echogarden 1.4.3 → 1.5.0

package/src/math/VectorMath.ts

@@ -1,3 +1,5 @@
+import { clip } from "../utilities/Utilities.js"
+
 export function covarianceMatrixOfSamples(samples: number[][], weights?: number[], biased = false) {
 	if (samples.length == 0) {
 		throw new Error('No vectors given')
@@ -175,7 +177,7 @@ export function scaleToSumTo1(vector: number[]) {
 	return scaledVector
 }
 
-export function normalizeVector(vector: number[], kind: 'population' | 'sample' = 'population') {
+export function normalizeVector(vector: ArrayLike<number>, kind: 'population' | 'sample' = 'population') {
 	if (vector.length == 0) {
 		throw new Error('Vector is empty')
 	}
@@ -327,15 +329,15 @@ export function varianceOfVectors(vectors: number[][], kind: 'population' | 'sam
 	return result
 }
 
-export function meanOfVector(vector: number[]) {
+export function meanOfVector(vector: ArrayLike<number>) {
 	if (vector.length == 0) {
-		throw new Error('Vector is empty')
+		return 0
 	}
 
 	return sumVector(vector) / vector.length
 }
 
-export function medianOfVector(vector: number[]) {
+export function medianOfVector(vector: ArrayLike<number>) {
 	if (vector.length == 0) {
 		throw new Error('Vector is empty')
 	}
@@ -343,13 +345,13 @@ export function medianOfVector(vector: number[]) {
 	return vector[Math.floor(vector.length / 2)]
 }
 
-export function stdDeviationOfVector(vector: number[], kind: 'population' | 'sample' = 'population', mean?: number) {
+export function stdDeviationOfVector(vector: ArrayLike<number>, kind: 'population' | 'sample' = 'population', mean?: number) {
 	return Math.sqrt(varianceOfVector(vector, kind, mean))
 }
 
-export function varianceOfVector(vector: number[], kind: 'population' | 'sample' = 'population', mean?: number) {
+export function varianceOfVector(vector: ArrayLike<number>, kind: 'population' | 'sample' = 'population', mean?: number) {
 	if (vector.length == 0) {
-		throw new Error('Vector is empty')
+		return 0
 	}
 
 	const sampleSizeMetric = kind == 'population' || vector.length == 1 ? vector.length : vector.length - 1
@@ -360,8 +362,8 @@ export function varianceOfVector(vector: number[], kind: 'population' | 'sample'
 
 	let result = 0.0
 
-	for (const value of vector) {
-		result += (value - mean) ** 2
+	for (let i = 0; i < vector.length; i++) {
+		result += (vector[i] - mean) ** 2
 	}
 
 	return result / sampleSizeMetric
@@ -434,7 +436,7 @@ export function averageMeanSquaredError(actual: number[][], expected: number[][]
 	return sum / vectorCount
 }
 
-export function meanSquaredError(actual: number[], expected: number[]) {
+export function meanSquaredError(actual: ArrayLike<number>, expected: ArrayLike<number>) {
 	if (actual.length != expected.length) {
 		throw new Error('Vectors are not the same length')
 	}
@@ -454,18 +456,18 @@ export function meanSquaredError(actual: number[], expected: number[]) {
 	return sum / featureCount
 }
 
-export function euclidianDistance(vector1: number[], vector2: number[]) {
+export function euclidianDistance(vector1: ArrayLike<number>, vector2: ArrayLike<number>) {
 	return Math.sqrt(squaredEuclidianDistance(vector1, vector2))
 }
 
-export function squaredEuclidianDistance(vector1: number[], vector2: number[]) {
-	if (vector1.length != vector2.length) {
+export function squaredEuclidianDistance(vector1: ArrayLike<number>, vector2: ArrayLike<number>) {
+	if (vector1.length !== vector2.length) {
 		throw new Error('Vectors are not the same length')
 	}
 
 	const elementCount = vector1.length
 
-	if (elementCount == 0) {
+	if (elementCount === 0) {
 		return 0
 	}
 
@@ -478,16 +480,41 @@ export function squaredEuclidianDistance(vector1: number[], vector2: number[]) {
 	return sum
 }
 
