echogarden 2.8.5 → 2.8.7

package/docs/Tasklist.md

@@ -76,7 +76,8 @@
 * `speak-youtube-subtitles`: To speak the subtitles of a YouTube video
 
 ### API
-* Validate timelines to ensure timestamps are always increasing: no negative timestamps or timestamps over the duration of the audio. No sentences without words, etc. and correct if needed
+* Allow callers from API to cancel a task via `AbortController` and `AbortSignal`
+* Validate timelines to ensure timestamps are always increasing: no undefined timestamps, no negative timestamps, out-of-order timestamps, or timestamps over the duration of the audio. No sentences without words, etc. Missing or incorrect word offsets, etc.
 * Add support for phrases in timelines
 * Accept voice list caching options in `SynthesisOptions`
 
@@ -100,7 +101,7 @@
 ### Synthesis
 * Option to disable alignment (only for some engines). Alternative: use a low granularity DTW setting that is very fast to compute
 * Find places to add commas (",") to improve speech fluency. VITS voices don't normally add speech breaks if there is no punctuation
-* An isolated dash " - " can be converted to a " , " to ensure there's a break in the speech
+* An isolated dash " - " can maybe be converted to a " , " to ensure there's a break in the speech
 * Find a way to manually reset voice list cache
 * When synthesized text isn't pre-split to sentences, apply sentence splits by using the existing method to convert the output of word timelines to sentence/segment timelines
 * Some `sapi` voices and `msspeech` languages output phones that are converted to Microsoft alphabet, not IPA symbols. Try to see if these can be translated to IPA
@@ -126,11 +127,15 @@
 * Option to add POS tags to timeline, if available
 
 ### Synthesis / VITS
-* Allow limiting how many models are cached in memory
+* Allow limiting how many models are cached in memory.
+* Ensure that caching behaves correctly when the same model is used with different execution providers
 * Custom model paths (decide how to implement)
 * Pull voice list from JSON file, or based on URL? Is that a good idea?
 * Add speaker names to voice list somehow
 
+### Synthesis / Kokoro
+* Ensure that caching behaves correctly when the same model is used with different execution providers
+
 ### Synthesis / Azure Cognitive Services
 * Currently, when input is set to be SSML, it is wrapped in a `<speak>` tag. Handle the case where the user made their own SSML document wrapped with a `<speak>` tag as well. Currently, it may send invalid input to Azure
 
@@ -139,8 +144,8 @@
 
 ### Recognition / Whisper
 * Whisper's Chinese and Japanese output can be split into words in a more accurate way. Consider using a dedicated segmentation library to perform the segmentation in character sequences that have no punctuation characters to aid on guessing word boundaries
-* Cache last model (if enough memory is available)
-* The segment output can be used to split into segments, otherwise it is possible to try to guess using pause lengths or voice activity detection
+* Cache last model (if enough memory is available). Ensure that caching works when switching between  different execution providers
+* Whisper timestamp tokens can be used to split into segments, otherwise it is possible to try to guess using pause lengths or voice activity detection
 * Bring back the option to use eSpeak DTW based alignment on segments, as an alternative approach
 
