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

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
@@ -86,6 +86,7 @@ var __require = import.meta.require;
86
86
  // src/core/turnDetection.ts
87
87
  var DEFAULT_SILENCE_MS = 700;
88
88
  var DEFAULT_SPEECH_THRESHOLD = 0.015;
89
+ var DEFAULT_SEMANTIC_VETO_RECHECK_MS = 1200;
89
90
  var toUint8Array = (audio) => {
90
91
  if (audio instanceof ArrayBuffer) {
91
92
  return new Uint8Array(audio);
@@ -3133,18 +3134,24 @@ var resolveAudioConditioningConfig = (config) => {
3133
3134
  var TURN_PROFILE_DEFAULTS = {
3134
3135
  balanced: {
3135
3136
  qualityProfile: "general",
3137
+ semanticVetoMaxMs: 0,
3138
+ semanticVetoRecheckMs: DEFAULT_SEMANTIC_VETO_RECHECK_MS,
3136
3139
  silenceMs: 1400,
3137
3140
  speechThreshold: 0.012,
3138
3141
  transcriptStabilityMs: 1000
3139
3142
  },
3140
3143
  fast: {
3141
3144
  qualityProfile: "general",
3145
+ semanticVetoMaxMs: 0,
3146
+ semanticVetoRecheckMs: DEFAULT_SEMANTIC_VETO_RECHECK_MS,
3142
3147
  silenceMs: 700,
3143
3148
  speechThreshold: 0.015,
3144
3149
  transcriptStabilityMs: 450
3145
3150
  },
3146
3151
  "long-form": {
3147
3152
  qualityProfile: "general",
3153
+ semanticVetoMaxMs: 0,
3154
+ semanticVetoRecheckMs: DEFAULT_SEMANTIC_VETO_RECHECK_MS,
3148
3155
  silenceMs: 2200,
3149
3156
  speechThreshold: 0.01,
3150
3157
  transcriptStabilityMs: 1500
@@ -3178,6 +3185,8 @@ var resolveTurnDetectionConfig = (config) => {
3178
3185
  return {
3179
3186
  profile,
3180
3187
  qualityProfile,
3188
+ semanticVetoMaxMs: config?.semanticVetoMaxMs ?? preset.semanticVetoMaxMs,
3189
+ semanticVetoRecheckMs: config?.semanticVetoRecheckMs ?? preset.semanticVetoRecheckMs,
3181
3190
  silenceMs: config?.silenceMs ?? quality.silenceMs ?? preset.silenceMs,
3182
3191
  speechThreshold: config?.speechThreshold ?? quality.speechThreshold ?? preset.speechThreshold,
3183
3192
  transcriptStabilityMs: config?.transcriptStabilityMs ?? quality.transcriptStabilityMs ?? preset.transcriptStabilityMs
@@ -6105,8 +6114,11 @@ var createVoiceSession = (options) => {
6105
6114
  const turnDetection = {
6106
6115
  silenceMs: options.turnDetection.silenceMs ?? DEFAULT_SILENCE_MS,
6107
6116
  speechThreshold: options.turnDetection.speechThreshold ?? DEFAULT_SPEECH_THRESHOLD,
6108
- transcriptStabilityMs: options.turnDetection.transcriptStabilityMs ?? DEFAULT_TRANSCRIPT_STABILITY_MS
6117
+ transcriptStabilityMs: options.turnDetection.transcriptStabilityMs ?? DEFAULT_TRANSCRIPT_STABILITY_MS,
6118
+ semanticVetoMaxMs: options.turnDetection.semanticVetoMaxMs ?? 0,
6119
+ semanticVetoRecheckMs: options.turnDetection.semanticVetoRecheckMs ?? DEFAULT_SEMANTIC_VETO_RECHECK_MS
6109
6120
  };
6121
+ let semanticVetoElapsedMs = 0;
6110
6122
  const sttFallback = options.sttFallback ? {
6111
6123
  adapter: options.sttFallback.adapter,
6112
6124
  completionTimeoutMs: options.sttFallback.completionTimeoutMs ?? DEFAULT_FALLBACK_COMPLETION_TIMEOUT_MS,
@@ -6621,10 +6633,51 @@ var createVoiceSession = (options) => {
6621
6633
  silenceTimer = setTimeout(() => {
6622
6634
  silenceTimer = null;
6623
6635
  pendingCommitReason = null;