-export function cosineDistance(vector1: number[], vector2: number[]) {
+export function euclidianDistance13Dim(vector1: ArrayLike<number>, vector2: ArrayLike<number>) {
+	return Math.sqrt(squaredEuclidianDistance13Dim(vector1, vector2))
+}
+
+export function squaredEuclidianDistance13Dim(vector1: ArrayLike<number>, vector2: ArrayLike<number>) {
+	// Assumes the input has 13 dimensions (optimized for 13-dimensional MFCC vectors)
+
+	const result =
+		(vector1[0] - vector2[0]) ** 2 +
+		(vector1[1] - vector2[1]) ** 2 +
+		(vector1[2] - vector2[2]) ** 2 +
+		(vector1[3] - vector2[3]) ** 2 +
+		(vector1[4] - vector2[4]) ** 2 +
+		(vector1[5] - vector2[5]) ** 2 +
+		(vector1[6] - vector2[6]) ** 2 +
+		(vector1[7] - vector2[7]) ** 2 +
+		(vector1[8] - vector2[8]) ** 2 +
+		(vector1[9] - vector2[9]) ** 2 +
+		(vector1[10] - vector2[10]) ** 2 +
+		(vector1[11] - vector2[11]) ** 2 +
+		(vector1[12] - vector2[12]) ** 2
+
+	return result
+}
+
+export function cosineDistance(vector1: ArrayLike<number>, vector2: ArrayLike<number>) {
 	return 1 - cosineSimilarity(vector1, vector2)
 }
 
-export function cosineSimilarity(vector1: number[], vector2: number[]) {
-	if (vector1.length != vector2.length) {
+export function cosineSimilarity(vector1: ArrayLike<number>, vector2: ArrayLike<number>) {
+	if (vector1.length !== vector2.length) {
 		throw new Error('Vectors are not the same length')
 	}
 
-	if (vector1.length == 0) {
+	if (vector1.length === 0) {
 		return 0
 	}
 
@@ -499,42 +526,28 @@ export function cosineSimilarity(vector1: number[], vector2: number[]) {
 	let squaredMagnitude2 = 0.0
 
 	for (let i = 0; i < elementCount; i++) {
-		dotProduct += vector1[i] * vector2[i]
-
-		squaredMagnitude1 += vector1[i] ** 2
-		squaredMagnitude2 += vector2[i] ** 2
-	}
-
-	const result = dotProduct / (Math.sqrt(squaredMagnitude1) * Math.sqrt(squaredMagnitude2) + 1e-40)
-
-	return zeroIfNaN(result)
-}
+		const element1 = vector1[i]
+		const element2 = vector2[i]
 
-export function minkowskiDistance(vector1: number[], vector2: number[], power: number) {
-	if (vector1.length != vector2.length) {
-		throw new Error('Vectors are not the same length')
-	}
-
-	const elementCount = vector1.length
+		dotProduct += element1 * element2
 
-	if (elementCount == 0) {
-		return 0
+		squaredMagnitude1 += element1 ** 2
+		squaredMagnitude2 += element2 ** 2
 	}
 
-	let sum = 0.0
+	let result = dotProduct / (Math.sqrt(squaredMagnitude1) * Math.sqrt(squaredMagnitude2) + 1e-40)
 
-	for (let i = 0; i < elementCount; i++) {
-		sum += Math.abs(vector1[i] - vector2[i]) ** power
-	}
+	result = zeroIfNaN(result)
+	result = clip(result, -1.0, 1.0)
 
-	return sum ** (1 / power)
+	return result
 }
 
-export function cosineDistancePrecomputedMagnitudes(vector1: number[], vector2: number[], magnitude1: number, magnitude2: number) {
+export function cosineDistancePrecomputedMagnitudes(vector1: ArrayLike<number>, vector2: ArrayLike<number>, magnitude1: number, magnitude2: number) {
 	return 1 - cosineSimilarityPrecomputedMagnitudes(vector1, vector2, magnitude1, magnitude2)
 }
 