 ### Alignment

package/package.json

@@ -1,6 +1,6 @@
 {
 	"name": "echogarden",
-	"version": "2.8.5",
+	"version": "2.8.7",
 	"description": "An easy-to-use speech toolset. Includes tools for synthesis, recognition, alignment, speech translation, language detection, source separation and more.",
 	"author": "Rotem Dan",
 	"license": "GPL-3.0",
@@ -54,20 +54,20 @@
 	"dependencies": {
 		"@aws-sdk/client-polly": "~3.812.0",
 		"@aws-sdk/client-transcribe-streaming": "~3.812.0",
-		"@echogarden/audio-io": "~0.4.1",
+		"@echogarden/audio-io": "~0.4.2",
 		"@echogarden/espeak-ng-emscripten": "~0.3.3",
 		"@echogarden/fasttext-wasm": "~0.1.0",
 		"@echogarden/flite-wasi": "~0.1.1",
 		"@echogarden/fvad-wasm": "~0.2.0",
 		"@echogarden/gnuspeech-wasm": "~0.2.1",
-		"@echogarden/icu-segmentation-wasm": "~0.2.1",
+		"@echogarden/icu-segmentation-wasm": "~0.2.2",
 		"@echogarden/pffft-wasm": "~0.4.2",
 		"@echogarden/rnnoise-wasm": "~0.2.0",
 		"@echogarden/rubberband-wasm": "~0.2.0",
 		"@echogarden/sonic-wasm": "~0.2.0",
 		"@echogarden/speex-resampler-wasm": "~0.3.0",
 		"@echogarden/svoxpico-wasm": "~0.2.0",
-		"@echogarden/text-segmentation": "~0.3.1",
+		"@echogarden/text-segmentation": "~0.3.3",
 		"@echogarden/transformers-nodejs-lite": "~2.17.1-lite.4",
 		"@echogarden/wave-codec": "~0.3.3",
 		"@mozilla/readability": "~0.6.0",
@@ -82,10 +82,10 @@
 		"jsdom": "~26.1.0",
 		"json5": "~2.2.3",
 		"kuromoji": "~0.1.2",
-		"microsoft-cognitiveservices-speech-sdk": "~1.43.1",
+		"microsoft-cognitiveservices-speech-sdk": "~1.44.0",
 		"msgpack-lite": "~0.1.26",
 		"onnxruntime-node": "~1.21.1",
-		"openai": "~4.100.0",
+		"openai": "~4.102.0",
 		"regexp-composer": "~0.3.0",
 		"sam-js": "~0.3.1",
 		"strip-ansi": "~7.1.0",
@@ -114,7 +114,7 @@
 		"@types/graceful-fs": "~4.1.9",
 		"@types/jsdom": "~21.1.7",
 		"@types/msgpack-lite": "~0.1.11",
-		"@types/node": "~22.15.18",
+		"@types/node": "~22.15.21",
 		"@types/tar": "~6.1.13",
 		"@types/ws": "~8.18.1",
 		"ts-json-schema-generator": "~2.4.0",

package/src/math/VectorMath.ts

@@ -345,7 +345,11 @@ export function medianOfVector(vector: ArrayLike<number>) {
 		throw new Error('Vector is empty')
 	}
 
-	return vector[Math.floor(vector.length / 2)]
+	const sortedArray = Array.from(vector).sort((a, b) => a - b)
+
+	const median = sortedArray[Math.floor(sortedArray.length / 2)]
+
+	return median
 }
 
 export function stdDeviationOfVector(vector: ArrayLike<number>, kind: 'population' | 'sample' = 'population', mean?: number) {

package/src/nlp/Segmentation.ts

@@ -10,11 +10,14 @@ import { splitJapaneseTextToWords_Kuromoji } from './JapaneseSegmentation.js'
 const log = logToStderr
 
 export const wordCharacterRegExp = /[\p{Letter}\p{Number}]/u
-export const emojiCharacterRegExp = /[\p{Emoji}]/u
+
+// See: https://mathiasbynens.be/notes/es-unicode-property-escapes
+export const emojiSequenceRegExp = /\p{Emoji_Modifier_Base}\p{Emoji_Modifier}?|\p{Emoji_Presentation}|\p{Emoji}\uFE0F/u
+
 export const punctuationRegExp = /[\p{Punctuation}]/u
 
 export const phraseSeparators = [',', '、', '，', '،', ';', '；', ':', '：', '—']
-export const symbolWords = ['$', '€', '¢', '£', '¥', '©', '®', '™', '%', '&', '#', '~', '@', '+', '±', '÷', '/', '*', '×', '=', '¼', '½', '¾']
+export const symbolWords = ['$', '€', '¢', '£', '¥', '©', '®', '™', '%', '&', '#', '~', '@', '+', '±', '÷', '/', '\\', '^', '*', '×', '=', '¼', '½', '¾']
 
 ///////////////////////////////////////////////////////////////////////////////////////////////
 // Predicates
@@ -36,7 +39,7 @@ export function includesPunctuation(str: string) {
 }
 
 export function includesEmoji(str: string) {
-	return emojiCharacterRegExp.test(str?.trim())
+	return emojiSequenceRegExp.test(str?.trim())
 }
 
 export function isWhitespace(str: string) {

package/src/recognition/WhisperSTT.ts

@@ -3,7 +3,7 @@ import type * as Onnx from 'onnxruntime-node'
 import { Logger } from '../utilities/Logger.js'
 import { computeMelSpectogramUsingFilterbanks, Filterbank } from '../dsp/MelSpectogram.js'
 import { clip, getIntegerRange, getTopKIndexes, splitFloat32Array, yieldToEventLoop } from '../utilities/Utilities.js'
-import { indexOfMax, logOfVector, logSumExp, meanOfVector, softmax, sumOfSquaresForVector, sumVector } from '../math/VectorMath.js'
+import { indexOfMax, logOfVector, logSumExp, meanOfVector, medianOfVector, softmax, sumOfSquaresForVector, sumVector } from '../math/VectorMath.js'
 
