@kuralle-syrinx/pipecat-smart-turn 4.2.0 → 4.4.0

package/src/index.ts

@@ -5,450 +5,19 @@
 // Mirrors Pipecat's LocalSmartTurnAnalyzerV3 turn-stop strategy:
 // Silero determines candidate speech boundaries, then Smart Turn v3 decides
 // whether a pause is an actual completed user turn.
+//
+// Node / local default: `new PipecatEOSPlugin()` uses LocalSmartTurnV3Predictor
+// (onnxruntime-node). Edge hosts must import from `@kuralle-syrinx/pipecat-smart-turn/eos`
+// and inject a non-ONNX predictor (e.g. WorkersAiSmartTurnPredictor).
 
-import {
-  fuseEndpointDecision,
-  latestTranscript,
-  scoreSemanticCompleteness,
-  type SemanticEndpointFusionConfig,
-} from "./semantic-completeness.js";
-import {
-  Route,
-  type EndOfSpeechPacket,
-  type FinalizeSttPacket,
-  type InterruptionDetectedPacket,
-  type InterimEndOfSpeechPacket,
-  type PipelineBus,
-  type PluginConfig,
-  type SttInterimPacket,
-  type SttResultPacket,
-  type TextToSpeechAudioPacket,
-  type TextToSpeechEndPacket,
-  type TextToSpeechPlayoutProgressPacket,
-  type VadAudioPacket,
-  type VadSpeechEndedPacket,
-  type VadSpeechStartedPacket,
-  type VoicePlugin,
-} from "@kuralle-syrinx/core";
-import { pcm16BytesToSamples } from "@kuralle-syrinx/core/audio";
+import { PipecatEOSPlugin as PipecatEOSPluginBase } from "./eos-plugin.js";
 import { LocalSmartTurnV3Predictor, type SmartTurnPredictor } from "./predictor.js";
 