6624
- api.commitTurn(reason);
6636
+ runScheduledCommit(reason);
6625
6637
  }, delayMs);
6626
6638
  };
6627
6639
  const scheduleSilenceCommit = (delayMs = turnDetection.silenceMs, reset = true) => scheduleTurnCommit(delayMs, "silence", reset);
6640
+ const shouldDeferSilenceCommit = async (reason) => {
6641
+ if (reason !== "silence" || turnDetection.semanticVetoMaxMs <= 0 || !options.semanticTurnDetector || semanticVetoElapsedMs >= turnDetection.semanticVetoMaxMs) {
6642
+ return false;
6643
+ }
6644
+ const session = await readSession();
6645
+ const { partialText, transcripts } = session.currentTurn;
6646
+ const userText = buildTurnText(transcripts, partialText, {
6647
+ partialEndedAtMs: session.currentTurn.partialEndedAt,
6648
+ partialStartedAtMs: session.currentTurn.partialStartedAt
6649
+ });
6650
+ if (!userText) {
6651
+ return false;
6652
+ }
6653
+ const silenceMs = session.currentTurn.silenceStartedAt !== undefined ? Date.now() - session.currentTurn.silenceStartedAt : turnDetection.silenceMs;
6654
+ let endOfTurn = true;
6655
+ try {
6656
+ const verdict = await Promise.resolve(options.semanticTurnDetector.evaluate({
6657
+ lastFinalTranscript: transcripts.at(-1),
6658
+ partialText,
6659
+ silenceMs,
6660
+ transcripts
6661
+ }));
6662
+ endOfTurn = verdict.endOfTurn;
6663
+ } catch {
6664
+ return false;
6665
+ }
6666
+ if (endOfTurn !== false) {
6667
+ return false;
6668
+ }
6669
+ const remaining = turnDetection.semanticVetoMaxMs - semanticVetoElapsedMs;
6670
+ const extendMs = Math.max(1, Math.min(turnDetection.semanticVetoRecheckMs, remaining));
6671
+ semanticVetoElapsedMs += extendMs;
6672
+ scheduleTurnCommit(extendMs, reason);
6673
+ return true;
6674
+ };
6675
+ const runScheduledCommit = async (reason) => {
6676
+ if (await shouldDeferSilenceCommit(reason)) {
6677
+ return;
6678
+ }
6679
+ await api.commitTurn(reason);
6680
+ };
6628
6681
  const requestTurnCommit = async (reason) => {
6629
6682
  const session = await readSession();
6630
6683
  const text = buildTurnText(session.currentTurn.transcripts, session.currentTurn.partialText, {
@@ -7336,6 +7389,7 @@ var createVoiceSession = (options) => {
7336
7389
  session2.lastActivityAt = Date.now();
7337
7390
  session2.status = "active";
7338
7391
  });
7392
+ semanticVetoElapsedMs = 0;
7339
7393
  if (silenceTimer && pendingCommitReason === "vendor") {
7340
7394
  scheduleTurnCommit(getVendorCommitDelayMs(), "vendor");
7341
7395
  }
@@ -8039,6 +8093,7 @@ var createVoiceSession = (options) => {
8039
8093
  };
8040
8094
  const commitTurnInternal = async (reason = "manual") => {
8041
8095
  clearSilenceTimer();
8096
+ semanticVetoElapsedMs = 0;
8042
8097
  backchannelDriver?.reset();
8043
8098
  amdLastTurnCommitAt = Date.now();
8044
8099
  const session = await readSession();
package/dist/vue/index.js CHANGED
@@ -11660,22 +11660,146 @@ var resolveAudioConditioningConfig = (config) => {
11660
11660
  };
11661
11661
  };
11662
11662
 
11663
+ // src/core/turnDetection.ts
11664
+ var DEFAULT_SILENCE_MS = 700;
11665
+ var DEFAULT_SPEECH_THRESHOLD = 0.015;
11666
+ var DEFAULT_SEMANTIC_VETO_RECHECK_MS = 1200;