 import { alignDTWWindowed } from '../alignment/DTWSequenceAlignmentWindowed.js'
 import { extendDeep } from '../utilities/ObjectUtilities.js'
@@ -11,7 +11,7 @@ import { Timeline, TimelineEntry } from '../utilities/Timeline.js'
 import { AlignmentPath } from '../alignment/SpeechAlignment.js'
 import { getRawAudioDuration, RawAudio, sliceRawAudio } from '../audio/AudioUtilities.js'
 import { readFileAsUtf8 } from '../utilities/FileSystem.js'
-import type { LanguageDetectionResults } from '../api/API.js'
+import { logLevelGreaterOrEqualTo, type LanguageDetectionResults } from '../api/API.js'
 import { formatLanguageCodeWithName, getShortLanguageCode, languageCodeToName } from '../utilities/Locale.js'
 import { loadPackage } from '../utilities/PackageManager.js'
 import chalk from 'chalk'
@@ -25,6 +25,7 @@ import { dmlProviderAvailable, getOnnxSessionOptions, makeOnnxLikeFloat32Tensor,
 import { murmurHash3_int32Input } from '../utilities/Hashing.js'
 import { containsInvalidCodepoint, getTokenRepetitionScore } from '../utilities/StringUtilities.js'
 import { joinPath } from '../utilities/PathUtilities.js'
+import { Timer } from '../utilities/Timer.js'
 
 export async function recognize(
 	sourceRawAudio: RawAudio,
@@ -36,7 +37,7 @@ export async function recognize(
 
 	options = extendDeep(defaultWhisperOptions, options)
 
-	if (sourceRawAudio.sampleRate != 16000) {
+	if (sourceRawAudio.sampleRate !== 16000) {
 		throw new Error('Source audio must have a sample rate of 16000 Hz')
 	}
 
@@ -46,7 +47,7 @@ export async function recognize(
 		throw new Error(`The language ${formatLanguageCodeWithName(sourceLanguage)} is not supported by the Whisper engine.`)
 	}
 
-	if (isEnglishOnlyModel(modelName) && sourceLanguage != 'en') {
+	if (isEnglishOnlyModel(modelName) && sourceLanguage !== 'en') {
 		throw new Error(`The model '${modelName}' can only be used with English inputs. However, the given source language was ${languageCodeToName(sourceLanguage)}.`)
 	}
 
@@ -60,7 +61,7 @@ export async function recognize(
 
 	// Workaround issue with large-v3-turbo that produces invalid results when a prompt is passed to it.
 	// Always disable autoprompting for that model.
-	if (options.autoPromptParts && modelName == 'large-v3-turbo') {
+	if (options.autoPromptParts && modelName === 'large-v3-turbo') {
 		options.autoPromptParts = false
 	}
 
@@ -96,7 +97,7 @@ export async function align(
 
 	options = extendDeep(defaultWhisperAlignmentOptions, options)
 
-	if (sourceRawAudio.sampleRate != 16000) {
+	if (sourceRawAudio.sampleRate !== 16000) {
 		throw new Error('Source audio must have a sample rate of 16000 Hz')
 	}
 
@@ -106,7 +107,7 @@ export async function align(
 		throw new Error(`The language ${formatLanguageCodeWithName(sourceLanguage)} is not supported by the Whisper engine.`)
 	}
 
-	if (isEnglishOnlyModel(modelName) && sourceLanguage != 'en') {
+	if (isEnglishOnlyModel(modelName) && sourceLanguage !== 'en') {
 		throw new Error(`The model '${modelName}' can only be used with English inputs. However, the given source language was ${languageCodeToName(sourceLanguage)}.`)
 	}
 
@@ -139,7 +140,7 @@ export async function alignEnglishTranslation(
 
 	options = extendDeep(defaultWhisperAlignmentOptions, options)
 
-	if (sourceRawAudio.sampleRate != 16000) {
+	if (sourceRawAudio.sampleRate !== 16000) {
 		throw new Error('Source audio must have a sample rate of 16000 Hz')
 	}
 