-export function cosineSimilarityPrecomputedMagnitudes(vector1: number[], vector2: number[], magnitude1: number, magnitude2: number) {
+export function cosineSimilarityPrecomputedMagnitudes(vector1: ArrayLike<number>, vector2: ArrayLike<number>, magnitude1: number, magnitude2: number) {
 	if (vector1.length != vector2.length) {
 		throw new Error('Vectors are not the same length')
 	}
@@ -551,11 +564,33 @@ export function cosineSimilarityPrecomputedMagnitudes(vector1: number[], vector2
 		dotProduct += vector1[i] * vector2[i]
 	}
 
-	const result = dotProduct / (magnitude1 * magnitude2)
+	let result = dotProduct / (magnitude1 * magnitude2 + 1e-40)
 
-	return zeroIfNaN(result)
+	result = zeroIfNaN(result)
+	result = clip(result, -1.0, 1.0)
+
+	return result
 }
 
+export function minkowskiDistance(vector1: ArrayLike<number>, vector2: ArrayLike<number>, power: number) {
+	if (vector1.length != vector2.length) {
+		throw new Error('Vectors are not the same length')
+	}
+
+	const elementCount = vector1.length
+
+	if (elementCount == 0) {
+		return 0
+	}
+
+	let sum = 0.0
+
+	for (let i = 0; i < elementCount; i++) {
+		sum += Math.abs(vector1[i] - vector2[i]) ** power
+	}
+
+	return sum ** (1 / power)
+}
 
 export function subtractVectors(vector1: number[], vector2: number[]) {
 	if (vector1.length != vector2.length) {
@@ -571,7 +606,7 @@ export function subtractVectors(vector1: number[], vector2: number[]) {
 	return result
 }
 
-export function sumVector(vector: number[]) {
+export function sumVector(vector: ArrayLike<number>) {
 	let result = 0.0
 
 	for (let i = 0; i < vector.length; i++) {
@@ -581,7 +616,7 @@ export function sumVector(vector: number[]) {
 	return result
 }
 
-export function dotProduct(vector1: number[], vector2: number[]) {
+export function dotProduct(vector1: ArrayLike<number>, vector2: ArrayLike<number>) {
 	if (vector1.length != vector2.length) {
 		throw new Error('Vectors are not the same length')
 	}
@@ -597,7 +632,7 @@ export function dotProduct(vector1: number[], vector2: number[]) {
 	return result
 }
 
-export function magnitude(vector: number[]) {
+export function magnitude(vector: ArrayLike<number>) {
 	const featureCount = vector.length
 
 	let squaredMagnitude = 0.0
@@ -609,11 +644,11 @@ export function magnitude(vector: number[]) {
 	return Math.sqrt(squaredMagnitude)
 }
 
-export function maxValue(vector: number[]) {
+export function maxValue(vector: ArrayLike<number>) {
 	return vector[indexOfMax(vector)]
 }
 
-export function indexOfMax(vector: number[]) {
+export function indexOfMax(vector: ArrayLike<number>) {
 	if (vector.length == 0) {
 		return -1
 	}
@@ -631,11 +666,11 @@ export function indexOfMax(vector: number[]) {
 	return result
 }
 
-export function minValue(vector: number[]) {
+export function minValue(vector: ArrayLike<number>) {
 	return vector[indexOfMin(vector)]
 }
 