-const SAMPLE_RATE = 16000;
-const MAX_AUDIO_SAMPLES = SAMPLE_RATE * 8;
-
-interface TurnState {
-  readonly contextId: string;
-  audio: number[];
-  finalPackets: SttResultPacket[];
-  finalSegments: string[];
-  latestInterim: string;
-  boundaryAnalyzed: boolean;
-  smartTurnComplete: boolean;
-  semanticComplete: boolean;
-  speechActive: boolean;
-  finalized: boolean;
-  analysisSequence: number;
-  finalizeTimer: ReturnType<typeof setTimeout> | null;
-  sttQuietTimer: ReturnType<typeof setTimeout> | null;
-  maxTimer: ReturnType<typeof setTimeout> | null;
-  deferTimer: ReturnType<typeof setTimeout> | null;
-}
-
-export class PipecatEOSPlugin implements VoicePlugin {
-  readonly endpointingCapability = { owner: "smart_turn" as const };
-
-  private bus: PipelineBus | null = null;
-  private disposers: Array<() => void> = [];
-  private turns = new Map<string, TurnState>();
-  private finalizeDelayMs = 250;
-  private sttQuietFallbackMs = 2500;
-  private maxDelayMs = 2000;
-  private incompleteFallbackMs = 2000;
-  private semanticShortcutDelayMs = 50;
-  private semanticDeferFallbackMs = 4000;
-  private semanticEndpointingEnabled = true;
-  private probabilityThreshold = 0.5;
-  private readonly lockedContextIds = new Set<string>();
-  private readonly lockedContextsWithAssistantAudio = new Set<string>();
-
-  constructor(private readonly predictor: SmartTurnPredictor = new LocalSmartTurnV3Predictor()) {}
-
-  async initialize(bus: PipelineBus, config: PluginConfig): Promise<void> {
-    this.bus = bus;
-    this.finalizeDelayMs = readNonNegativeNumber(config["finalize_delay_ms"], 250);
-    this.sttQuietFallbackMs = readNonNegativeNumber(config["stt_quiet_fallback_ms"], 2500);
-    this.maxDelayMs = readNonNegativeNumber(config["max_delay_ms"], 2000);
-    this.incompleteFallbackMs = readNonNegativeNumber(config["incomplete_fallback_ms"], 2000);
-    this.semanticShortcutDelayMs = readNonNegativeNumber(config["semantic_shortcut_delay_ms"], 50);
-    this.semanticDeferFallbackMs = readNonNegativeNumber(config["semantic_defer_fallback_ms"], 4000);
-    this.semanticEndpointingEnabled = readBooleanConfig(config["semantic_endpointing_enabled"], true);
-    this.probabilityThreshold = readProbability(config["probability_threshold"], 0.5);
-    await this.predictor.initialize(config);
-
-    this.disposers.push(
-      bus.on("vad.audio", (pkt) => {
-        this.handleAudio(pkt as VadAudioPacket);
-      }),
-      bus.on("stt.interim", (pkt) => {
-        this.handleInterim(pkt as SttInterimPacket);
-      }),
-      bus.on("stt.result", (pkt) => {
-        this.handleFinal(pkt as SttResultPacket);
-      }),
-      bus.on("vad.speech_started", (pkt) => {
-        this.handleSpeechStarted(pkt as VadSpeechStartedPacket);
-      }),
-      bus.on("vad.speech_ended", async (pkt) => {
-        await this.handleSpeechEnded(pkt as VadSpeechEndedPacket);
-      }),
-      bus.on("tts.audio", (pkt) => {
-        this.handleTtsAudio(pkt as TextToSpeechAudioPacket);
-      }),
-      bus.on("tts.end", (pkt) => {
-        this.handleTtsEnd(pkt as TextToSpeechEndPacket);
-      }),
-      bus.on("tts.playout_progress", (pkt) => {
-        this.handleTtsPlayoutProgress(pkt as TextToSpeechPlayoutProgressPacket);
-      }),
-      bus.on("interrupt.detected", (pkt) => {
-        this.releaseContextLock((pkt as InterruptionDetectedPacket).contextId);
-      }),
-    );
-  }
-
-  async close(): Promise<void> {
-    for (const dispose of this.disposers.splice(0)) {
-      dispose();
-    }
-    for (const state of this.turns.values()) {
-      clearTurnTimers(state);
-    }
-    this.turns.clear();
-    this.lockedContextIds.clear();
-    this.lockedContextsWithAssistantAudio.clear();
-    await this.predictor.close();
-    this.bus = null;