@@ -184,7 +185,7 @@ export async function detectLanguage(
 
 	options = extendDeep(defaultWhisperLanguageDetectionOptions, options)
 
-	if (sourceRawAudio.sampleRate != 16000) {
+	if (sourceRawAudio.sampleRate !== 16000) {
 		throw new Error('Source audio must have a sample rate of 16000 Hz')
 	}
 
@@ -232,7 +233,7 @@ export async function detectVoiceActivity(
 
 	options = extendDeep(defaultWhisperVADOptions, options)
 
-	if (sourceRawAudio.sampleRate != 16000) {
+	if (sourceRawAudio.sampleRate !== 16000) {
 		throw new Error('Source audio must have a sample rate of 16000 Hz')
 	}
 
@@ -441,7 +442,11 @@ export class Whisper {
 			let {
 				decodedTokens: partTokens,
 				decodedTokensConfidence: partTokensConfidence,
-				decodedTokensCrossAttentionQKs: partCrossAttentionQKs,
+				decodedTokensCrossAttentionQKs: partTokensCrossAttentionQKs,
+
+				decodedTokensDecodingTime: partTokensDecodingTime,
+				decodedTokensInferenceTime: partTokensInferenceTime,
+				decodedTokensOverheadTime: partTokensOverheadTime,
 			} = await this.decodeTokens(
 				audioPartFeatures,
 				initialTokens,
@@ -470,19 +475,19 @@ export class Whisper {
 
 			await logger.startAsync(`Extract timeline for part (timestamp accuracy: ${options.timestampAccuracy!})`)
 
-			if (partTokens.length != partCrossAttentionQKs.length) {
-				throw new Error('Unexpected: partTokens.length != partCrossAttentionQKs.length')
+			if (partTokens.length !== partTokensCrossAttentionQKs.length) {
+				throw new Error('Unexpected: partTokens.length !== partCrossAttentionQKs.length')
 			}
 
 			// Prepare tokens
 			partTokens = partTokens.slice(initialTokens.length)
 			partTokensConfidence = partTokensConfidence.slice(initialTokens.length)
-			partCrossAttentionQKs = partCrossAttentionQKs.slice(initialTokens.length)
+			partTokensCrossAttentionQKs = partTokensCrossAttentionQKs.slice(initialTokens.length)
 
 			// Find alignment path
 			let alignmentHeads: number[] | undefined
 
-			if (options.timestampAccuracy === 'medium' || options.model == 'large-v3-turbo') {
+			if (options.timestampAccuracy === 'medium' || options.model === 'large-v3-turbo') {
 				alignmentHeads = this.alignmentHeadIndexes
 			} else if (options.timestampAccuracy === 'high') {
 				alignmentHeads = undefined
@@ -490,7 +495,7 @@ export class Whisper {
 				throw new Error(`Unsupported timestamp accuracy '${options.timestampAccuracy}', can only be 'medium' or 'high'.`)
 			}
 
-			const alignmentPath = await this.findAlignmentPathFromQKs(partCrossAttentionQKs, partTokens, 0, segmentFrameCount, alignmentHeads)
+			const alignmentPath = await this.findAlignmentPathFromQKs(partTokensCrossAttentionQKs, partTokens, 0, segmentFrameCount, alignmentHeads)
 
 			// Generate timeline from alignment path
 			const partTimeline = await this.getTokenTimelineFromAlignmentPath(alignmentPath, partTokens, segmentStartTime, segmentEndTime, partTokensConfidence)
@@ -514,6 +519,15 @@ export class Whisper {
 			audioOffset = audioEndOffset
 
 			logger.end()
+
+			if (logLevelGreaterOrEqualTo('trace')) {
+				const promptDecodingTime = partTokensDecodingTime[0]
+				const medianTokenDecodingTime = medianOfVector(partTokensDecodingTime.slice(1))
+				const medianTokenInferenceTime = medianOfVector(partTokensInferenceTime.slice(1))
+				const medianOverheadTime = medianOfVector(partTokensOverheadTime.slice(1))
+
+				logger.log(`${chalk.blueBright('Context')}: ${initialTokens.length + partTokens.length} tokens (${initialTokens.length} prompt, ${partTokens.length} decoded)\n${chalk.blueBright('Prompt decode time')}: ${promptDecodingTime.toFixed(1)}ms\n${chalk.blueBright('Median token decode time')}: ${medianTokenDecodingTime.toFixed(1)}ms (${medianTokenInferenceTime.toFixed(1)}ms inference, ${medianOverheadTime.toFixed(2)}ms overhead)`, 'trace')
+			}
 		}
 
