@livekit/agents-plugin-livekit 0.1.3 → 1.0.0-next.1

package/src/turn_detector/index.ts

@@ -0,0 +1,21 @@
+// SPDX-FileCopyrightText: 2024 LiveKit, Inc.
+//
+// SPDX-License-Identifier: Apache-2.0
+import { InferenceRunner } from '@livekit/agents';
+import { INFERENCE_METHOD_EN } from './english.js';
+import { INFERENCE_METHOD_MULTILINGUAL } from './multilingual.js';
+
+export { EOUModel } from './base.js';
+export { EnglishModel } from './english.js';
+export { MultilingualModel } from './multilingual.js';
+export { getUnicodeCategory, normalizeText } from './utils.js';
+
+InferenceRunner.registerRunner(
+  INFERENCE_METHOD_EN,
+  new URL('./english.js', import.meta.url).toString(),
+);
+
+InferenceRunner.registerRunner(
+  INFERENCE_METHOD_MULTILINGUAL,
+  new URL('./multilingual.js', import.meta.url).toString(),
+);

package/src/turn_detector/multilingual.ts

@@ -0,0 +1,145 @@
+// SPDX-FileCopyrightText: 2025 LiveKit, Inc.
+//
+// SPDX-License-Identifier: Apache-2.0
+import type { llm } from '@livekit/agents';
+import { getJobContext, log } from '@livekit/agents';
+import { EOUModel, EOURunnerBase } from './base.js';
+import { MAX_HISTORY_TURNS } from './constants.js';
+
+const REMOTE_INFERENCE_TIMEOUT = 2000;
+
+export const INFERENCE_METHOD_MULTILINGUAL = 'lk_end_of_utterance_multilingual';
+
+export class EUORunnerMultilingual extends EOURunnerBase {
+  constructor() {
+    super('multilingual');
+  }
+}
+
+export class MultilingualModel extends EOUModel {
+  #logger = log();
+
+  constructor(unlikelyThreshold?: number) {
+    super({
+      modelType: 'multilingual',
+      unlikelyThreshold,
+    });
+  }
+
+  inferenceMethod(): string {
+    return INFERENCE_METHOD_MULTILINGUAL;
+  }
+
+  async unlikelyThreshold(language?: string): Promise<number | undefined> {
+    if (!language) {
+      return undefined;
+    }
+
+    let threshold = await super.unlikelyThreshold(language);
+    if (threshold === undefined) {
+      const url = remoteInferenceUrl();
+      if (!url) return undefined;
+
+      const resp = await fetch(url, {
+        method: 'POST',
+        body: JSON.stringify({
+          language,
+        }),
+        headers: {
+          'Content-Type': 'application/json',
+        },
+        signal: AbortSignal.timeout(REMOTE_INFERENCE_TIMEOUT),
+      });
+
+      if (!resp.ok) {
+        throw new Error(`Failed to fetch threshold: ${resp.statusText}`);
+      }
+
+      const data = (await resp.json()) as { threshold: number | undefined };
+      threshold = data.threshold;
+      if (threshold) {
+        const languages = await this.languagesFuture.await;
+        languages[language] = { threshold };
+      }
+    }
+
+    return threshold;
+  }
+
+  async predictEndOfTurn(chatCtx: llm.ChatContext, timeout: number = 3): Promise<number> {
+    const url = remoteInferenceUrl();
+    if (!url) {
+      return await super.predictEndOfTurn(chatCtx, timeout);
+    }
+
+    // Copy and process chat context similar to Python implementation
+    const messages = chatCtx
+      .copy({
+        excludeFunctionCall: true,
+        excludeInstructions: true,
+        excludeEmptyMessage: true,
+      })
+      .truncate(MAX_HISTORY_TURNS);
+
+    // Get job context and build request
+    const ctx = getJobContext();
+    const request: any = {
+      ...messages.toJSON({
+        excludeImage: true,
+        excludeAudio: true,
+        excludeTimestamp: true,
+      }),
+      jobId: ctx.job.id,
+      workerId: ctx.workerId,
+    };
+
+    // Add agentId from environment variable if available
+    const agentId = process.env.LIVEKIT_AGENT_ID;
+    if (agentId) {
+      request.agentId = agentId;
+    }
+
+    const startedAt = performance.now();
+
+    this.#logger.debug({ url, request }, '=== remote EOU inference');
+
+    const resp = await fetch(url, {
+      method: 'POST',
+      body: JSON.stringify(request),
+      headers: {
+        'Content-Type': 'application/json',
+      },
+      signal: AbortSignal.timeout(REMOTE_INFERENCE_TIMEOUT),
+    });
+
+    if (!resp.ok) {
+      throw new Error(`Failed to predict end of turn: ${resp.statusText}`);
+    }
+
+    const data = await resp.json();
+    const probability = data.probability;
+    if (typeof probability === 'number' && probability >= 0) {
+      this.#logger.debug(
+        {
+          eouProbability: probability,
+          duration: (performance.now() - startedAt) / 1000,
+        },
+        'eou prediction',
+      );
+      return probability;
+    }
+
+    // default to indicate no prediction
+    return 1;
+  }
+}
+
+function remoteInferenceUrl() {
+  const urlBase = process.env.LIVEKIT_REMOTE_EOT_URL;
+  if (!urlBase) {
+    return undefined;
+  }
+  return `${urlBase}/eot/multi`;
+}
+
+export default EUORunnerMultilingual;