-export function indexOfMin(vector: number[]) {
+export function indexOfMin(vector: ArrayLike<number>) {
 	let minValue = Infinity
 	let result = -1
 

package/src/recognition/WhisperCppSTT.ts

@@ -13,7 +13,7 @@ import { splitToLines } from '../nlp/Segmentation.js'
 import { extendDeep } from '../utilities/ObjectUtilities.js'
 import { formatLanguageCodeWithName, getShortLanguageCode } from '../utilities/Locale.js'
 import { loadPackage } from '../utilities/PackageManager.js'
-import { detectSpeechLanguageByParts } from '../api/LanguageDetection.js'
+import { detectSpeechLanguageByParts } from '../api/SpeechLanguageDetection.js'
 
 export async function recognize(
 	sourceRawAudio: RawAudio,
@@ -201,7 +201,7 @@ export async function recognize(
 
 export async function detectLanguage(sourceRawAudio: RawAudio, modelName: WhisperModelName, modelPath: string) {
 	if (sourceRawAudio.sampleRate != 16000) {
-		throw new Error('Source audio must have a sampling rate of 16000')
+		throw new Error('Source audio must have a sample rate of 16000')
 	}
 
 	async function detectLanguageForPart(partAudio: RawAudio) {

package/src/recognition/WhisperSTT.ts

@@ -17,7 +17,7 @@ import { formatLanguageCodeWithName, getShortLanguageCode, languageCodeToName }
 import { loadPackage } from '../utilities/PackageManager.js'
 import chalk from 'chalk'
 import { XorShift32RNG } from '../utilities/RandomGenerator.js'
-import { detectSpeechLanguageByParts } from '../api/LanguageDetection.js'
+import { detectSpeechLanguageByParts } from '../api/SpeechLanguageDetection.js'
 import { type Tiktoken } from 'tiktoken/lite'
 import { isPunctuation, isWhitespace, isWord, splitToSentences, splitToWords } from '../nlp/Segmentation.js'
 import { medianOf5Filter } from '../math/MedianFilter.js'
@@ -35,7 +35,7 @@ export async function recognize(
 	options = extendDeep(defaultWhisperOptions, options)
 
 	if (sourceRawAudio.sampleRate != 16000) {
-		throw new Error('Source audio must have a sampling rate of 16000 Hz')
+		throw new Error('Source audio must have a sample rate of 16000 Hz')
 	}
 
 	sourceLanguage = getShortLanguageCode(sourceLanguage)
@@ -87,7 +87,7 @@ export async function align(
 	options = extendDeep(defaultWhisperAlignmentOptions, options)
 
 	if (sourceRawAudio.sampleRate != 16000) {
-		throw new Error('Source audio must have a sampling rate of 16000')
+		throw new Error('Source audio must have a sample rate of 16000 Hz')
 	}
 
 	sourceLanguage = getShortLanguageCode(sourceLanguage)
@@ -128,7 +128,7 @@ export async function alignEnglishTranslation(
 	options = extendDeep(defaultWhisperAlignmentOptions, options)
 
 	if (sourceRawAudio.sampleRate != 16000) {
-		throw new Error('Source audio must have a sampling rate of 16000')
+		throw new Error('Source audio must have a sample rate of 16000 Hz')
 	}
 
 	sourceLanguage = getShortLanguageCode(sourceLanguage)
@@ -167,7 +167,7 @@ export async function detectLanguage(
 	options = extendDeep(defaultWhisperLanguageDetectionOptions, options)
 
 	if (sourceRawAudio.sampleRate != 16000) {
-		throw new Error('Source audio must have a sampling rate of 16000')
+		throw new Error('Source audio must have a sample rate of 16000 Hz')
 	}
 
 	if (!isMultilingualModel(modelName)) {
@@ -213,7 +213,7 @@ export async function detectVoiceActivity(
 	options = extendDeep(defaultWhisperVADOptions, options)
 
 	if (sourceRawAudio.sampleRate != 16000) {
-		throw new Error('Source audio must have a sampling rate of 16000')
+		throw new Error('Source audio must have a sample rate of 16000 Hz')
 	}
 
 	if (options.temperature! < 0) {
@@ -1089,6 +1089,10 @@ export class Whisper {
 		const maxAudioSamples = sampleRate * 30
 		const maxAudioFrames = 3000
 