 		// Convert token timeline to word timeline
@@ -812,8 +826,17 @@ export class Whisper {
 
 		const maxTokensPerPart = Math.min(options.maxTokensPerPart!, largestMaximumTokensPerPart)
 
+		let decodedTokensInferenceTime: number[] = []
+		let decodedTokensDecodingTime: number[] = []
+
+		const tokenDecodingTimeTimer = new Timer()
+
 		// Start decoding loop
 		for (let decodedTokenCount = 0; decodedTokenCount < maxTokensPerPart; decodedTokenCount++) {
+			if (decodedTokenCount > 0) {
+				decodedTokensDecodingTime.push(tokenDecodingTimeTimer.getElapsedTimeAndRestart())
+			}
+
 			const isInitialState = decodedTokens.length === initialTokens.length
 			const atLeastOneTextTokenDecoded = decodedTokens.slice(initialTokens.length).some(token => this.isTextToken(token))
 
@@ -847,9 +870,13 @@ export class Whisper {
 				offset: offsetTensor
 			}
 
-			// Run decoder model
+			//// Infer with ONNX decoder model
+			const tokenInferenceTimeTimer = new Timer()
+
 			const decoderOutputs = await this.textDecoder!.run(decoderInputs)
 
+			decodedTokensInferenceTime.push(tokenInferenceTimeTimer.elapsedTime)
+
 			// Extract decoder model results
 			const logitsBuffer = decoderOutputs['logits'].data as Float32Array
 			kvCacheTensor = decoderOutputs['output_kv_cache'] as any
@@ -1142,6 +1169,10 @@ export class Whisper {
 			await yieldToEventLoop()
 		}
 
+		if (decodedTokensDecodingTime.length === decodedTokensInferenceTime.length - 1) {
+			decodedTokensDecodingTime.push(tokenDecodingTimeTimer.getElapsedTimeAndRestart())
+		}
+
 		// If at least two timestamp tokens were decoded and it's not the final part,
 		// truncate up to the last timestamp token
 		if (timestampTokenSeenCount >= 2 && !isFinalPart) {
@@ -1151,8 +1182,13 @@ export class Whisper {
 			decodedTokensTimestampLogits = decodedTokensTimestampLogits.slice(0, sliceEndTokenIndex)
 			decodedTokensCrossAttentionQKs = decodedTokensCrossAttentionQKs.slice(0, sliceEndTokenIndex)
 			decodedTokensConfidence = decodedTokensConfidence.slice(0, sliceEndTokenIndex)
+
+			decodedTokensDecodingTime = decodedTokensDecodingTime.slice(0, sliceEndTokenIndex)
+			decodedTokensInferenceTime = decodedTokensInferenceTime.slice(0, sliceEndTokenIndex)
 		}
 
+		const decodedTokensOverheadTime = decodedTokensDecodingTime.map((time, index) => time - decodedTokensInferenceTime[index])
+
 		logger.write('\n')
 		logger.end()
 
@@ -1162,6 +1198,10 @@ export class Whisper {
 			decodedTokensTimestampLogits,
 			decodedTokensConfidence,
 			decodedTokensCrossAttentionQKs,
+
+			decodedTokensDecodingTime,
+			decodedTokensInferenceTime,
+			decodedTokensOverheadTime,
 		}
 	}
 
@@ -1186,7 +1226,7 @@ export class Whisper {
 		const maxAudioSamples = sampleRate * 30
 		const maxAudioFrames = 3000
 
-		if (sampleRate != 16000) {
+		if (sampleRate !== 16000) {
 			throw new Error('Audio must have a sample rate of 16000 Hz')
 		}
 
@@ -1267,7 +1307,7 @@ export class Whisper {
 			return isSeparatorCharacter(text[text.length - 1])
 		}
 
-		if (language != 'zh' && language != 'ja') {
+		if (language !== 'zh' && language !== 'ja') {
 			tokenTimeline = tokenTimeline.filter(entry => this.isTextToken(entry.id!))
 		}
 
@@ -1282,7 +1322,7 @@ export class Whisper {
 			const text = entry.text
 			const previousEntryText = previousEntry?.text
 