-  }
-
-  private handleAudio(pkt: VadAudioPacket): void {
-    if (this.lockedContextIds.has(pkt.contextId)) return;
-    if (pkt.audio.byteLength % 2 !== 0) return;
-    const state = this.stateFor(pkt.contextId);
-    const samples = pcm16BytesToSamples(pkt.audio);
-    for (const sample of samples) {
-      state.audio.push(sample / 32768);
-    }
-    if (state.audio.length > MAX_AUDIO_SAMPLES) {
-      state.audio.splice(0, state.audio.length - MAX_AUDIO_SAMPLES);
-    }
-  }
-
-  private handleInterim(pkt: SttInterimPacket): void {
-    if (this.lockedContextIds.has(pkt.contextId)) return;
-    if (!pkt.text.trim()) return;
-    const state = this.stateFor(pkt.contextId);
-    state.latestInterim = pkt.text.trim();
-    if (state.sttQuietTimer) this.armSttQuietFallback(state);
-    this.bus?.push(Route.Main, {
-      kind: "eos.interim",
-      contextId: pkt.contextId,
-      timestampMs: Date.now(),
-      text: pkt.text,
-    } satisfies InterimEndOfSpeechPacket);
+/** Node-facing EOS plugin: defaults to the local ONNX Smart Turn predictor. */
+export class PipecatEOSPlugin extends PipecatEOSPluginBase {
+  constructor(predictor: SmartTurnPredictor = new LocalSmartTurnV3Predictor()) {
+    super(predictor);
   }
-
-  private handleFinal(pkt: SttResultPacket): void {
-    if (this.lockedContextIds.has(pkt.contextId)) return;
-    if (!pkt.text.trim()) return;
-    const state = this.stateFor(pkt.contextId);
-    appendFinalPacket(state, pkt);
-    state.latestInterim = "";
-
-    const transcript = latestTranscript(state.finalSegments, state.latestInterim);
-    const semantic = scoreSemanticCompleteness(transcript);
-    state.semanticComplete = semantic.complete;
-
-    if (state.boundaryAnalyzed && state.smartTurnComplete && state.semanticComplete) {
-      this.scheduleFinalize(state, this.finalizeDelayMs);
-      return;
-    }
-    if (state.smartTurnComplete && state.semanticComplete) {
-      this.scheduleFinalize(state, this.finalizeDelayMs);
-      return;
-    }
-    if (state.boundaryAnalyzed && state.smartTurnComplete && !state.semanticComplete) {
-      return;
-    }
-    if (state.boundaryAnalyzed && !state.smartTurnComplete && state.semanticComplete) {
-      this.scheduleSemanticShortcut(state);
-      return;
-    }
-    if (state.boundaryAnalyzed) {
-      this.scheduleIncompleteFallback(state);
-      return;
-    }
-    if (state.speechActive) {
-      this.armSttQuietFallback(state);
-      return;
-    }
-    this.scheduleMaxFinalize(state);
-  }
-
-  private handleSpeechStarted(pkt: VadSpeechStartedPacket): void {
-    if (this.lockedContextIds.has(pkt.contextId)) return;
-    const state = this.stateFor(pkt.contextId);
-    if (state.sttQuietTimer) {
-      clearTimeout(state.sttQuietTimer);
-      state.sttQuietTimer = null;
-    }
-    state.boundaryAnalyzed = false;
-    state.smartTurnComplete = false;
-    state.semanticComplete = false;
-    state.speechActive = true;
-    state.latestInterim = "";
-    state.analysisSequence += 1;
-    clearTurnTimers(state);
-  }
-
-  private async handleSpeechEnded(pkt: VadSpeechEndedPacket): Promise<void> {
-    if (this.lockedContextIds.has(pkt.contextId)) return;
-    const state = this.stateFor(pkt.contextId);
-    if (state.sttQuietTimer) {
-      clearTimeout(state.sttQuietTimer);
-      state.sttQuietTimer = null;
-    }
-    await this.analyzeBoundary(state);
-  }
-
-  // Shared end-of-speech boundary analysis: used by the VAD speech_ended path
-  // and by the STT-quiet fallback (when the provider transcript has gone quiet
-  // but the VAD never closed the segment — e.g. model state saturation on long
-  // telephony audio). The turn must never be held hostage by a wedged VAD.
-  private async analyzeBoundary(state: TurnState): Promise<void> {