+		if (sampleRate != 16000) {
+			throw new Error('Audio must have a sample rate of 16000 Hz')
+		}
+
 		if (audioSamples.length > maxAudioSamples) {
 			throw new Error(`Audio part is longer than 30 seconds`)
 		}

package/src/server/Client.ts

@@ -7,9 +7,10 @@ import { OpenPromise } from '../utilities/OpenPromise.js'
 import { AudioSourceParam, RawAudio } from '../audio/AudioUtilities.js'
 import { AlignmentOptions, AlignmentResult } from '../api/Alignment.js'
 import { RecognitionOptions, RecognitionResult } from '../api/Recognition.js'
-import { SpeechTranslationOptions, SpeechTranslationResult } from '../api/Translation.js'
+import { SpeechTranslationOptions, SpeechTranslationResult } from '../api/SpeechTranslation.js'
 import { Worker as WorkerThread } from 'node:worker_threads'
-import { SpeechLanguageDetectionOptions, SpeechLanguageDetectionResult, TextLanguageDetectionOptions, TextLanguageDetectionResult } from '../api/LanguageDetection.js'
+import { SpeechLanguageDetectionOptions, SpeechLanguageDetectionResult } from '../api/SpeechLanguageDetection.js'
+import { TextLanguageDetectionOptions, TextLanguageDetectionResult } from '../api/TextLanguageDetection.js'
 
 const log = logToStderr
 

package/src/server/Worker.ts

@@ -4,13 +4,14 @@ import { logToStderr, yieldToEventLoop } from '../utilities/Utilities.js'
 import { AudioSourceParam } from '../audio/AudioUtilities.js'
 import { RecognitionOptions, RecognitionResult, recognize } from '../api/Recognition.js'
 import { AlignmentOptions, AlignmentResult, align } from '../api/Alignment.js'
-import { SpeechTranslationOptions, SpeechTranslationResult, translateSpeech } from '../api/Translation.js'
+import { SpeechTranslationOptions, SpeechTranslationResult, translateSpeech } from '../api/SpeechTranslation.js'
 import { resetActiveLogger } from '../utilities/Logger.js'
 import { writeToStderr } from '../utilities/Utilities.js'
 import { Worker, SHARE_ENV } from 'node:worker_threads'
-import { SpeechLanguageDetectionOptions, SpeechLanguageDetectionResult, TextLanguageDetectionOptions, TextLanguageDetectionResult, detectSpeechLanguage, detectTextLanguage } from '../api/LanguageDetection.js'
+import { SpeechLanguageDetectionOptions, SpeechLanguageDetectionResult, detectSpeechLanguage } from '../api/SpeechLanguageDetection.js'
 import chalk from 'chalk'
 import { resolveToModuleRootDir } from '../utilities/PathUtilities.js'
+import { TextLanguageDetectionOptions, TextLanguageDetectionResult, detectTextLanguage } from '../api/TextLanguageDetection.js'
 
 const log = logToStderr
 

package/src/source-separation/MDXNetSourceSeparation.ts

@@ -32,7 +32,7 @@ export class MDXNet {
 		}
 
 		if (rawAudio.sampleRate != 44100) {
-			throw new Error(`Input audio must have a 44100 Hz sampling rate`)
+			throw new Error(`Input audio must have a sample rate of 44100 Hz`)
 		}
 
 		await this.initializeSessionIfNeeded()

