Speech Recognition — bu nima, texnologiyalari va ishlash tamoyillari

Speech Recognition nima: ASR tamoyillari

Automatic Speech Recognition (ASR) — pipeline: audio → dastlabki ishlov (16 kHz mono, noise reduction, VAD — ovoz faolligi) → feature extraction (MFCC, LogMel FilterBank) → acoustic model (neyron tarmoq: CTC / RNNT / Transducer) → language model (LM) → dekodlash (matn). Zamonaviy end-to-end modellar (Whisper, Google USM, Conformer CTC) acoustic + language model'ni bitta Transformer'ga birlashtiradi, bu pipeline'ni soddalashtiradi va aniqlikni oshiradi.

Mobil qurilmalar uchun ASR arxitekturalari: CTC (Connectionist Temporal Classification) — tezkor, audio va matnni moslashtirishni talab qilmaydi, Vosk, Android native'da ishlatiladi. RNNT (Recurrent Neural Network Transducer) — streaming ASR, past latency (Google USM). Conformer — CNN + Transformer gibridi, eng yaxshi aniqlik, lekin og'irroq: Whisper Medium (769M params) — 30–60x latency. On-device uchun CTC afzal: Vosk CTC modellari 50–100 MB joy egallaydi, latency 0.3–0.8x real-time.

ASR metrikalari: WER (Word Error Rate) — etalon bilan solishtirganda almashtirilgan, o'chirilgan, qo'shilgan so'zlar foizi. Google Cloud ASR — 7% WER (EN), 12% (RU). Vosk — 13% (RU), 9% (EN). Whisper Small — 6% (EN), 10% (RU). CER (Character Error Rate) — xuddi shunday, belgilar darajasida. Real-Time Factor (RTF) — ishlov berish vaqti / audio davomiyligi. RTF < 1 — real vaqtdan tezroq. RTF < 0.3 — real-time ASR. Ovozli kiritish uchun RTF < 1 kerak, transkripsiya uchun — RTF < 3.

VAD (Voice Activity Detection) — nutqni sukunat va shovqindan ajratuvchi ovoz faolligi detektori. WebRTC VAD — mobil ASR uchun standart: 30 ms freymlar, uchta rejim (0, 1, 2 — konservativdan agressivgacha). VAD RTF'ni 1.5–3x kamaytiradi (nutqsiz qismlarni o'tkazib yuboradi). Silero VAD (MIT) — neyron tarmoq asosidagi VAD, WebRTC'dan aniqroq (94% vs 85% ROC AUC), 5–10 ms latency, TFLite va Core ML. Production ASR uchun VAD → ASR kaskadini qo'llang: bu noto'g'ri ishga tushirishlarni kamaytiradi va ishlovni tezlashtiradi.

Android SpeechRecognizer API

Android SpeechRecognizer — Android SDK'ga o'rnatilgan nutqni tanib olish klassi. Ikki rejimda ishlaydi: cloud (Google server) — yuqori aniqlik, internet talab qiladi; on-device (Android 10+ dan) — GBoard embedded ASR modeli, 20+ til, WER 15–18%. SpeechRecognizer nutq paytida oraliq natijalarni (partial results) va yakuniy matnni qaytaradi. RecognizerIntent.EXTRA_LANGUAGE orqali 60+ tilni qo'llab-quvvatlaydi. Ovozli kiritish uchun tavsiya etiladi — tezkor integratsiya, bepul.

SpeechRecognizer API: SpeechRecognizer.createSpeechRecognizer(context) orqali yaratish. RecognizerIntent.ACTION_RECOGNIZE_SPEECH bilan Intent, extra: LANGUAGE_MODEL_FREE_FORM (erkin matn) yoki LANGUAGE_MODEL_WEB_SEARCH (qidiruv so'rovlari). Natija RecognitionListener orqali: onReadyForSpeech → onBeginningOfSpeech → onRmsChanged → onPartialResults → onResults. Continuous recognition ("har doim tinglash" rejimi) uchun — onResults'dan keyin recognizer'ni qayta ishga tushiring. Batareyani tejash uchun onDeviceOnly = true dan foydalaning.

// Android SpeechRecognizer
val recognizer = SpeechRecognizer.createSpeechRecognizer(context)
recognizer.setRecognitionListener(object : RecognitionListener {
    override fun onResults(results: Bundle) {
        val text = results
            .getStringArrayList(SpeechRecognizer.RESULTS_RECOGNITION)
            ?.firstOrNull() ?: ""
        showResult(text)
    }
    override fun onPartialResults(partialResults: Bundle) {
        val partial = partialResults
            .getStringArrayList(SpeechRecognizer.RESULTS_RECOGNITION)
            ?.firstOrNull() ?: ""
        showPartial(partial)
    }
})

val intent = Intent(RecognizerIntent.ACTION_RECOGNIZE_SPEECH).apply {
    putExtra(RecognizerIntent.EXTRA_LANGUAGE_MODEL,
        RecognizerIntent.LANGUAGE_MODEL_FREE_FORM)
    putExtra(RecognizerIntent.EXTRA_LANGUAGE, "ru-RU")
}
recognizer.startListening(intent)

On-device vs Cloud Android'da: on-device internetsiz ishlaydi (Android 10+, Pixel 4+, Samsung S20+). Cloud — aniqroq (Google Neural Network ASM Model), lekin latency yuqoriroq (tarmoq bilan 300–800 ms). Ovozli kiritish uchun gibriddan foydalaning: tarmoq yo'q bo'lganda on-device'ga fallback qiladigan cloud. Android SpeechRecognizer rejimni RecognizerIntent.EXTRA_PREFER_OFFLINE ga qarab avtomatik tanlaydi. Maksimal maxfiylik uchun — onDeviceOnly = true qo'ying (lekin aniqlik rus tilida 15–18% WER).

iOS SFSpeechRecognizer va Speech Framework

SFSpeechRecognizer — Apple'ning iOS, macOS, watchOS uchun nutqni tanib olish frameworki. iOS 10 dan mavjud. On-device rejim — iOS 17 dan (mahalliy model, internetsiz). 60+ tilni qo'llab-quvvatlaydi. Aniqlik: WER 7–12% (on-device — 12–15%). SFSpeechRecognizer Speech.framework orqali mavjud — qo'shimcha SDK talab qilmaydi. Ruxsat so'rash SFSpeechRecognizer.requestAuthorization orqali.