package/src/turn_detector/utils.test.ts

@@ -0,0 +1,231 @@
+// SPDX-FileCopyrightText: 2025 LiveKit, Inc.
+//
+// SPDX-License-Identifier: Apache-2.0
+import { describe, expect, it } from 'vitest';
+import { getUnicodeCategory, normalizeText } from './utils.js';
+
+describe('getUnicodeCategory', () => {
+  it('should identify basic ASCII punctuation', () => {
+    expect(getUnicodeCategory('!')).toBe('P');
+    expect(getUnicodeCategory('"')).toBe('P');
+    expect(getUnicodeCategory('#')).toBe('P');
+    expect(getUnicodeCategory('$')).toBe('P');
+    expect(getUnicodeCategory('%')).toBe('P');
+    expect(getUnicodeCategory('&')).toBe('P');
+    expect(getUnicodeCategory("'")).toBe('P');
+    expect(getUnicodeCategory('(')).toBe('P');
+    expect(getUnicodeCategory(')')).toBe('P');
+    expect(getUnicodeCategory('*')).toBe('P');
+    expect(getUnicodeCategory('+')).toBe('P');
+    expect(getUnicodeCategory(',')).toBe('P');
+    expect(getUnicodeCategory('-')).toBe('P');
+    expect(getUnicodeCategory('.')).toBe('P');
+    expect(getUnicodeCategory('/')).toBe('P');
+  });
+
+  it('should identify colon/semicolon punctuation', () => {
+    expect(getUnicodeCategory(':')).toBe('P');
+    expect(getUnicodeCategory(';')).toBe('P');
+    expect(getUnicodeCategory('<')).toBe('P');
+    expect(getUnicodeCategory('=')).toBe('P');
+    expect(getUnicodeCategory('>')).toBe('P');
+    expect(getUnicodeCategory('?')).toBe('P');
+    expect(getUnicodeCategory('@')).toBe('P');
+  });
+
+  it('should identify bracket punctuation', () => {
+    expect(getUnicodeCategory('[')).toBe('P');
+    expect(getUnicodeCategory('\\')).toBe('P');
+    expect(getUnicodeCategory(']')).toBe('P');
+    expect(getUnicodeCategory('^')).toBe('P');
+    expect(getUnicodeCategory('_')).toBe('P');
+    expect(getUnicodeCategory('`')).toBe('P');
+  });
+
+  it('should identify brace punctuation', () => {
+    expect(getUnicodeCategory('{')).toBe('P');
+    expect(getUnicodeCategory('|')).toBe('P');
+    expect(getUnicodeCategory('}')).toBe('P');
+    expect(getUnicodeCategory('~')).toBe('P');
+  });
+
+  it('should identify extended punctuation', () => {
+    expect(getUnicodeCategory('¡')).toBe('P');
+    expect(getUnicodeCategory('¿')).toBe('P');
+    expect(getUnicodeCategory('«')).toBe('P');
+    expect(getUnicodeCategory('»')).toBe('P');
+  });
+
+  it('should not identify letters as punctuation', () => {
+    expect(getUnicodeCategory('a')).toBe('');
+    expect(getUnicodeCategory('A')).toBe('');
+    expect(getUnicodeCategory('z')).toBe('');
+    expect(getUnicodeCategory('Z')).toBe('');
+  });
+
+  it('should not identify numbers as punctuation', () => {
+    expect(getUnicodeCategory('0')).toBe('');
+    expect(getUnicodeCategory('1')).toBe('');
+    expect(getUnicodeCategory('9')).toBe('');
+  });
+
+  it('should not identify whitespace as punctuation', () => {
+    expect(getUnicodeCategory(' ')).toBe('');
+    expect(getUnicodeCategory('\t')).toBe('');
+    expect(getUnicodeCategory('\n')).toBe('');
+  });
+
+  it('should handle empty string', () => {
+    expect(getUnicodeCategory('')).toBe('');
+  });
+
+  it('should handle unicode characters', () => {
+    expect(getUnicodeCategory('é')).toBe('');
+    expect(getUnicodeCategory('ñ')).toBe('');
+    expect(getUnicodeCategory('ç')).toBe('');
+  });
+});
+
+describe('normalizeText', () => {
+  describe('basic functionality', () => {
+    it('should convert to lowercase', () => {
+      expect(normalizeText('HELLO')).toBe('hello');
+      expect(normalizeText('HeLLo')).toBe('hello');
+      expect(normalizeText('WORLD')).toBe('world');
+    });
+
+    it('should remove basic punctuation', () => {
+      expect(normalizeText('Hello!')).toBe('hello');
+      expect(normalizeText('Hello?')).toBe('hello');
+      expect(normalizeText('Hello.')).toBe('hello');
+      expect(normalizeText('Hello,')).toBe('hello');
+    });
+
+    it('should preserve apostrophes', () => {
+      expect(normalizeText("I'm happy")).toBe("i'm happy");
+      expect(normalizeText("don't worry")).toBe("don't worry");
+      expect(normalizeText("it's great")).toBe("it's great");
+    });
+
+    it('should preserve hyphens', () => {
+      expect(normalizeText('well-trained')).toBe('well-trained');
+      expect(normalizeText('state-of-the-art')).toBe('state-of-the-art');
+      expect(normalizeText('co-worker')).toBe('co-worker');
+    });
+
+    it('should collapse multiple whitespace', () => {
+      expect(normalizeText('hello    world')).toBe('hello world');