package/src/speech-embeddings/WavToVec2BertFeatureEmbeddings.ts

@@ -0,0 +1,107 @@
+import type * as Onnx from 'onnxruntime-node'
+import { OnnxExecutionProvider, getOnnxSessionOptions } from '../utilities/OnnxUtilities.js';
+
+import { RawAudio } from "../audio/AudioUtilities.js";
+import { computeMelSpectogram } from "../dsp/MelSpectogram.js";
+import { Logger } from '../utilities/Logger.js';
+import { concatFloat32Arrays, splitFloat32Array } from '../utilities/Utilities.js';
+import { applyEmphasis } from '../dsp/MFCC.js';
+
+export function computeEmbeddings(audioSamples: RawAudio, modelFilePath: string, executionProviders: OnnxExecutionProvider[]) {
+	const wav2vecBert = new Wav2Vec2BertFeatureEmbeddings(
+		modelFilePath,
+		executionProviders,
+	)
+
+	const result = wav2vecBert.computeEmbeddings(audioSamples)
+
+	return result
+}
+
+export class Wav2Vec2BertFeatureEmbeddings {
+	session?: Onnx.InferenceSession
+
+	constructor(
+		public readonly modelFilePath: string,
+		public readonly executionProviders: OnnxExecutionProvider[]) {
+	}
+
+	async computeEmbeddings(rawAudio: RawAudio) {
+		const logger = new Logger()
+
+		rawAudio.audioChannels[0] = applyEmphasis(rawAudio.audioChannels[0], 0.97)
+
+		const { melSpectogram } = await computeMelSpectogram(
+			rawAudio,
+			512,
+			400,
+			160,
+			80,
+			20,
+			8000,
+			'povey')
+
+		// Ensure even length
+		if (melSpectogram.length % 2 != 0) {
+			melSpectogram.push(new Float32Array(80))
+		}
+
+		// Normalize filterbanks
+		for (let filterbankIndex = 0; filterbankIndex < 80; filterbankIndex++) {
+			let sum = 0
+			let sumOfSquares = 0
+
+			for (let i = 0; i < melSpectogram.length; i++) {
+				const value = melSpectogram[i][filterbankIndex]
+
+				sum += value
+				sumOfSquares += value ** 2
+			}
+
+			const mean = sum / melSpectogram.length
+			const normalizationFactor = 1 / (Math.sqrt(sumOfSquares / melSpectogram.length) + 1e-40)
+
+			for (let i = 0; i < melSpectogram.length; i++) {
+				melSpectogram[i][filterbankIndex] -= mean
+				melSpectogram[i][filterbankIndex] *= normalizationFactor
+			}
+		}
+
+		// Flatten
+		const flattenedMelSpectogram = concatFloat32Arrays(melSpectogram)
+
+		// Initialize session
+		await this.initializeSessionIfNeeded()
+
+		const session = this.session!
+
+		const Onnx = await import('onnxruntime-node')
+
+		const inputTensor = new Onnx.Tensor('float32', flattenedMelSpectogram, [1, melSpectogram.length / 2, 80 * 2])
+
+		const attentionMask = new Int32Array(melSpectogram.length / 2).fill(1)
+		const attentionMaskTensor = new Onnx.Tensor('int32', attentionMask, [1, attentionMask.length])
+
+		// Run inference
+		const outputs = await session.run({ 'input_features': inputTensor, 'attention_mask': attentionMaskTensor })
+
+		// Return output
+		const lastHiddenStateData = outputs['last_hidden_state'].data as Float32Array
+
+		const outputEmbeddings = splitFloat32Array(lastHiddenStateData, outputs['last_hidden_state'].dims[2])
+
+		return outputEmbeddings
+	}
+
+	private async initializeSessionIfNeeded() {
+		if (this.session) {
+			return
+		}
+
+		const Onnx = await import('onnxruntime-node')
+
+		const onnxSessionOptions = getOnnxSessionOptions({ executionProviders: this.executionProviders })
+
+		this.session = await Onnx.InferenceSession.create(this.modelFilePath, onnxSessionOptions)
+	}
+}

package/src/speech-language-detection/SileroLanguageDetection.ts

@@ -3,9 +3,10 @@ import { softmax } from '../math/VectorMath.js'
 import { Logger } from '../utilities/Logger.js'
 import { RawAudio } from '../audio/AudioUtilities.js'
 import { readAndParseJsonFile } from '../utilities/FileSystem.js'
-import { detectSpeechLanguageByParts, type LanguageDetectionResults } from '../api/LanguageDetection.js'
+import { detectSpeechLanguageByParts } from '../api/SpeechLanguageDetection.js'
 import { languageCodeToName } from '../utilities/Locale.js'
 import { OnnxExecutionProvider, getOnnxSessionOptions } from '../utilities/OnnxUtilities.js'
+import { LanguageDetectionResults } from '../api/LanguageDetectionCommon.js'
 
 export async function detectLanguage(
 	rawAudio: RawAudio,

package/src/synthesis/EspeakTTS.ts

@@ -140,13 +140,11 @@ export async function preprocessAndSynthesize(text: string, language: string, es
 }
 
 export async function synthesizeFragments(fragments: string[], espeakOptions: EspeakOptions) {
-	const logger = new Logger()
-
 	espeakOptions = extendDeep(defaultEspeakOptions, espeakOptions)
 
