDetection.transcriptStabilityMs ?? DEFAULT_TRANSCRIPT_STABILITY_MS,
-    semanticVetoMaxMs: options.turnDetection.semanticVetoMaxMs ?? 0,
-    semanticVetoRecheckMs: options.turnDetection.semanticVetoRecheckMs ?? DEFAULT_SEMANTIC_VETO_RECHECK_MS
+    transcriptStabilityMs: options.turnDetection.transcriptStabilityMs ?? DEFAULT_TRANSCRIPT_STABILITY_MS
+  };
+  let lastTurnCompleteConfidence = null;
+  const adaptiveSilenceMs = () => {
+    const { minSilenceMs, silenceMs } = turnDetection;
+    if (lastTurnCompleteConfidence === null || silenceMs <= minSilenceMs) {
+      return silenceMs;
+    }
+    const complete = Math.max(0, Math.min(1, lastTurnCompleteConfidence));
+    return Math.round(minSilenceMs + (silenceMs - minSilenceMs) * (1 - complete));
   };
-  let semanticVetoElapsedMs = 0;
   const sttFallback = options.sttFallback ? {
     adapter: options.sttFallback.adapter,
     completionTimeoutMs: options.sttFallback.completionTimeoutMs ?? DEFAULT_FALLBACK_COMPLETION_TIMEOUT_MS,
@@ -6693,47 +6698,8 @@ var createVoiceSession = (options) => {
       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,
-        ...getTurnAudioForDetector()
-      }));
-      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 scheduleSilenceCommit = (delayMs = adaptiveSilenceMs(), reset = true) => scheduleTurnCommit(delayMs, "silence", reset);
   const runScheduledCommit = async (reason) => {
-    if (await shouldDeferSilenceCommit(reason)) {
-      return;
-    }
     await api.commitTurn(reason);
   };
   const requestTurnCommit = async (reason) => {
@@ -7473,7 +7439,7 @@ var createVoiceSession = (options) => {
       session2.lastActivityAt = Date.now();
       session2.status = "active";
     });
-    semanticVetoElapsedMs = 0;
+    lastTurnCompleteConfidence = null;
     if (silenceTimer && pendingCommitReason === "vendor") {
       scheduleTurnCommit(getVendorCommitDelayMs(), "vendor");
     }
@@ -7503,6 +7469,12 @@ var createVoiceSession = (options) => {
         transcripts: session.currentTurn.transcripts,
         ...getTurnAudioForDetector()
       }));
+      if (typeof verdict.confidence === "number") {
+        lastTurnCompleteConfidence = verdict.confidence;
+        if (silenceTimer && pendingCommitReason === "silence") {
+          scheduleSilenceCommit();
+        }
+      }
       if (verdict.endOfTurn) {
         clearSilenceTimer();
         await requestTurnCommit("vendor");
@@ -8198,7 +8170,7 @@ var createVoiceSession = (options) => {
   };
   const commitTurnInternal = async (reason = "manual") => {
     clearSilenceTimer();
-    semanticVetoElapsedMs = 0;
+    lastTurnCompleteConfidence = null;
     backchannelDriver?.reset();
     amdLastTurnCommitAt = Date.now();
     const session = await readSession();

package/dist/vue/index.js

@@ -11689,146 +11689,25 @@ 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,
+    minSilenceMs: 400,
     silenceMs: 1400,
     speechThreshold: 0.012,
     transcriptStabilityMs: 1000
   },
   fast: {
     qualityProfile: "general",
-    semanticVetoMaxMs: 0,
-    semanticVetoRecheckMs: DEFAULT_SEMANTIC_VETO_RECHECK_MS,
+    minSilenceMs: 300,
     silenceMs: 700,
     speechThreshold: 0.015,
     transcriptStabilityMs: 450
   },
   "long-form": {
     qualityProfile: "general",
-    semanticVetoMaxMs: 0,
-    semanticVetoRecheckMs: DEFAULT_SEMANTIC_VETO_RECHECK_MS,
+    minSilenceMs: 600,
     silenceMs: 2200,
     speechThreshold: 0.01,
     transcriptStabilityMs: 1500
@@ -11859,12 +11738,12 @@ var resolveTurnDetectionConfig = (config) => {
   const qualityProfile = config?.qualityProfile ?? DEFAULT_QUALITY_PROFILE;
   const preset = TURN_PROFILE_DEFAULTS[profile];
   const quality = QUALITY_PROFILE_DEFAULTS[qualityProfile];
+  const silenceMs = config?.silenceMs ?? quality.silenceMs ?? preset.silenceMs;
   return {
     profile,
     qualityProfile,
-    semanticVetoMaxMs: config?.semanticVetoMaxMs ?? preset.semanticVetoMaxMs,
-    semanticVetoRecheckMs: config?.semanticVetoRecheckMs ?? preset.semanticVetoRecheckMs,
-    silenceMs: config?.silenceMs ?? quality.silenceMs ?? preset.silenceMs,
+    minSilenceMs: Math.min(silenceMs, config?.minSilenceMs ?? quality.minSilenceMs ?? preset.minSilenceMs),
+    silenceMs,
     speechThreshold: config?.speechThreshold ?? quality.speechThreshold ?? preset.speechThreshold,
     transcriptStabilityMs: config?.transcriptStabilityMs ?? quality.transcriptStabilityMs ?? preset.transcriptStabilityMs
   };

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@absolutejs/voice",
-  "version": "0.0.22-beta.598",
+  "version": "0.0.22-beta.599",
   "description": "Voice primitives and Elysia plugin for AbsoluteJS",
   "repository": {
     "type": "git",