+      expect(normalizeText('multiple   spaces   here')).toBe('multiple spaces here');
+      expect(normalizeText('tab\t\tspaces')).toBe('tab spaces');
+      expect(normalizeText('newline\n\nspaces')).toBe('newline spaces');
+    });
+
+    it('should trim leading and trailing whitespace', () => {
+      expect(normalizeText('  hello  ')).toBe('hello');
+      expect(normalizeText('\t\nhello\t\n')).toBe('hello');
+      expect(normalizeText('   hello world   ')).toBe('hello world');
+    });
+  });
+
+  describe('comprehensive test cases', () => {
+    it('should handle the basic greeting case', () => {
+      expect(normalizeText('Hi, how can I help you today?')).toBe('hi how can i help you today');
+    });
+
+    it('should handle contractions and hyphens', () => {
+      expect(normalizeText("I'm a well-trained assistant!")).toBe("i'm a well-trained assistant");
+    });
+
+    it('should remove various punctuation types', () => {
+      expect(normalizeText('Hello!!! What??? Price: $19.99 (20% off).')).toBe(
+        'hello what price 1999 20 off',
+      );
+    });
+
+    it('should handle multiple spaces', () => {
+      expect(normalizeText('Multiple    spaces   here')).toBe('multiple spaces here');
+    });
+
+    it('should handle unicode characters', () => {
+      expect(normalizeText('Café entrées naïve résumé')).toBe('café entrées naïve résumé');
+    });
+
+    it('should handle mixed punctuation and unicode', () => {
+      expect(normalizeText('¿Cómo estás? ¡Muy bien!')).toBe('cómo estás muy bien');
+    });
+  });
+
+  describe('edge cases', () => {
+    it('should handle empty string', () => {
+      expect(normalizeText('')).toBe('');
+    });
+
+    it('should handle whitespace-only string', () => {
+      expect(normalizeText('   ')).toBe('');
+      expect(normalizeText('\t\n  ')).toBe('');
+    });
+
+    it('should handle punctuation-only string', () => {
+      expect(normalizeText('!!!')).toBe('');
+      expect(normalizeText('???')).toBe('');
+      expect(normalizeText('...')).toBe('');
+    });
+
+    it('should handle mixed punctuation and preserved characters', () => {
+      expect(normalizeText('!@#$%^&*()_+-={}[]|\\:;"\'<>?,./')).toBe("-'");
+    });
+
+    it('should handle numbers with punctuation', () => {
+      expect(normalizeText('$19.99')).toBe('1999');
+      expect(normalizeText('(555) 123-4567')).toBe('555 123-4567');
+    });
+
+    it('should handle special unicode punctuation', () => {
+      expect(normalizeText('Hello… world!')).toBe('hello world');
+      expect(normalizeText('"Quoted text"')).toBe('quoted text');
+      expect(normalizeText("'Single quotes'")).toBe("'single quotes'");
+    });
+  });
+
+  describe('unicode normalization', () => {
+    it('should apply NFKC normalization', () => {
+      expect(normalizeText('café')).toBe('café');
+      expect(normalizeText('naïve')).toBe('naïve');
+    });
+
+    it('should handle combining characters', () => {
+      const eWithCombiningAcute = 'caf\u0065\u0301';
+      const precomposedE = 'café';
+      expect(normalizeText(eWithCombiningAcute)).toBe(normalizeText(precomposedE));
+    });
+  });
+
+  describe('real-world examples', () => {
+    it('should handle typical assistant responses', () => {
+      expect(normalizeText('Hello! How can I assist you today?')).toBe(
+        'hello how can i assist you today',
+      );
+      expect(normalizeText("I'm here to help with any questions you might have.")).toBe(
+        "i'm here to help with any questions you might have",
+      );
+    });
+
+    it('should handle typical user queries', () => {
+      expect(normalizeText("What's the weather like?")).toBe("what's the weather like");
+      expect(normalizeText('Can you help me with my order?')).toBe('can you help me with my order');
+      expect(normalizeText('I need assistance, please!')).toBe('i need assistance please');
+    });
+
+    it('should handle incomplete sentences', () => {
+      expect(normalizeText('What is the weather in')).toBe('what is the weather in');
+      expect(normalizeText('I am looking for')).toBe('i am looking for');
+      expect(normalizeText('Could you please')).toBe('could you please');
+    });
+
+    it('should handle multilingual text', () => {
+      expect(normalizeText('Bonjour! Comment ça va?')).toBe('bonjour comment ça va');
+      expect(normalizeText('¡Hola! ¿Cómo estás?')).toBe('hola cómo estás');
+      expect(normalizeText('Guten Tag! Wie geht es Ihnen?')).toBe('guten tag wie geht es ihnen');
+    });
+  });
+});