-    state.speechActive = false;
-    const sequence = ++state.analysisSequence;
-    const probability = await this.predictor.predict(Float32Array.from(state.audio));
-    if (state.finalized || state.analysisSequence !== sequence) return;
-
-    state.boundaryAnalyzed = true;
-    state.smartTurnComplete = probability > this.probabilityThreshold;
-
-    const transcript = latestTranscript(state.finalSegments, state.latestInterim);
-    const semantic = transcript.trim()
-      ? scoreSemanticCompleteness(transcript)
-      : { complete: state.smartTurnComplete, label: "complete" as const, confidence: 0 };
-    state.semanticComplete = semantic.complete;
-
-    const fusion = transcript.trim()
-      ? fuseEndpointDecision(state.smartTurnComplete, semantic, this.fusionConfig())
-      : {
-          release: state.smartTurnComplete,
-          requestFinalize: state.smartTurnComplete,
-          finalizeDelayMs: this.finalizeDelayMs,
-        };
-
-    if (state.maxTimer) {
-      clearTimeout(state.maxTimer);
-      state.maxTimer = null;
-    }
-
-    if (fusion.deferReason) {
-      this.scheduleSemanticDefer(state);
-      return;
-    }
-
-    if (!fusion.release) {
-      this.scheduleIncompleteFallback(state);
-      return;
-    }
-
-    if (fusion.requestFinalize) {
-      this.requestSttFinalize(state.contextId);
-    }
-    if (state.finalPackets.length > 0) {
-      this.scheduleFinalize(state, fusion.finalizeDelayMs);
-    }
-  }
-
-  private stateFor(contextId: string): TurnState {
-    const existing = this.turns.get(contextId);
-    if (existing) return existing;
-
-    const state: TurnState = {
-      contextId,
-      audio: [],
-      finalPackets: [],
-      finalSegments: [],
-      latestInterim: "",
-      boundaryAnalyzed: false,
-      smartTurnComplete: false,
-      semanticComplete: false,
-      speechActive: false,
-      finalized: false,
-      analysisSequence: 0,
-      finalizeTimer: null,
-      sttQuietTimer: null,
-      maxTimer: null,
-      deferTimer: null,
-    };
-    this.turns.set(contextId, state);
-    return state;
-  }
-
-  private scheduleFinalize(state: TurnState, delayMs: number): void {
-    if (state.finalized || state.finalizeTimer) return;
-    state.finalizeTimer = setTimeout(() => {
-      state.finalizeTimer = null;
-      this.finalize(state);
-    }, delayMs);
-  }
-
-  private scheduleIncompleteFallback(state: TurnState): void {
-    if (state.finalized || state.finalizeTimer) return;
-    state.finalizeTimer = setTimeout(() => {
-      state.finalizeTimer = null;
-      state.smartTurnComplete = true;
-      if (state.finalPackets.length > 0) {
-        this.finalize(state);
-        return;
-      }
-      this.requestSttFinalize(state.contextId);
-    }, this.incompleteFallbackMs);
-  }
-
-  private scheduleSemanticShortcut(state: TurnState): void {
-    if (state.finalized || state.finalizeTimer) return;
-    state.finalizeTimer = setTimeout(() => {
-      state.finalizeTimer = null;
-      state.smartTurnComplete = true;
-      this.requestSttFinalize(state.contextId);
-      if (state.finalPackets.length > 0) {
-        this.finalize(state);
-      }
-    }, this.semanticShortcutDelayMs);
-  }
-
-  private scheduleSemanticDefer(state: TurnState): void {
-    if (state.finalized || state.deferTimer) return;
-    state.deferTimer = setTimeout(() => {
-      state.deferTimer = null;
-      if (state.finalized || state.speechActive) return;
-      const transcript = latestTranscript(state.finalSegments, state.latestInterim);
-      const semantic = scoreSemanticCompleteness(transcript);
-      state.smartTurnComplete = true;
-      this.requestSttFinalize(state.contextId);
-      if (state.finalPackets.length > 0) {
-        if (semantic.complete) {
