@absolutejs/voice 0.0.22-beta.597 → 0.0.22-beta.599

package/dist/angular/index.js

@@ -1227,146 +1227,25 @@ 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
@@ -1397,12 +1276,12 @@ 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/client/htmxBootstrap.js

@@ -1107,31 +1107,25 @@ var resolveAudioConditioningConfig = (config) => {
   };
 };
 
-// src/core/turnDetection.ts
-var DEFAULT_SEMANTIC_VETO_RECHECK_MS = 1200;
-
 // 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
@@ -1162,12 +1156,12 @@ 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/client/index.js

@@ -1678,146 +1678,25 @@ 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
@@ -1848,12 +1727,12 @@ 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/core/semanticTurn.d.ts

@@ -1,10 +1,20 @@
-import type { Transcript } from "./types";
+import type { AudioFormat, Transcript } from "./types";
 export type VoiceSemanticTurnInput = {
     audioLevel?: number;
     lastFinalTranscript?: Transcript;
     partialText: string;
     silenceMs: number;
     transcripts: Transcript[];
+    /**
+     * The current turn's buffered user audio (PCM chunks, oldest→newest) and its
+     * format. Lets an AUDIO-based end-of-turn detector (e.g. a smart-turn / Whisper
+     * EOT model) judge completion from prosody — pitch, pace, trailing intonation —
+     * which a transcript-only judge fundamentally cannot see. Undefined when no
+     * audio was buffered for the turn (the runtime only stores chunks above the
+     * speech threshold).
+     */
+    turnAudio?: ReadonlyArray<Uint8Array>;
+    turnAudioFormat?: AudioFormat;
 };
 export type VoiceSemanticTurnVerdict = {
     confidence?: number;

package/dist/core/turnDetection.d.ts

@@ -1,7 +1,7 @@
 import type { AudioChunk, Transcript } from "./types";
 export declare const DEFAULT_SILENCE_MS = 700;
 export declare const DEFAULT_SPEECH_THRESHOLD = 0.015;
-export declare const DEFAULT_SEMANTIC_VETO_RECHECK_MS = 1200;
+export declare const DEFAULT_MIN_SILENCE_MS = 400;
 export declare const measureAudioLevel: (audio: AudioChunk) => number;
 export declare const selectPreferredTranscriptText: (currentText: string, nextText: string) => string;
 export declare const buildTurnText: (transcripts: Transcript[], partialText: string, options?: {

package/dist/core/types.d.ts

@@ -436,19 +436,17 @@ export type VoiceTurnDetectionConfig = {
     profile?: VoiceTurnProfile;
     qualityProfile?: VoiceTurnQualityProfile;
     silenceMs?: number;
+    minSilenceMs?: number;
     speechThreshold?: number;
     transcriptStabilityMs?: number;
-    semanticVetoMaxMs?: number;
-    semanticVetoRecheckMs?: number;
 };
 export type VoiceResolvedTurnDetectionConfig = {
     qualityProfile: VoiceTurnQualityProfile;
     profile: VoiceTurnProfile;
     silenceMs: number;
+    minSilenceMs: number;
     speechThreshold: number;
     transcriptStabilityMs: number;
-    semanticVetoMaxMs: number;
-    semanticVetoRecheckMs: number;
 };
 export type VoiceAudioConditioningConfig = {
     enabled?: boolean;

package/dist/embed/index.js

@@ -1104,31 +1104,25 @@ var resolveAudioConditioningConfig = (config) => {
   };
 };
 
-// src/core/turnDetection.ts
-var DEFAULT_SEMANTIC_VETO_RECHECK_MS = 1200;
-
 // 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
@@ -1159,12 +1153,12 @@ 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
   };