package/src/turn_detector/utils.ts

@@ -0,0 +1,76 @@
+// SPDX-FileCopyrightText: 2025 LiveKit, Inc.
+//
+// SPDX-License-Identifier: Apache-2.0
+
+/**
+ * Text normalization utilities for EOU turn detector
+ */
+
+/**
+ * Simple unicode category detection for punctuation
+ * Mimics Python's unicodedata.category() for punctuation detection
+ */
+export function getUnicodeCategory(char: string): string {
+  const code = char.codePointAt(0);
+  if (!code) return '';
+
+  // Basic punctuation ranges (simplified version to match Python unicodedata.category)
+  if (
+    (code >= 0x21 && code <= 0x2f) || // !"#$%&'()*+,-./
+    (code >= 0x3a && code <= 0x40) || // :;<=>?@
+    (code >= 0x5b && code <= 0x60) || // [\]^_`
+    (code >= 0x7b && code <= 0x7e) || // {|}~
+    (code >= 0xa0 && code <= 0xbf) || // Latin-1 punctuation
+    (code >= 0x2000 && code <= 0x206f) || // General punctuation
+    (code >= 0x3000 && code <= 0x303f)
+  ) {
+    // CJK symbols and punctuation
+    return 'P';
+  }
+  return '';
+}
+
+/**
+ * Normalizes text to match the training data format used by the EOU model
+ *
+ * This function applies the following transformations:
+ * 1. Converts to lowercase
+ * 2. Applies Unicode NFKC normalization
+ * 3. Removes all punctuation except apostrophes (') and hyphens (-)
+ * 4. Collapses multiple whitespace characters into single spaces
+ * 5. Trims leading and trailing whitespace
+ *
+ * @param text - The input text to normalize
+ * @returns The normalized text
+ *
+ * @example
+ * ```typescript
+ * normalizeText("Hi, how can I help you today?")
+ * // Returns: "hi how can i help you today"
+ *
+ * normalizeText("I'm a well-trained assistant!")
+ * // Returns: "i'm a well-trained assistant"
+ *
+ * normalizeText("Price: $19.99 (20% off).")
+ * // Returns: "price 1999 20 off"
+ * ```
+ */
+export function normalizeText(text: string): string {
+  if (!text) return '';
+
+  let normalized = text.toLowerCase().normalize('NFKC');
+
+  // Remove punctuation except apostrophes and hyphens
+  // Using character-by-character approach to match Python logic
+  normalized = Array.from(normalized)
+    .filter((ch) => {
+      const category = getUnicodeCategory(ch);
+      return !(category.startsWith('P') && ch !== "'" && ch !== '-');
+    })
+    .join('');
+
+  // Collapse whitespace and trim
+  normalized = normalized.replace(/\s+/g, ' ').trim();
+
+  return normalized;
+}

