echogarden 1.0.4 → 1.1.0

package/src/utilities/Utilities.ts

@@ -5,7 +5,7 @@ import { inspect } from 'node:util'
 import { RandomGenerator } from './RandomGenerator.js'
 import { randomUUID, randomBytes } from 'node:crypto'
 import { Logger } from './Logger.js'
-import { ChildProcessWithoutNullStreams } from 'node:child_process'
+import { ChildProcessWithoutNullStreams, exec } from 'node:child_process'
 
 const log = logToStderr
 
@@ -69,6 +69,10 @@ export function simplifyPunctuationCharacters(text: string) {
 		.replaceAll(`：`, `:`)
 		.replaceAll(`；`, `;`)
 		.replaceAll(`。`, `.`)
+
+		.replaceAll(`？`, `?`)
+		.replaceAll(`！`, `!`)
+		.replaceAll(`؟`, `?`)
 }
 
 export function writeToStderr(message: any) {
@@ -427,7 +431,7 @@ export async function resolveModuleMainPath(moduleName: string) {
 	const { resolve } = await import('import-meta-resolve')
 	const { fileURLToPath } = await import('url')
 
-	return fileURLToPath(await resolve(moduleName, import.meta.url))
+	return fileURLToPath(resolve(moduleName, import.meta.url))
 }
 
 export function getWithDefault<T>(value: T | undefined, defaultValue: T) {
@@ -480,89 +484,47 @@ export function getUTF32Chars(str: string) {
 	return { utf32chars, mapping }
 }
 
-export function getRepetitionScoreRelativeToFirstSubstring(tokens: string[] | number[]) {
-	const maxOrder = Math.floor(tokens.length)
+export function getTokenRepetitionScore(tokens: string[] | number[]) {
+	const maxCycleLength = Math.floor(tokens.length / 2)
 
-	const matchCountForOrder: number[] = []
+	const matchLengthForCycleLength: number[] = [0]
 
-	for (let i = 0; i <= maxOrder; i++) {
-		matchCountForOrder.push(0)
-	}
+	for (let cycleLength = 1; cycleLength <= maxCycleLength; cycleLength++) {
+		let matchCount = 0
 
-	for (let offset = 0; offset < tokens.length; offset++) {
-		for (let order = 1; order <= maxOrder; order++) {
-			const referenceToken = tokens[-1 + order]
-			const targetToken = tokens[offset + order]
+		for (let leftIndex = cycleLength; leftIndex < tokens.length; leftIndex++) {
+			const referenceIndex = leftIndex - cycleLength
 
-			if (targetToken == referenceToken) {
-				matchCountForOrder[order] += 1
-			} else {
+			if (tokens[leftIndex] !== tokens[referenceIndex]) {
 				break
 			}
-		}
-	}
-
-	const scores = matchCountForOrder.map((count, index) => count * index)
-
-	let maxScoreOrder = -1
-	let maxScore = -Infinity
-
-	for (let i = 1; i <= matchCountForOrder.length; i++) {
-		const score = scores[i]
-
-		if (score > maxScore) {
-			maxScoreOrder = i
-			maxScore = score
-		}
-	}
-
-	return { maxScore, maxScoreOrder }
-}
-
-export function getConsecutiveRepetitionScoreRelativeToFirstSubstring(tokens: string[] | number[]) {
-	function countRepetitionForLength(len: number) {
-		let count = 0
-
-		for (let offset = len; offset < tokens.length; offset += len) {
-			for (let i = 0; i < len; i++) {
-				if (tokens[i] != tokens[offset + i]) {
-					return count
-				}
-			}
 
-			count += 1
+			matchCount += 1
 		}
 
-		return count
-	}
-
-	const maxLength = Math.floor(tokens.length / 2)
+		const score = matchCount
 
-	const repeatCountForLength: number[] = []
-
-	for (let i = 0; i <= maxLength; i++) {
-		repeatCountForLength.push(0)
+		matchLengthForCycleLength.push(score)
 	}
 
-	for (let targetLen = 1; targetLen <= maxLength; targetLen++) {
-		repeatCountForLength[targetLen] = countRepetitionForLength(targetLen)
-	}
+	let longestMatch = -Infinity
+	let longestCycleRepetition = -Infinity
 
-	const scores = repeatCountForLength.map((count, index) => count * index)
+	for (let i = 1; i <= matchLengthForCycleLength.length; i++) {
+		const matchLength = matchLengthForCycleLength[i]
 