11667
+ var toUint8Array = (audio) => {
11668
+ if (audio instanceof ArrayBuffer) {
11669
+ return new Uint8Array(audio);
11670
+ }
11671
+ return new Uint8Array(audio.buffer, audio.byteOffset, audio.byteLength);
11672
+ };
11673
+ var measureAudioLevel = (audio) => {
11674
+ const bytes = toUint8Array(audio);
11675
+ if (bytes.byteLength < 2) {
11676
+ return 0;
11677
+ }
11678
+ const samples = new Int16Array(bytes.buffer, bytes.byteOffset, Math.floor(bytes.byteLength / 2));
11679
+ if (samples.length === 0) {
11680
+ return 0;
11681
+ }
11682
+ let sumSquares = 0;
11683
+ for (const sample of samples) {
11684
+ const normalized = sample / 32768;
11685
+ sumSquares += normalized * normalized;
11686
+ }
11687
+ return Math.sqrt(sumSquares / samples.length);
11688
+ };
11689
+ var normalizeText = (value) => value.trim().replace(/\s+/g, " ");
11690
+ var countWords = (value) => value.length > 0 ? value.split(" ").length : 0;
11691
+ var selectPreferredTranscriptText = (currentText, nextText) => {
11692
+ const current = normalizeText(currentText);
11693
+ const next = normalizeText(nextText);
11694
+ if (!current) {
11695
+ return next;
11696
+ }
11697
+ if (!next) {
11698
+ return current;
11699
+ }
11700
+ if (current === next || current.includes(next)) {
11701
+ return current;
11702
+ }
11703
+ if (next.includes(current)) {
11704
+ return next;
11705
+ }
11706
+ if (countWords(next) > countWords(current)) {
11707
+ return next;
11708
+ }
11709
+ if (countWords(next) === countWords(current) && next.length > current.length) {
11710
+ return next;
11711
+ }
11712
+ return current;
11713
+ };
11714
+ var mergeSequentialTranscriptText = (currentText, nextText) => {
11715
+ const current = normalizeText(currentText);
11716
+ const next = normalizeText(nextText);
11717
+ if (!current) {
11718
+ return next;
11719
+ }
11720
+ if (!next) {
11721
+ return current;
11722
+ }
11723
+ const currentWords = current.split(" ");
11724
+ const nextWords = next.split(" ");
11725
+ const maxOverlap = Math.min(currentWords.length, nextWords.length);
11726
+ for (let overlap = maxOverlap;overlap > 0; overlap -= 1) {
11727
+ const currentSuffix = currentWords.slice(-overlap).join(" ");
11728
+ const nextPrefix = nextWords.slice(0, overlap).join(" ");
11729
+ if (currentSuffix === nextPrefix) {
11730
+ return [...currentWords, ...nextWords.slice(overlap)].join(" ");
11731
+ }
11732
+ }
11733
+ return `${current} ${next}`.trim();
11734
+ };
11735
+ var countCommonPrefixWords = (currentText, nextText) => {
11736
+ const currentWords = normalizeText(currentText).split(" ").filter(Boolean);
11737
+ const nextWords = normalizeText(nextText).split(" ").filter(Boolean);
11738
+ const maxWords = Math.min(currentWords.length, nextWords.length);
11739
+ let count = 0;
11740
+ for (let index = 0;index < maxWords; index += 1) {
11741
+ if (currentWords[index] !== nextWords[index]) {
11742
+ break;
11743
+ }
11744
+ count += 1;
11745
+ }
11746
+ return count;
11747
+ };
11748
+ var mergeTranscriptTexts = (transcripts) => {
11749
+ const merged = [];
11750
+ for (const transcript of transcripts) {