package/dist/turn_detector.cjs

@@ -1,129 +0,0 @@
-"use strict";
-var __defProp = Object.defineProperty;
-var __getOwnPropDesc = Object.getOwnPropertyDescriptor;
-var __getOwnPropNames = Object.getOwnPropertyNames;
-var __hasOwnProp = Object.prototype.hasOwnProperty;
-var __export = (target, all) => {
-  for (var name in all)
-    __defProp(target, name, { get: all[name], enumerable: true });
-};
-var __copyProps = (to, from, except, desc) => {
-  if (from && typeof from === "object" || typeof from === "function") {
-    for (let key of __getOwnPropNames(from))
-      if (!__hasOwnProp.call(to, key) && key !== except)
-        __defProp(to, key, { get: () => from[key], enumerable: !(desc = __getOwnPropDesc(from, key)) || desc.enumerable });
-  }
-  return to;
-};
-var __toCommonJS = (mod) => __copyProps(__defProp({}, "__esModule", { value: true }), mod);
-var turn_detector_exports = {};
-__export(turn_detector_exports, {
-  EOUModel: () => EOUModel,
-  EOURunner: () => EOURunner,
-  default: () => turn_detector_default
-});
-module.exports = __toCommonJS(turn_detector_exports);
-var getImportMetaUrl = () => typeof document === "undefined" ? new URL(`file:${__filename}`).href : document.currentScript && document.currentScript.src || new URL("main.js", document.baseURI).href;
-var importMetaUrl = /* @__PURE__ */ getImportMetaUrl();
-var import_transformers = require("@huggingface/transformers");
-var import_agents = require("@livekit/agents");
-var import_node_url = require("node:url");
-var import_onnxruntime_node = require("onnxruntime-node");
-const MAX_HISTORY = 4;
-class EOURunner extends import_agents.InferenceRunner {
-  static INFERENCE_METHOD = "lk_end_of_utterance";
-  #tokenizerPromise;
-  #session;
-  #tokenizer;
-  #logger = (0, import_agents.log)();
-  constructor() {
-    super();
-    this.#tokenizerPromise = import_transformers.AutoTokenizer.from_pretrained("livekit/turn-detector", {
-      revision: "v1.2.0"
-      // local_files_only: true, // TODO(nbsp): can't find it
-    });
-    this.#session = import_onnxruntime_node.InferenceSession.create(
-      (0, import_node_url.fileURLToPath)(new URL("turn_detector.onnx", importMetaUrl).href),
-      {
-        executionProviders: [{ name: "cpu" }]
-      }
-    );
-  }
-  async initialize() {
-    this.#tokenizer = await this.#tokenizerPromise;
-  }
-  async run(data) {
-    const text = this.#formatChatContext(data);
-    const startTime = Date.now();
-    const inputs = this.#tokenizer.encode(text, { add_special_tokens: false });
-    const outputs = await this.#session.then(
-      (session) => session.run({ input_ids: new import_onnxruntime_node.Tensor("int64", inputs, [1, inputs.length]) }, ["prob"])
-    );
-    const endTime = Date.now();
-    const logits = outputs.prob;
-    const eouProbability = logits.data[0];
-    this.#logger.child({ eouProbability, input: text, duration: endTime - startTime }).debug("eou prediction");
-    return eouProbability;
-  }
-  #formatChatContext(ctx) {
-    const newCtx = [];
-    for (const msg of ctx) {
-      if (!msg.content) continue;
-      newCtx.push(msg);
-    }
-    const convoText = this.#tokenizer.apply_chat_template(newCtx, {
-      add_generation_prompt: false,
-      tokenize: false
-    });
-    return convoText.slice(0, convoText.lastIndexOf("<|im_end|>"));
-  }
-  async close() {
-    await this.#session.then((session) => session.release());
-  }
-}
-class EOUModel {
-  unlikelyThreshold;
-  #executor;
-  constructor(unlikelyThreshold = 0.15) {
-    this.unlikelyThreshold = unlikelyThreshold;
-    this.#executor = import_agents.CurrentJobContext.getCurrent().inferenceExecutor;
-  }
-  supportsLanguage(language) {
-    if (!language) return false;
-    const parts = language.toLowerCase().split("-");
-    return parts[0] === "en" || parts[0] === "english";
-  }
-  async predictEndOfTurn(chatCtx) {
-    let messages = [];
-    for (const msg of chatCtx.messages) {
-      if (msg.role !== import_agents.llm.ChatRole.ASSISTANT && msg.role !== import_agents.llm.ChatRole.USER) {
-        continue;
-      }
-      if (typeof msg.content === "string") {
-        messages.push({
-          role: msg.role === import_agents.llm.ChatRole.ASSISTANT ? "assistant" : "user",
-          content: msg.content
-        });
-      } else if (Array.isArray(msg.content)) {
-        for (const content of msg.content) {
-          if (typeof content === "string") {
-            messages.push({
-              role: msg.role === import_agents.llm.ChatRole.ASSISTANT ? "assistant" : "user",
-              content
-            });
-          }
-        }
-      }
-    }
-    messages = messages.slice(-MAX_HISTORY);
-    const result = await this.#executor.doInference(EOURunner.INFERENCE_METHOD, messages);
-    return result;
-  }
-}
-var turn_detector_default = EOURunner;
-// Annotate the CommonJS export names for ESM import in node:
-0 && (module.exports = {
-  EOUModel,
-  EOURunner
-});
-//# sourceMappingURL=turn_detector.cjs.map