-          state.semanticComplete = true;
-          this.scheduleFinalize(state, this.finalizeDelayMs);
-        } else {
-          this.finalize(state);
-        }
-      }
-    }, this.semanticDeferFallbackMs);
-  }
-
-  private fusionConfig(): SemanticEndpointFusionConfig {
-    return {
-      enabled: this.semanticEndpointingEnabled,
-      finalizeDelayMs: this.finalizeDelayMs,
-      semanticShortcutDelayMs: this.semanticShortcutDelayMs,
-      incompleteFallbackMs: this.incompleteFallbackMs,
-    };
-  }
-
-  private armSttQuietFallback(state: TurnState): void {
-    if (this.sttQuietFallbackMs <= 0 || state.finalized) return;
-    if (state.sttQuietTimer) clearTimeout(state.sttQuietTimer);
-    state.sttQuietTimer = setTimeout(() => {
-      state.sttQuietTimer = null;
-      if (state.finalized || !state.speechActive || state.finalPackets.length === 0) return;
-      this.bus?.push(Route.Main, {
-        kind: "metric.conversation",
-        contextId: state.contextId,
-        timestampMs: Date.now(),
-        name: "eos.stt_quiet_fallback",
-        value: String(this.sttQuietFallbackMs),
-      });
-      void this.analyzeBoundary(state);
-    }, this.sttQuietFallbackMs);
-  }
-
-  private scheduleMaxFinalize(state: TurnState): void {
-    if (state.finalized || state.maxTimer || this.maxDelayMs <= 0) return;
-    state.maxTimer = setTimeout(() => {
-      state.maxTimer = null;
-      this.finalize(state);
-    }, this.maxDelayMs);
-  }
-
-  private handleTtsAudio(pkt: TextToSpeechAudioPacket): void {
-    if (this.lockedContextIds.has(pkt.contextId)) {
-      this.lockedContextsWithAssistantAudio.add(pkt.contextId);
-    }
-  }
-
-  private handleTtsEnd(pkt: TextToSpeechEndPacket): void {
-    if (!this.lockedContextsWithAssistantAudio.has(pkt.contextId)) {
-      this.releaseContextLock(pkt.contextId);
-    }
-  }
-
-  private handleTtsPlayoutProgress(pkt: TextToSpeechPlayoutProgressPacket): void {
-    if (pkt.complete) this.releaseContextLock(pkt.contextId);
-  }
-
-  private releaseContextLock(contextId: string): void {
-    this.lockedContextIds.delete(contextId);
-    this.lockedContextsWithAssistantAudio.delete(contextId);
-  }
-
-  private finalize(state: TurnState): void {
-    const text = state.finalSegments.join(" ").replace(/\s+/g, " ").trim();
-    if (state.finalized || state.finalPackets.length === 0 || !text) return;
-    state.finalized = true;
-    this.lockedContextIds.add(state.contextId);
-    clearTurnTimers(state);
-    this.bus?.push(Route.Main, {
-      kind: "eos.turn_complete",
-      contextId: state.contextId,
-      timestampMs: Date.now(),
-      text,
-      transcripts: state.finalPackets,
-    } satisfies EndOfSpeechPacket);
-    this.turns.delete(state.contextId);
-  }
-
-  private requestSttFinalize(contextId: string): void {
-    this.bus?.push(Route.Critical, {
-      kind: "stt.finalize",
-      contextId,
-      timestampMs: Date.now(),
-    } satisfies FinalizeSttPacket);
-  }
-}
-
-function appendFinalPacket(state: TurnState, packet: SttResultPacket): void {
-  const text = packet.text.trim();
-  if (state.finalSegments.at(-1) === text) return;
-  state.finalSegments.push(text);
-  state.finalPackets.push(packet);
-}
-
-function clearTurnTimers(state: TurnState): void {
-  if (state.finalizeTimer) clearTimeout(state.finalizeTimer);
-  if (state.sttQuietTimer) clearTimeout(state.sttQuietTimer);
-  if (state.maxTimer) clearTimeout(state.maxTimer);
-  if (state.deferTimer) clearTimeout(state.deferTimer);
-  state.finalizeTimer = null;
-  state.sttQuietTimer = null;
-  state.maxTimer = null;
-  state.deferTimer = null;
-}
-
-function readBooleanConfig(value: unknown, fallback: boolean): boolean {
-  if (typeof value === "boolean") return value;
-  return fallback;
 }
 
