import { exec } from 'child_process'
import { readFile, unlink } from 'fs/promises'
import { promisify } from 'util'
import type { WhisperResult, LanguageCode } from './types'
import { cleanTranscript } from './transcript-cleaner'
const execAsync = promisify(exec)
type WhisperModel = 'base' | 'large' | 'medium' | 'small' | 'turbo'
const COMPLEX_LANGUAGES: LanguageCode[] = ['zh', 'ja', 'ko', 'ar', 'hi', 'th']
const selectWhisperModel = (detectedLanguage?: string): WhisperModel =>
(detectedLanguage && COMPLEX_LANGUAGES.includes(detectedLanguage as LanguageCode) ? 'base' : 'turbo')
const getAudioQuality = (detectedLanguage?: string): string =>
(detectedLanguage && ['ja', 'ko', 'zh'].includes(detectedLanguage) ? '192k' : '128k')
const buildAudioCommand = (url: string, tempAudioFile: string, quality: string): string => {
const args = [
'yt-dlp',
'-x',
'--audio-format wav',
`--audio-quality ${quality}`,
'--no-warnings',
'-o',
`"${tempAudioFile}.%(ext)s"`,
`"${url}"`,
]
return args.join(' ')
}
const buildWhisperCommand = (tempAudioFile: string, model: WhisperModel, detectedLanguage?: string): string => {
const args = [
'whisper',
`"${tempAudioFile}.wav"`,
`--model ${model}`,
'--output_format txt',
detectedLanguage ? `--language ${detectedLanguage}` : '--language auto',
'--fp16 True',
'--verbose False',
'--word_timestamps False',
'--no_speech_threshold 0.6',
'--logprob_threshold -1.0',
]
return args.join(' ')
}
const extractLanguageFromStderr = (stderr: string): string | undefined => {
const languageMatch = stderr.match(/Detected language: (\w+)/)
return languageMatch ? languageMatch[1] : undefined
}
const cleanupTempFiles = async (tempAudioFile: string): Promise<void> => {
try {
await execAsync(`rm -f "${tempAudioFile}.wav" "${tempAudioFile}.txt"`)
} catch {
// Ignore cleanup errors
}
}
export const extractWithWhisper = async (url: string, detectedLanguage?: string): Promise<WhisperResult> => {
const model = selectWhisperModel(detectedLanguage)
const tempAudioFile = `temp_audio_${Date.now()}`
try {
const audioQuality = getAudioQuality(detectedLanguage)
const audioCommand = buildAudioCommand(url, tempAudioFile, audioQuality)
await execAsync(audioCommand)
const whisperCommand = buildWhisperCommand(tempAudioFile, model, detectedLanguage)
console.log(`Using Whisper model: ${model} for language: ${detectedLanguage || 'auto'}`)
const { stderr } = await execAsync(whisperCommand)
const finalDetectedLanguage = detectedLanguage || extractLanguageFromStderr(stderr)
const textFile = `${tempAudioFile}.txt`
const rawTranscript = await readFile(textFile, 'utf-8')
const cleanedTranscript = cleanTranscript(rawTranscript.trim())
await unlink(`${tempAudioFile}.wav`)
await unlink(textFile)
console.log(`🎤 Whisper extraction successful (${model})`)
console.log(`📊 Original length: ${rawTranscript.length} chars`)
console.log(`📊 Cleaned length: ${cleanedTranscript.length} chars`)
return {
transcript: cleanedTranscript,
source: 'whisper',
detectedLanguage: finalDetectedLanguage,
model,
}
} catch (error) {
await cleanupTempFiles(tempAudioFile)
const errorMessage = error instanceof Error ? error.message : 'Unknown error'
throw new Error(`Failed to extract with Whisper: ${errorMessage}`)
}
}