package/dist/turn_detector.cjs.map

@@ -1 +0,0 @@
-{"version":3,"sources":["../src/turn_detector.ts","../../../node_modules/.pnpm/tsup@8.3.5_@microsoft+api-extractor@7.43.7_@types+node@22.5.5__postcss@8.4.38_tsx@4.19.2_typescript@5.4.5/node_modules/tsup/assets/cjs_shims.js"],"sourcesContent":["// SPDX-FileCopyrightText: 2024 LiveKit, Inc.\n//\n// SPDX-License-Identifier: Apache-2.0\nimport type { PreTrainedTokenizer } from '@huggingface/transformers';\nimport { AutoTokenizer } from '@huggingface/transformers';\nimport type { ipc } from '@livekit/agents';\nimport { CurrentJobContext, InferenceRunner, llm, log } from '@livekit/agents';\nimport { fileURLToPath } from 'node:url';\nimport { InferenceSession, Tensor } from 'onnxruntime-node';\n\nconst MAX_HISTORY = 4;\n\ntype RawChatContext = { role: string; content: string }[];\n\nexport class EOURunner extends InferenceRunner {\n  static INFERENCE_METHOD = 'lk_end_of_utterance';\n  #tokenizerPromise: Promise<PreTrainedTokenizer>;\n  #session: Promise<InferenceSession>;\n  #tokenizer?: PreTrainedTokenizer;\n  #logger = log();\n\n  constructor() {\n    super();\n    this.#tokenizerPromise = AutoTokenizer.from_pretrained('livekit/turn-detector', {\n      revision: 'v1.2.0',\n      // local_files_only: true, // TODO(nbsp): can't find it\n    });\n    this.#session = InferenceSession.create(\n      fileURLToPath(new URL('turn_detector.onnx', import.meta.url).href),\n      {\n        executionProviders: [{ name: 'cpu' }],\n      },\n    );\n  }\n\n  async initialize() {\n    this.#tokenizer = await this.#tokenizerPromise;\n  }\n\n  async run(data: RawChatContext): Promise<number | undefined> {\n    const text = this.#formatChatContext(data);\n    const startTime = Date.now();\n    const inputs = this.#tokenizer!.encode(text, { add_special_tokens: false });\n    const outputs = await this.#session.then((session) =>\n      session.run({ input_ids: new Tensor('int64', inputs, [1, inputs.length]) }, ['prob']),\n    );\n    const endTime = Date.now();\n    const logits = outputs.prob!;\n    const eouProbability = logits.data[0] as number;\n    this.#logger\n      .child({ eouProbability, input: text, duration: endTime - startTime })\n      .debug('eou prediction');\n    return eouProbability;\n  }\n\n  #formatChatContext(ctx: RawChatContext): string {\n    const newCtx: RawChatContext = [];\n    for (const msg of ctx) {\n      if (!msg.content) continue;\n      newCtx.push(msg);\n    }\n\n    const convoText = this.#tokenizer!.apply_chat_template(newCtx, {\n      add_generation_prompt: false,\n      tokenize: false,\n    }) as string;\n    // remove EOU token from current utterance\n    return convoText.slice(0, convoText.lastIndexOf('<|im_end|>'));\n  }\n\n  async close() {\n    await this.#session.then((session) => session.release());\n  }\n}\n\nexport class EOUModel {\n  readonly unlikelyThreshold: number;\n  #executor: ipc.InferenceExecutor;\n\n  constructor(unlikelyThreshold = 0.15) {\n    this.unlikelyThreshold = unlikelyThreshold;\n    this.#executor = CurrentJobContext.getCurrent().inferenceExecutor;\n  }\n\n  supportsLanguage(language?: string) {\n    if (!language) return false;\n    const parts = language.toLowerCase().split('-');\n    return parts[0] === 'en' || parts[0] === 'english';\n  }\n\n  async predictEndOfTurn(chatCtx: llm.ChatContext): Promise<number> {\n    let messages: RawChatContext = [];\n\n    for (const msg of chatCtx.messages) {\n      if (msg.role !== llm.ChatRole.ASSISTANT && msg.role !== llm.ChatRole.USER) {\n        continue;\n      }\n\n      if (typeof msg.content === 'string') {\n        messages.push({\n          role: msg.role === llm.ChatRole.ASSISTANT ? 