-			if (groups.length == 0 ||
+			if (groups.length === 0 ||
 				text === '' ||
 				startsWithSeparatorCharacter(text) ||
 				(previousEntryText != null && endsWithSeparatorCharacter(previousEntryText))) {
@@ -1346,14 +1386,14 @@ export class Whisper {
 	}
 
 	async getTokenTimelineFromAlignmentPath(alignmentPath: AlignmentPath, tokens: number[], startTimeOffset: number, endTimeOffset: number, tokensConfidence?: number[], correctionAmount = 0.0) {
-		if (alignmentPath.length == 0) {
+		if (alignmentPath.length === 0) {
 			return []
 		}
 
 		const tokenTimeline: Timeline = []
 
 		for (let pathIndex = 0; pathIndex < alignmentPath.length; pathIndex++) {
-			if (pathIndex != 0 && alignmentPath[pathIndex].source == alignmentPath[pathIndex - 1].source) {
+			if (pathIndex !== 0 && alignmentPath[pathIndex].source === alignmentPath[pathIndex - 1].source) {
 				continue
 			}
 
@@ -1604,7 +1644,7 @@ export class Whisper {
 
 		const onnxSessionOptions = getOnnxSessionOptions({ executionProviders: this.encoderExecutionProviders })
 
-		const onnxProvidersString = onnxSessionOptions.executionProviders!.length > 0 ? `${ onnxSessionOptions.executionProviders!.join(', ') }` : `default`
+		const onnxProvidersString = onnxSessionOptions.executionProviders!.length > 0 ? `${onnxSessionOptions.executionProviders!.join(', ')}` : `default`
 
 		await logger.startAsync(`Create encoder inference session for model '${this.modelName}' (ONNX provider: ${onnxProvidersString})`)
 
@@ -1642,15 +1682,15 @@ export class Whisper {
 	getKvDimensions(groupCount: number, length: number) {
 		const modelName = this.modelName
 
-		if (modelName == 'tiny' || modelName == 'tiny.en') {
+		if (modelName === 'tiny' || modelName === 'tiny.en') {
 			return [8, groupCount, length, 384]
-		} else if (modelName == 'base' || modelName == 'base.en') {
+		} else if (modelName === 'base' || modelName === 'base.en') {
 			return [12, groupCount, length, 512]
-		} else if (modelName == 'small' || modelName == 'small.en') {
+		} else if (modelName === 'small' || modelName === 'small.en') {
 			return [24, groupCount, length, 768]
-		} else if (modelName == 'medium' || modelName == 'medium.en') {
+		} else if (modelName === 'medium' || modelName === 'medium.en') {
 			return [48, groupCount, length, 1024]
-		} else if (modelName == 'large-v1' || modelName == 'large-v2' || modelName == 'large-v3' || modelName == 'large-v3-turbo') {
+		} else if (modelName === 'large-v1' || modelName === 'large-v2' || modelName === 'large-v3' || modelName === 'large-v3-turbo') {
 			return [64, groupCount, length, 1280]
 		} else {
 			throw new Error(`Unsupported model: ${modelName}`)
@@ -1664,7 +1704,7 @@ export class Whisper {
 
 		if (this.isMultiligualModel) {
 			const languageToken = this.tokenConfig.languageTokensStart + languageIdLookup[language]
-			const taskToken = task == 'translate' ? this.tokenConfig.translateTaskToken : this.tokenConfig.transcribeTaskToken
+			const taskToken = task === 'translate' ? this.tokenConfig.translateTaskToken : this.tokenConfig.transcribeTaskToken
 
 			tokens = [startOfTextToken, languageToken, taskToken]
 		} else {
@@ -1900,7 +1940,7 @@ export async function loadPackagesAndGetPaths(modelName: WhisperModelName | unde
 		if (languageCode) {
 			const shortLanguageCode = getShortLanguageCode(languageCode)
 
-			modelName = shortLanguageCode == 'en' ? 'tiny.en' : 'tiny'
+			modelName = shortLanguageCode === 'en' ? 'tiny.en' : 'tiny'
 		} else {
 			modelName = 'tiny'
 		}
@@ -1918,7 +1958,7 @@ export async function loadPackagesAndGetPaths(modelName: WhisperModelName | unde
 }
 
 export function normalizeWhisperModelName(modelName: WhisperModelName, languageCode: string | undefined): WhisperModelName {
-	if (languageCode != 'en' && modelName.endsWith('.en')) {
+	if (languageCode !== 'en' && modelName.endsWith('.en')) {
 		const originalModelName = modelName
 		modelName = modelName.slice(0, modelName.length - 3) as WhisperModelName
 

package/src/utilities/Timer.ts

@@ -1,8 +1,5 @@
 import { logToStderr, roundToDigits } from './Utilities.js'
 