-	const sampleRate = await getSampleRate()
+	const voice = espeakOptions.voice
 
-	//fragments = fragments.filter(fragment => fragment.trim() != '')
+	const sampleRate = await getSampleRate()
 
 	if (fragments.length === 0) {
 		return {
@@ -156,9 +154,11 @@ export async function synthesizeFragments(fragments: string[], espeakOptions: Es
 		}
 	}
 
+	const canInsertSeparators = !['roa/an', 'art/eo', 'trk/ky', 'zlw/pl', 'zle/uk'].includes(voice)
+
 	let textWithMarkers: string
 
-	if (espeakOptions.insertSeparators || !espeakOptions.voice.endsWith('/pl')) {
+	if (canInsertSeparators) {
 		textWithMarkers = `() | `
 	} else {
 		textWithMarkers = `() `
@@ -173,7 +173,7 @@ export async function synthesizeFragments(fragments: string[], espeakOptions: Es
 			.replaceAll('<', '&lt;')
 			.replaceAll('>', '&gt;')
 
-		if (espeakOptions.insertSeparators) {
+		if (espeakOptions.insertSeparators && canInsertSeparators) {
 			const separator = ` | `
 
 			textWithMarkers += `<mark name="s-${i}"/>${separator}${fragment}${separator}<mark name="e-${i}"/>`
@@ -186,8 +186,6 @@ export async function synthesizeFragments(fragments: string[], espeakOptions: Es
 		}
 	}
 
-	//log(textWithMarkers)
-
 	const { rawAudio, events } = await synthesize(textWithMarkers, { ...espeakOptions, ssml: true })
 
 	// Add first marker if missing

package/src/tests/Test.ts

@@ -1,6 +1,5 @@
-import { getTokenRepetitionScore, logToStderr, setupProgramTerminationListeners, writeToStderr } from '../utilities/Utilities.js'
 import { makeTarballsForInstalledPackages } from '../utilities/TarballMaker.js'
-import { testEspeakSynthesisWithPrePhonemizedInputs, testKirshenbaumPhonemization } from '../synthesis/EspeakTTS.js'
+import { logToStderr, setupProgramTerminationListeners } from '../utilities/Utilities.js'
 
 const log = logToStderr
 
@@ -9,16 +8,6 @@ setupProgramTerminationListeners()
 //process.env.NODE_TLS_REJECT_UNAUTHORIZED = '0'
 //process.env.http_proxy = 'http://localhost:8080'
 
-//const testText = `After a while, finding that nothing more happened, she decided on going into the garden at once; but, alas for poor Alice! when she got to the door, she found she had forgotten the little golden key, and when she went back to the table for it, she found she could not possibly reach it: she could see it quite plainly through the glass, and she tried her best to climb up one of the legs of the table, but it was too slippery; and when she had tired herself out with trying, the poor little thing sat down and cried.`
-//await testKirshenbaumPhonemization(testText)
-
-//getRepetitionScoreRelativeToFirstSubstring(['a', 'b', 'c', 'a', 'b', 'c'])
-//getRepetitionScoreRelativeToFirstSubstring(['a', 'b', 'a', 'd', 'a', 'b', 'a', 'd'])
-//getRepetitionScoreRelativeToFirstSubstring(['a', 'b', 'a', 'b', 'c', 'a', 'b', 'a', 'b'])
-//getRepetitionScoreRelativeToFirstSubstring(['a', 'a', 'a', 'b', 'a', 'a', 'a', 'b'])
-//getRepetitionScoreRelativeToFirstSubstring(['a', 'b', 'a', 'c', 'a', 'b', 'a', 'c', 'a'])
-//getRepetitionScoreRelativeToFirstSubstring(['a', 'a', 'a', 'b', 'b', 'a', 'a', 'a', 'b'])
-
 //await makeTarballsForInstalledPackages(true)
 