 export {
@@ -466,13 +35,3 @@ export {
   SmartTurnInteractionPolicy,
   type SmartTurnInteractionPolicyConfig,
 } from "./interaction-policy.js";
-
-function readNonNegativeNumber(value: unknown, fallback: number): number {
-  if (typeof value !== "number" || !Number.isFinite(value)) return fallback;
-  return Math.max(0, Math.floor(value));
-}
-
-function readProbability(value: unknown, fallback: number): number {
-  if (typeof value !== "number" || !Number.isFinite(value)) return fallback;
-  return Math.min(1, Math.max(0, value));
-}

package/src/interaction-policy.ts

@@ -22,6 +22,7 @@ const DEFAULT_MAX_AUDIO_SAMPLES = SAMPLE_RATE * 8;
 interface SmartTurnState {
   readonly contextId: string;
   audio: number[];
+  speechMs: number;
   finalSegments: string[];
   latestInterim: string;
   speechActive: boolean;
@@ -40,6 +41,7 @@ export interface SmartTurnInteractionPolicyConfig {
   readonly semantic_shortcut_delay_ms?: number;
   readonly incomplete_fallback_ms?: number;
   readonly semantic_defer_fallback_ms?: number;
+  readonly min_speech_ms?: number;
   readonly max_audio_samples?: number;
 }
 
@@ -51,6 +53,7 @@ export class SmartTurnInteractionPolicy implements LifecycleInteractionPolicy {
   private semanticShortcutDelayMs = 50;
   private incompleteFallbackMs = 2000;
   private semanticDeferFallbackMs = 4000;
+  private minSpeechMs = 0;
   private maxAudioSamples = DEFAULT_MAX_AUDIO_SAMPLES;
   private initialized = false;
 
@@ -63,6 +66,7 @@ export class SmartTurnInteractionPolicy implements LifecycleInteractionPolicy {
     this.semanticShortcutDelayMs = readNonNegativeNumber(config["semantic_shortcut_delay_ms"], 50);
     this.incompleteFallbackMs = readNonNegativeNumber(config["incomplete_fallback_ms"], 2000);
     this.semanticDeferFallbackMs = readNonNegativeNumber(config["semantic_defer_fallback_ms"], 4000);
+    this.minSpeechMs = readNonNegativeNumber(config["min_speech_ms"], 0);
     this.maxAudioSamples = readPositiveNumber(config["max_audio_samples"], DEFAULT_MAX_AUDIO_SAMPLES);
     await this.predictor.initialize(config);
     this.initialized = true;
@@ -72,7 +76,7 @@ export class SmartTurnInteractionPolicy implements LifecycleInteractionPolicy {
     const state = this.stateFor(observation.contextId);
     switch (observation.kind) {
       case "audio_frame":
-        this.appendAudio(state, observation.audio);
+        this.appendAudio(state, observation.audio, observation.sampleRateHz);
         break;
       case "stt_partial":
         if (observation.text.trim()) state.latestInterim = observation.text.trim();
@@ -118,6 +122,7 @@ export class SmartTurnInteractionPolicy implements LifecycleInteractionPolicy {
     const state: SmartTurnState = {
       contextId,
       audio: [],
+      speechMs: 0,
       finalSegments: [],
       latestInterim: "",
       speechActive: false,
@@ -135,6 +140,7 @@ export class SmartTurnInteractionPolicy implements LifecycleInteractionPolicy {
   private beginSpeech(state: SmartTurnState): void {
     this.clearFallback(state);
     state.audio = [];
+    state.speechMs = 0;
     state.finalSegments = [];
     state.latestInterim = "";
     state.speechActive = true;
@@ -145,9 +151,12 @@ export class SmartTurnInteractionPolicy implements LifecycleInteractionPolicy {
     state.pendingDecisions = [];
   }
 
-  private appendAudio(state: SmartTurnState, audio?: Int16Array): void {
+  private appendAudio(state: SmartTurnState, audio: Int16Array | undefined, sampleRateHz = SAMPLE_RATE): void {
     if (!audio?.length) return;
     for (const sample of audio) state.audio.push(sample / 32768);
+    if (state.speechActive && sampleRateHz > 0) {
+      state.speechMs += (audio.length / sampleRateHz) * 1000;
+    }
     const overflow = state.audio.length - this.maxAudioSamples;
     if (overflow > 0) state.audio.splice(0, overflow);
   }
@@ -177,13 +186,16 @@ export class SmartTurnInteractionPolicy implements LifecycleInteractionPolicy {
     const smartTurnComplete = state.probability > this.probabilityThreshold;
     const transcript = latestTranscript(state.finalSegments, state.latestInterim);
     if (!transcript) {
-      if (smartTurnComplete) this.issueTakeTurn(state, state.probability);
-      else this.scheduleFallback(state, this.incompleteFallbackMs);
+      if (smartTurnComplete && state.speechMs >= this.minSpeechMs) {
+        this.issueTakeTurn(state, state.probability);
+      } else {
+        this.scheduleFallback(state, this.incompleteFallbackMs);
+      }
       return;
     }
 