-	let maxScoreLength = -1
-	let maxScore = -Infinity
+		if (matchLength > longestMatch) {
+			longestMatch = matchLength
+		}
 
-	for (let i = 1; i <= repeatCountForLength.length; i++) {
-		const score = scores[i]
+		const cycleCount = matchLength / i
 
-		if (score > maxScore) {
-			maxScoreLength = i
-			maxScore = score
+		if (cycleCount > longestCycleRepetition) {
+			longestCycleRepetition = cycleCount
 		}
 	}
 
-	return { maxScore, maxScoreLength, repeatCountForLength }
+	return { longestMatch, longestCycleRepetition }
 }
 
 export async function resolveModuleScriptPath(moduleName: string) {
@@ -647,3 +609,13 @@ export function getIntegerRange(start: number, end: number): number[] {
 
 	return result
 }
+
+export function containsInvalidCodepoint(str: string) {
+	for (const char of str) {
+		if (char.codePointAt(0) === 65533) {
+			return true
+		}
+	}
+
+	return false
+}

package/src/voice-activity-detection/SileroVAD.ts

@@ -2,10 +2,14 @@ import type * as Onnx from 'onnxruntime-node'
 
 import { concatFloat32Arrays } from '../utilities/Utilities.js'
 import { RawAudio } from '../audio/AudioUtilities.js'
-
-let sileroVad: SileroVAD
-
-export async function detectVoiceActivity(rawAudio: RawAudio, modelPath: string, frameDuration: 30 | 60 | 90) {
+import { OnnxExecutionProvider, getOnnxSessionOptions } from '../utilities/OnnxUtilities.js'
+
+export async function detectVoiceActivity(
+	rawAudio: RawAudio,
+	modelPath: string,
+	frameDuration: 30 | 60 | 90,
+	executionProviders: OnnxExecutionProvider[]) {
+		
 	if (rawAudio.sampleRate != 16000) {
 		throw new Error('Audio sample rate must be 16KHz')
 	}
@@ -14,9 +18,7 @@ export async function detectVoiceActivity(rawAudio: RawAudio, modelPath: string,
 
 	const frameLength = Math.floor(16000 * (frameDuration / 1000))
 
-	if (!sileroVad) {
-		sileroVad = new SileroVAD(modelPath)
-	}
+	const sileroVad = new SileroVAD(modelPath, executionProviders)
 
 	const frameProbabilities: number[] = []
 
@@ -42,10 +44,9 @@ export class SileroVAD {
 	modelStateC?: Onnx.Tensor
 	modelSampleRate?: Onnx.Tensor
 
-	modelPath: string
-
-	constructor(modelPath: string) {
-		this.modelPath = modelPath
+	constructor(
+		public readonly modelPath: string,
+		public readonly executionProviders: OnnxExecutionProvider[]) {
 	}
 
 	async predictAudioFrame(frame: Float32Array) {
@@ -66,6 +67,7 @@ export class SileroVAD {
 
 		return probability
 	}
+
 	private async initializeIfNeeded() {
 		if (this.session) {
 			return
@@ -81,10 +83,8 @@ export class SileroVAD {
 
 		this.modelSampleRate = new Onnx.Tensor('int64', new BigInt64Array([BigInt(16000)]), [])
 
-		const onnxOptions: Onnx.InferenceSession.SessionOptions = {
-			logSeverityLevel: 3
-		}
+		const onnxSessionOptions = getOnnxSessionOptions({ executionProviders: this.executionProviders })
 
-		this.session = await Onnx.InferenceSession.create(this.modelPath, onnxOptions)
+		this.session = await Onnx.InferenceSession.create(this.modelPath, onnxSessionOptions)
 	}
 }

package/dist/utilities/NdArrayUtilities.d.ts

@@ -1,3 +0,0 @@
-import ndarray from 'ndarray';
-export declare function ndarraySoftMax(vector: ndarray.NdArray, temperature?: number): void;
-export declare function ndarrayMedianFilter(vector: ndarray.NdArray, width: number): void;