 process.exit(0)

package/src/text-language-detection/FastTextLanguageDetection.ts

@@ -1,8 +1,8 @@
 import path from 'path'
-import type { LanguageDetectionResults } from '../api/LanguageDetection.js'
 import { languageCodeToName } from '../utilities/Locale.js'
 import { OpenPromise } from '../utilities/OpenPromise.js'
 import { resolveModuleMainPath, roundToDigits } from '../utilities/Utilities.js'
+import { LanguageDetectionResults } from '../api/LanguageDetectionCommon.js'
 
 let fastTextLanguageDetectionModel: any
 

package/src/text-language-detection/TinyLDLanguageDetection.ts

@@ -1,6 +1,6 @@
-import type { LanguageDetectionResults } from '../api/LanguageDetection.js'
 import { detectAll } from 'tinyld'
 import { languageCodeToName } from '../utilities/Locale.js'
+import { LanguageDetectionResults } from '../api/LanguageDetectionCommon.js'
 
 export async function detectLanguage(text: string) {
 	const tinyldResults = detectAll(text)

package/src/text-translation/NLLBTextTranslation.ts

@@ -6,7 +6,7 @@ export async function translateText(sourceText: string, sourceLanguage: string,
 
 	const { AutoTokenizer,  M2M100ForConditionalGeneration } = await import('@echogarden/transformers-nodejs-lite')
 
-	const modelPath = await loadPackage(`xenova-nllb-200-distilled-600M-quantized`)
+	const modelPath = await loadPackage(`xenova-nllb-200-distilled-600M-q8`)
 
 	const tokenizer = await AutoTokenizer.from_pretrained(modelPath)
 	const model = await M2M100ForConditionalGeneration.from_pretrained(modelPath)

package/src/utilities/Locale.ts

@@ -22,7 +22,7 @@ export function formatLanguageCodeWithName(languageCode: string, styleId: 1 | 2
 	}
 }
 
-export async function normalizeIdentifierToLangaugeCode(langIdentifier: string) {
+export async function normalizeIdentifierToLanguageCode(langIdentifier: string) {
 	const result = await parseLangIdentifier(langIdentifier)
 
 	return result.Name

package/src/utilities/OnnxUtilities.ts

@@ -65,4 +65,4 @@ export interface OnnxSessionOptions {
 	logSeverityLevel?: 0 | 1 | 2 | 3 | 4
 }
 
-export type OnnxExecutionProvider = 'cpu' | 'dml'
+export type OnnxExecutionProvider = 'cpu' | 'dml' | 'cuda'

package/src/utilities/PackageManager.ts

@@ -183,6 +183,13 @@ const packageVersionTagResolutionLookup: { [packageName: string]: string } = {
 	'whisper.cpp-binaries-windows-x64-cublas-12.4.0-latest-patched': '20240409',
 	'whisper.cpp-binaries-windows-x64-cublas-11.8.0-latest-patched': '20240409',
 
-	'xenova-multilingual-e5-small-quantized': '20240504',
-	'xenova-nllb-200-distilled-600M-quantized': '20240505',
+	'xenova-multilingual-e5-small-q8': '20240504',
+	'xenova-nllb-200-distilled-600M-q8': '20240505',
+	'xenova-multilingual-e5-small-fp16': '20240514',
+	'xenova-multilingual-e5-base-fp16': '20240514',
+	'xenova-multilingual-e5-base-q8': '20240514',
+	'xenova-multilingual-e5-large-q8': '20240514',
+
+	'w2v-bert-2.0-int8': '20240517',
+	'w2v-bert-2.0-uint8': '20240517',
 }