     const semantic = scoreSemanticCompleteness(transcript);
-    const fusion = fuseEndpointDecision(smartTurnComplete, semantic, this.fusionConfig());
+    const fusion = fuseEndpointDecision(smartTurnComplete, semantic, state.speechMs, this.fusionConfig());
     if (fusion.release) {
       const confidence = smartTurnComplete ? Math.max(state.probability, semantic.confidence) : semantic.confidence;
       this.issueTakeTurn(state, confidence);
@@ -237,6 +249,7 @@ export class SmartTurnInteractionPolicy implements LifecycleInteractionPolicy {
       finalizeDelayMs: this.finalizeDelayMs,
       semanticShortcutDelayMs: this.semanticShortcutDelayMs,
       incompleteFallbackMs: this.incompleteFallbackMs,
+      minSpeechMs: this.minSpeechMs,
     };
   }
 }

package/src/predictor.ts

@@ -3,6 +3,9 @@
 import { fileURLToPath } from "node:url";
 
 import { optionalStringConfig, type PluginConfig } from "@kuralle-syrinx/core";
+import type { SmartTurnPredictor } from "./smart-turn-types.js";
+
+export type { SmartTurnPredictor } from "./smart-turn-types.js";
 
 type Ort = typeof import("onnxruntime-node");
 type InferenceSession = import("onnxruntime-node").InferenceSession;
@@ -15,12 +18,6 @@ const SAMPLE_RATE = 16000;
 const MAX_AUDIO_SAMPLES = SAMPLE_RATE * 8;
 const DEFAULT_MODEL_PATH = fileURLToPath(new URL("../models/smart-turn-v3.2-cpu.onnx", import.meta.url));
 
-export interface SmartTurnPredictor {
-  initialize(config: PluginConfig): Promise<void>;
-  predict(audio: Float32Array): Promise<number>;
-  close(): Promise<void>;
-}
-
 export class LocalSmartTurnV3Predictor implements SmartTurnPredictor {
   private ort: Ort | null = null;
   private session: InferenceSession | null = null;

package/src/semantic-completeness.ts

@@ -11,6 +11,7 @@ export interface SemanticEndpointFusionConfig {
   readonly finalizeDelayMs: number;
   readonly semanticShortcutDelayMs: number;
   readonly incompleteFallbackMs: number;
+  readonly minSpeechMs?: number;
 }
 
 export interface EndpointFusionDecision {
@@ -111,6 +112,7 @@ export function scoreSemanticCompleteness(text: string): SemanticCompletenessSco
 export function fuseEndpointDecision(
   smartTurnComplete: boolean,
   semantic: SemanticCompletenessScore,
+  speechMs: number,
   config: SemanticEndpointFusionConfig,
 ): EndpointFusionDecision {
   if (!config.enabled) {
@@ -121,7 +123,7 @@ export function fuseEndpointDecision(
     };
   }
 
-  if (smartTurnComplete && semantic.complete) {
+  if (smartTurnComplete && semantic.complete && speechMs >= (config.minSpeechMs ?? 0)) {
     return {
       release: true,
       requestFinalize: true,

package/src/smart-turn-types.ts

@@ -0,0 +1,10 @@
+// SPDX-License-Identifier: MIT
+
+import type { PluginConfig } from "@kuralle-syrinx/core";
+
+/** Host- or model-agnostic Smart Turn probability scorer. */
+export interface SmartTurnPredictor {
+  initialize(config: PluginConfig): Promise<void>;
+  predict(audio: Float32Array): Promise<number>;
+  close(): Promise<void>;
+}