'assistant' : 'user',\n          content: msg.content,\n        });\n      } else if (Array.isArray(msg.content)) {\n        for (const content of msg.content) {\n          if (typeof content === 'string') {\n            messages.push({\n              role: msg.role === llm.ChatRole.ASSISTANT ? 'assistant' : 'user',\n              content: content,\n            });\n          }\n        }\n      }\n    }\n    messages = messages.slice(-MAX_HISTORY);\n    const result = await this.#executor.doInference(EOURunner.INFERENCE_METHOD, messages);\n    return result as any;\n  }\n}\n\nexport default EOURunner;\n","// Shim globals in cjs bundle\n// There's a weird bug that esbuild will always inject importMetaUrl\n// if we export it as `const importMetaUrl = ... __filename ...`\n// But using a function will not cause this issue\n\nconst getImportMetaUrl = () =>\n  typeof document === 'undefined'\n    ? new URL(`file:${__filename}`).href\n    : (document.currentScript && document.currentScript.src) ||\n      new URL('main.js', document.baseURI).href\n\nexport const importMetaUrl = /* @__PURE__ */ getImportMetaUrl()\n"],"mappings":";;;;;;;;;;;;;;;;;;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;ACKA,IAAM,mBAAmB,MACvB,OAAO,aAAa,cAChB,IAAI,IAAI,QAAQ,UAAU,EAAE,EAAE,OAC7B,SAAS,iBAAiB,SAAS,cAAc,OAClD,IAAI,IAAI,WAAW,SAAS,OAAO,EAAE;AAEpC,IAAM,gBAAgC,iCAAiB;ADP9D,0BAA8B;AAE9B,oBAA6D;AAC7D,sBAA8B;AAC9B,8BAAyC;AAEzC,MAAM,cAAc;AAIb,MAAM,kBAAkB,8BAAgB;AAAA,EAC7C,OAAO,mBAAmB;AAAA,EAC1B;AAAA,EACA;AAAA,EACA;AAAA,EACA,cAAU,mBAAI;AAAA,EAEd,cAAc;AACZ,UAAM;AACN,SAAK,oBAAoB,kCAAc,gBAAgB,yBAAyB;AAAA,MAC9E,UAAU;AAAA;AAAA,IAEZ,CAAC;AACD,SAAK,WAAW,yCAAiB;AAAA,UAC/B,+BAAc,IAAI,IAAI,sBAAsB,aAAe,EAAE,IAAI;AAAA,MACjE;AAAA,QACE,oBAAoB,CAAC,EAAE,MAAM,MAAM,CAAC;AAAA,MACtC;AAAA,IACF;AAAA,EACF;AAAA,EAEA,MAAM,aAAa;AACjB,SAAK,aAAa,MAAM,KAAK;AAAA,EAC/B;AAAA,EAEA,MAAM,IAAI,MAAmD;AAC3D,UAAM,OAAO,KAAK,mBAAmB,IAAI;AACzC,UAAM,YAAY,KAAK,IAAI;AAC3B,UAAM,SAAS,KAAK,WAAY,OAAO,MAAM,EAAE,oBAAoB,MAAM,CAAC;AAC1E,UAAM,UAAU,MAAM,KAAK,SAAS;AAAA,MAAK,CAAC,YACxC,QAAQ,IAAI,EAAE,WAAW,IAAI,+BAAO,SAAS,QAAQ,CAAC,GAAG,OAAO,MAAM,CAAC,EAAE,GAAG,CAAC,MAAM,CAAC;AAAA,IACtF;AACA,UAAM,UAAU,KAAK,IAAI;AACzB,UAAM,SAAS,QAAQ;AACvB,UAAM,iBAAiB,OAAO,KAAK,CAAC;AACpC,SAAK,QACF,MAAM,EAAE,gBAAgB,OAAO,MAAM,UAAU,UAAU,UAAU,CAAC,EACpE,MAAM,gBAAgB;AACzB,WAAO;AAAA,EACT;AAAA,EAEA,mBAAmB,KAA6B;AAC9C,UAAM,SAAyB,CAAC;AAChC,eAAW,OAAO,KAAK;AACrB,UAAI,CAAC,IAAI,QAAS;AAClB,aAAO,KAAK,GAAG;AAAA,IACjB;AAEA,UAAM,YAAY,KAAK,WAAY,oBAAoB,QAAQ;AAAA,MAC7D,uBAAuB;AAAA,MACvB,UAAU;AAAA,IACZ,CAAC;AAED,WAAO,UAAU,MAAM,GAAG,UAAU,YAAY,YAAY,CAAC;AAAA,EAC/D;AAAA,EAEA,MAAM,QAAQ;AACZ,UAAM,KAAK,SAAS,KAAK,CAAC,YAAY,QAAQ,QAAQ,CAAC;AAAA,EACzD;AACF;AAEO,MAAM,SAAS;AAAA,EACX;AAAA,EACT;AAAA,EAEA,YAAY,oBAAoB,MAAM;AACpC,SAAK,oBAAoB;AACzB,SAAK,YAAY,gCAAkB,WAAW,EAAE;AAAA,EAClD;AAAA,EAEA,iBAAiB,UAAmB;AAClC,QAAI,CAAC,SAAU,QAAO;AACtB,UAAM,QAAQ,SAAS,YAAY,EAAE,MAAM,GAAG;AAC9C,WAAO,MAAM,CAAC,MAAM,QAAQ,MAAM,CAAC,MAAM;AAAA,EAC3C;AAAA,EAEA,MAAM,iBAAiB,SAA2C;AAChE,QAAI,WAA2B,CAAC;AAEhC,eAAW,OAAO,QAAQ,UAAU;AAClC,UAAI,IAAI,SAAS,kBAAI,SAAS,aAAa,IAAI,SAAS,kBAAI,SAAS,MAAM;AACzE;AAAA,MACF;AAEA,UAAI,OAAO,IAAI,YAAY,UAAU;AACnC,iBAAS,KAAK;AAAA,UACZ,MAAM,IAAI,SAAS,kBAAI,SAAS,YAAY,cAAc;AAAA,UAC1D,SAAS,IAAI;AAAA,QACf,CAAC;AAAA,MACH,WAAW,MAAM,QAAQ,IAAI,OAAO,GAAG;AACrC,mBAAW,WAAW,IAAI,SAAS;AACjC,cAAI,OAAO,YAAY,UAAU;AAC/B,qBAAS,KAAK;AAAA,cACZ,MAAM,IAAI,SAAS,kBAAI,SAAS,YAAY,cAAc;AAAA,cAC1D;AAAA,YACF,CAAC;AAAA,UACH;AAAA,QACF;AAAA,MACF;AAAA,IACF;AACA,eAAW,SAAS,MAAM,CAAC,WAAW;AACtC,UAAM,SAAS,MAAM,KAAK,UAAU,YAAY,UAAU,kBAAkB,QAAQ;AACpF,WAAO;AAAA,EACT;AACF;AAEA,IAAO,wBAAQ;","names":[]}