package/dist/utilities/NdArrayUtilities.js

@@ -1,23 +0,0 @@
-// import ops from 'ndarray-ops'
-import { medianFilter, softmax } from '../math/VectorMath.js';
-export function ndarraySoftMax(vector, temperature = 1.0) {
-    const vectorAsArray = new Array(vector.shape[0]);
-    for (let i = 0; i < vectorAsArray.length; i++) {
-        vectorAsArray[i] = vector.get(i);
-    }
-    const result = softmax(vectorAsArray, temperature);
-    for (let i = 0; i < vectorAsArray.length; i++) {
-        vector.set(i, result[i]);
-    }
-}
-export function ndarrayMedianFilter(vector, width) {
-    const vectorAsArray = new Array(vector.shape[0]);
-    for (let i = 0; i < vectorAsArray.length; i++) {
-        vectorAsArray[i] = vector.get(i);
-    }
-    const result = medianFilter(vectorAsArray, width);
-    for (let i = 0; i < vectorAsArray.length; i++) {
-        vector.set(i, result[i]);
-    }
-}
-//# sourceMappingURL=NdArrayUtilities.js.map

package/dist/utilities/NdArrayUtilities.js.map

@@ -1 +0,0 @@
-{"version":3,"file":"NdArrayUtilities.js","sourceRoot":"","sources":["../../src/utilities/NdArrayUtilities.ts"],"names":[],"mappings":"AACA,gCAAgC;AAChC,OAAO,EAAE,YAAY,EAAE,OAAO,EAAE,MAAM,uBAAuB,CAAA;AAE7D,MAAM,UAAU,cAAc,CAAC,MAAuB,EAAE,WAAW,GAAG,GAAG;IACxE,MAAM,aAAa,GAAG,IAAI,KAAK,CAAC,MAAM,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC,CAAA;IAEhD,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,aAAa,CAAC,MAAM,EAAE,CAAC,EAAE,EAAE,CAAC;QAC/C,aAAa,CAAC,CAAC,CAAC,GAAG,MAAM,CAAC,GAAG,CAAC,CAAC,CAAC,CAAA;IACjC,CAAC;IAED,MAAM,MAAM,GAAG,OAAO,CAAC,aAAa,EAAE,WAAW,CAAC,CAAA;IAElD,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,aAAa,CAAC,MAAM,EAAE,CAAC,EAAE,EAAE,CAAC;QAC/C,MAAM,CAAC,GAAG,CAAC,CAAC,EAAE,MAAM,CAAC,CAAC,CAAC,CAAC,CAAA;IACzB,CAAC;AACF,CAAC;AAED,MAAM,UAAU,mBAAmB,CAAC,MAAuB,EAAE,KAAa;IACzE,MAAM,aAAa,GAAG,IAAI,KAAK,CAAC,MAAM,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC,CAAA;IAEhD,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,aAAa,CAAC,MAAM,EAAE,CAAC,EAAE,EAAE,CAAC;QAC/C,aAAa,CAAC,CAAC,CAAC,GAAG,MAAM,CAAC,GAAG,CAAC,CAAC,CAAC,CAAA;IACjC,CAAC;IAED,MAAM,MAAM,GAAG,YAAY,CAAC,aAAa,EAAE,KAAK,CAAC,CAAA;IAEjD,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,aAAa,CAAC,MAAM,EAAE,CAAC,EAAE,EAAE,CAAC;QAC/C,MAAM,CAAC,GAAG,CAAC,CAAC,EAAE,MAAM,CAAC,CAAC,CAAC,CAAC,CAAA;IACzB,CAAC;AACF,CAAC"}

package/src/utilities/NdArrayUtilities.ts

@@ -1,31 +0,0 @@
-import ndarray from 'ndarray'
-// import ops from 'ndarray-ops'
-import { medianFilter, softmax } from '../math/VectorMath.js'
-
-export function ndarraySoftMax(vector: ndarray.NdArray, temperature = 1.0) {
-	const vectorAsArray = new Array(vector.shape[0])
-
-	for (let i = 0; i < vectorAsArray.length; i++) {
-		vectorAsArray[i] = vector.get(i)
-	}
-
-	const result = softmax(vectorAsArray, temperature)
-
-	for (let i = 0; i < vectorAsArray.length; i++) {
-		vector.set(i, result[i])
-	}
-}
-
-export function ndarrayMedianFilter(vector: ndarray.NdArray, width: number) {
-	const vectorAsArray = new Array(vector.shape[0])
-
-	for (let i = 0; i < vectorAsArray.length; i++) {
-		vectorAsArray[i] = vector.get(i)
-	}
-
-	const result = medianFilter(vectorAsArray, width)
-
-	for (let i = 0; i < vectorAsArray.length; i++) {
-		vector.set(i, result[i])
-	}
-}