-declare const chrome: any
-declare const process: any
-
 export class Timer {
 	startTime = 0
 
@@ -46,50 +43,37 @@ export class Timer {
 	}
 
 	static get currentTime(): number {
-		if (!this.timestampFunc) {
-			this.createGlobalTimestampFunction()
+		if (!this.getTimestamp) {
+			this.createTimestampFunction()
 		}
 
-		return this.timestampFunc()
+		return this.getTimestamp()
 	}
 
 	static get microsecondTimestamp(): number {
 		return Math.floor(Timer.currentTime * 1000)
 	}
 
-	private static createGlobalTimestampFunction() {
+	private static createTimestampFunction() {
 		if (typeof process === 'object' && typeof process.hrtime === 'function') {
 			let baseTimestamp = 0
 
-			this.timestampFunc = () => {
-				const nodeTimeStamp = process.hrtime()
-				const millisecondTime = (nodeTimeStamp[0] * 1000) + (nodeTimeStamp[1] / 1000000)
+			this.getTimestamp = () => {
+				const nodeTimeNanoSeconds = process.hrtime.bigint()
+				const nodeTimeMilliseconds = Number(nodeTimeNanoSeconds) / 1_000_000
 
-				return baseTimestamp + millisecondTime
+				return baseTimestamp + nodeTimeMilliseconds
 			}
 
-			baseTimestamp = Date.now() - this.timestampFunc()
-		}
-		else if (typeof chrome === 'object' && chrome.Interval) {
-			const baseTimestamp = Date.now()
-
-			const chromeIntervalObject = new chrome.Interval()
-			chromeIntervalObject.start()
-
-			this.timestampFunc = () => baseTimestamp + chromeIntervalObject.microseconds() / 1000
-		}
-		else if (typeof performance === 'object' && performance.now) {
+			baseTimestamp = Date.now() - this.getTimestamp()
+		} else if (typeof performance === 'object' && performance.now) {
 			const baseTimestamp = Date.now() - performance.now()
 
-			this.timestampFunc = () => baseTimestamp + performance.now()
-		}
-		else if (Date.now) {
-			this.timestampFunc = () => Date.now()
-		}
-		else {
-			this.timestampFunc = () => (new Date()).getTime()
+			this.getTimestamp = () => baseTimestamp + performance.now()
+		} else {
+			this.getTimestamp = () => Date.now()
 		}
 	}
 
-	private static timestampFunc: () => number
+	private static getTimestamp: () => number
 }

package/src/utilities/Utilities.ts

@@ -8,6 +8,7 @@ import { ChildProcessWithoutNullStreams } from 'node:child_process'
 import { inspect } from 'node:util'
 import { TypedArray, TypedArrayConstructor } from '../typings/TypedArray.js'
 import { encodeHex } from '../encodings/Hex.js'
+import { Timer } from './Timer.js'
 
 const log = logToStderr
 
@@ -123,9 +124,21 @@ export function roundToDigits(val: number, digits = 3) {
 	return Math.round(val * multiplier) / multiplier
 }
 
-export function delay(timeMs: number) {
-	return new Promise((resolve) => {
-		setTimeout(resolve, timeMs)
+export function sleep(timeMs: number) {
+	const timer = new Timer()
+
+	return new Promise<void>((resolve) => {
+		const tickCallback = () => {
+			if (timer.elapsedTime < timeMs) {
+				//setImmediate(tickCallback)
+				setTimeout(tickCallback, 0)
+			} else {
+				resolve()
+			}
+		}
+
+		//setImmediate(tickCallback)
+		setTimeout(tickCallback, 0)
 	})
 }
 
@@ -440,7 +453,7 @@ export async function runOperationWithRetries<R>(
 			logger.log('', 'error')
 			logger.logTitledMessage(`${operationName} failed`, `Trying again in ${delayBetweenRetries}ms..`, chalk.redBright, 'error')
 
-			await delay(delayBetweenRetries)
+			await sleep(delayBetweenRetries)
 
 			logger.log(``, 'warning')
 			logger.logTitledMessage(`Starting retry attempt`, `${retryIndex} / ${maxRetries}`, chalk.yellowBright, 'warning')