11751
+ const nextText = normalizeText(transcript.text);
11752
+ if (!nextText) {
11753
+ continue;
11754
+ }
11755
+ const previous = merged.at(-1);
11756
+ if (!previous) {
11757
+ merged.push(nextText);
11758
+ continue;
11759
+ }
11760
+ if (nextText === previous || previous.includes(nextText)) {
11761
+ continue;
11762
+ }
11763
+ if (nextText.includes(previous)) {
11764
+ merged[merged.length - 1] = nextText;
11765
+ continue;
11766
+ }
11767
+ merged.push(nextText);
11768
+ }
11769
+ return merged.join(" ").trim();
11770
+ };
11771
+ var buildTurnText = (transcripts, partialText, options = {}) => {
11772
+ const finalText = mergeTranscriptTexts(transcripts);
11773
+ const nextPartial = normalizeText(partialText);
11774
+ const lastFinalEndedAtMs = [...transcripts].reverse().find((transcript) => typeof transcript.endedAtMs === "number")?.endedAtMs;
11775
+ if (finalText && nextPartial && typeof lastFinalEndedAtMs === "number" && typeof options.partialStartedAtMs === "number" && options.partialStartedAtMs - lastFinalEndedAtMs >= 250 && countCommonPrefixWords(finalText, nextPartial) === 0) {
11776
+ return mergeSequentialTranscriptText(finalText, nextPartial);
11777
+ }
11778
+ return selectPreferredTranscriptText(finalText, nextPartial);
11779
+ };
11780
+
11663
11781
  // src/core/turnProfiles.ts
11664
11782
  var TURN_PROFILE_DEFAULTS = {
11665
11783
  balanced: {
11666
11784
  qualityProfile: "general",
11785
+ semanticVetoMaxMs: 0,
11786
+ semanticVetoRecheckMs: DEFAULT_SEMANTIC_VETO_RECHECK_MS,
11667
11787
  silenceMs: 1400,
11668
11788
  speechThreshold: 0.012,
11669
11789
  transcriptStabilityMs: 1000
11670
11790
  },
11671
11791
  fast: {
11672
11792
  qualityProfile: "general",
11793
+ semanticVetoMaxMs: 0,
11794
+ semanticVetoRecheckMs: DEFAULT_SEMANTIC_VETO_RECHECK_MS,
11673
11795
  silenceMs: 700,
11674
11796
  speechThreshold: 0.015,
11675
11797
  transcriptStabilityMs: 450
11676
11798
  },
11677
11799
  "long-form": {
11678
11800
  qualityProfile: "general",
11801
+ semanticVetoMaxMs: 0,
11802
+ semanticVetoRecheckMs: DEFAULT_SEMANTIC_VETO_RECHECK_MS,
11679
11803
  silenceMs: 2200,
11680
11804
  speechThreshold: 0.01,
11681
11805
  transcriptStabilityMs: 1500
@@ -11709,6 +11833,8 @@ var resolveTurnDetectionConfig = (config) => {
11709
11833
  return {
11710
11834
  profile,
11711
11835
  qualityProfile,
11836
+ semanticVetoMaxMs: config?.semanticVetoMaxMs ?? preset.semanticVetoMaxMs,
11837
+ semanticVetoRecheckMs: config?.semanticVetoRecheckMs ?? preset.semanticVetoRecheckMs,
11712
11838
  silenceMs: config?.silenceMs ?? quality.silenceMs ?? preset.silenceMs,
11713
11839
  speechThreshold: config?.speechThreshold ?? quality.speechThreshold ?? preset.speechThreshold,
11714
11840
  transcriptStabilityMs: config?.transcriptStabilityMs ?? quality.transcriptStabilityMs ?? preset.transcriptStabilityMs
package/package.json CHANGED
@@ -1,6 +1,6 @@
1
1
  {
2
2
  "name": "@absolutejs/voice",
3
- "version": "0.0.22-beta.584",
3
+ "version": "0.0.22-beta.585",
4
4
  "description": "Voice primitives and Elysia plugin for AbsoluteJS",
5
5
  "repository": {
6
6
  "type": "git",