package/dist/turn_detector.d.cts

@@ -1,22 +0,0 @@
-import { InferenceRunner, llm } from '@livekit/agents';
-type RawChatContext = {
-    role: string;
-    content: string;
-}[];
-export declare class EOURunner extends InferenceRunner {
-    #private;
-    static INFERENCE_METHOD: string;
-    constructor();
-    initialize(): Promise<void>;
-    run(data: RawChatContext): Promise<number | undefined>;
-    close(): Promise<void>;
-}
-export declare class EOUModel {
-    #private;
-    readonly unlikelyThreshold: number;
-    constructor(unlikelyThreshold?: number);
-    supportsLanguage(language?: string): boolean;
-    predictEndOfTurn(chatCtx: llm.ChatContext): Promise<number>;
-}
-export default EOURunner;
-//# sourceMappingURL=turn_detector.d.ts.map

package/dist/turn_detector.d.ts

@@ -1,22 +0,0 @@
-import { InferenceRunner, llm } from '@livekit/agents';
-type RawChatContext = {
-    role: string;
-    content: string;
-}[];
-export declare class EOURunner extends InferenceRunner {
-    #private;
-    static INFERENCE_METHOD: string;
-    constructor();
-    initialize(): Promise<void>;
-    run(data: RawChatContext): Promise<number | undefined>;
-    close(): Promise<void>;
-}
-export declare class EOUModel {
-    #private;
-    readonly unlikelyThreshold: number;
-    constructor(unlikelyThreshold?: number);
-    supportsLanguage(language?: string): boolean;
-    predictEndOfTurn(chatCtx: llm.ChatContext): Promise<number>;
-}
-export default EOURunner;
-//# sourceMappingURL=turn_detector.d.ts.map

package/dist/turn_detector.d.ts.map

@@ -1 +0,0 @@
-{"version":3,"file":"turn_detector.d.ts","sourceRoot":"","sources":["../src/turn_detector.ts"],"names":[],"mappings":"AAMA,OAAO,EAAqB,eAAe,EAAE,GAAG,EAAO,MAAM,iBAAiB,CAAC;AAM/E,KAAK,cAAc,GAAG;IAAE,IAAI,EAAE,MAAM,CAAC;IAAC,OAAO,EAAE,MAAM,CAAA;CAAE,EAAE,CAAC;AAE1D,qBAAa,SAAU,SAAQ,eAAe;;IAC5C,MAAM,CAAC,gBAAgB,SAAyB;;IAoB1C,UAAU;IAIV,GAAG,CAAC,IAAI,EAAE,cAAc,GAAG,OAAO,CAAC,MAAM,GAAG,SAAS,CAAC;IA+BtD,KAAK;CAGZ;AAED,qBAAa,QAAQ;;IACnB,QAAQ,CAAC,iBAAiB,EAAE,MAAM,CAAC;gBAGvB,iBAAiB,SAAO;IAKpC,gBAAgB,CAAC,QAAQ,CAAC,EAAE,MAAM;IAM5B,gBAAgB,CAAC,OAAO,EAAE,GAAG,CAAC,WAAW,GAAG,OAAO,CAAC,MAAM,CAAC;CA4BlE;AAED,eAAe,SAAS,CAAC"}