SFSpeechRecognizer API: SFSpeechRecognizer(locale: Locale(identifier: "ru_RU")) → SFSpeechAudioBufferRecognitionRequest (live audio) yoki SFSpeechURLRecognitionRequest (audio fayl). Streaming recognitionTask(with:delegate:) orqali SFSpeechRecognitionTaskDelegate bilan (didHypothesizeTranscription — partial, didFinishRecognition — final). On-device rejim: request.requiresOnDeviceRecognition = true. iOS 15+ uchun: request.shouldReportPartialResults = true (past kechikishli streaming natija).

let recognizer = SFSpeechRecognizer(locale: Locale(identifier: "ru_RU"))
guard recognizer.isAvailable else { return }

let request = SFSpeechAudioBufferRecognitionRequest()
request.requiresOnDeviceRecognition = true
request.shouldReportPartialResults = true

recognitionTask = recognizer.recognitionTask(with: request) { result, error in
    if let result = result {
        let text = result.bestTranscription.formattedString
        let isFinal = result.isFinal
        DispatchQueue.main.async {
            textView.text = text
        }
    }
}

audioEngine.inputNode.installTap(config: bufferConfig) { buffer, _ in
    request.append(buffer)
}
audioEngine.prepare()
try audioEngine.start()

SFSpeechRecognizer vs Android SpeechRecognizer: SFSpeechRecognizer qayta ishga tushirmasdan mahalliy streaming'ga ega (Android qayta startListening talab qiladi). iOS'da on-device iOS 17 dan mavjud (Android — Android 10 dan). Ikkalasi ham foydalanuvchi ruxsatini talab qiladi (iOS uchun NSMicrophoneUsageDescription + NSSpeechRecognitionUsageDescription, Android uchun RECORD_AUDIO). SFSpeechRecognizer ingliz tilida aniqroq (on-device), Android SpeechRecognizer rus tilida aniqroq (cloud). iOS 17 dan oldingi versiyalarda on-device mavjud emas — internet talab qilinadi. iOS 17+ uchun on-device rus tilida WER 12–14% beradi.

Vosk: 20+ tilda oflayn tanib olish

Vosk — Alpha Cephei kompaniyasining offline nutqni tanib olish uchun open-source ASR kutubxonasi. Kaldi ASR toolkit + CTC modellari asosida qurilgan. 20+ tilni qo'llab-quvvatlaydi: rus, ingliz, nemis, fransuz, ispan, xitoy, arab. Modellar 50 MB (small — 50 MB, ru) dan 1.2 GB (large — 1.2 GB, en) gacha. Vosk Android (JNI), iOS (C++ wrapper), Linux, Windows, Raspberry Pi'da ishlaydi. RTF: 0.3–0.8 flagman qurilmalarda. Vosk — to'liq offline ASR uchun eng yaxshi tanlov.

Vosk API: VoskWaveformModelLoader (model yuklash) → VoskWaveformRecognizer (tanib oluvchini yaratish) → recognizer.createGrammar(list) yoki recognizer.getResult() / recognizer.getPartialResult(). Grammatikalarni qo'llab-quvvatlash (cheklangan buyruqlar to'plami) — GrammarRecogniser — RTF 0.1–0.3 beradi (3x tezroq). Ovozli kiritish uchun erkin tanib olishdan (getResult) foydalaning. Ovozli buyruqlar uchun — GrammarRecogniser (buyruqlarda 99% aniqlik). Vosk shuningdek Speaker Identification'ni (ovozni vektorli tahlil) qo'llab-quvvatlaydi.

// Vosk oflayn ASR
val model = VoskModel("model/vosk-model-small-ru-0.22")
model.load()

val recognizer = VoskRecognizer(model, 16000.0f)
recognizer.setWords(true)

audioRecord.startRecording()
while (isListening) {
    val buffer = ByteArray(3200) // 100ms audio
    audioRecord.read(buffer, 0, buffer.size)
    if (recognizer.acceptWaveform(buffer)) {
        val result = recognizer.result // JSON
        text += JSONObject(result).getString("text")
    }
}

Vosk vs Cloud ASR: Vosk butunlay oflayn — maxfiylik, zero latency, bepul. Cloud (Google, Yandex) murakkab sahnalarda (shovqin, aksent) aniqroq — WER 3–5% past. Vosk quyidagilar uchun idealdir: internetsiz ovozli kiritish, accessibility ilovalari, qo'ng'iroqlarni transkripsiya qilish (maxfiylik). Cloud ASR — ovozli yordamchilar uchun, bu yerda aniqlik maxfiylikdan muhimroq. Tavsiya: offline uchun Vosk, online uchun SFSpeechRecognizer (iOS) / SpeechRecognizer (Android) — turli stsenariylar uchun yondashuvlarni kombinatsiya qiling.

Whisper OpenAI mobil qurilmalarda

Whisper — OpenAI'ning nutqni tanib olish modeli, 680,000 soat audio bilan o'qitilgan (ko'p tilli, 99 til). O'lchamlari: tiny (39M, WER 10% EN, 15% RU), small (244M, WER 6% EN, 10% RU), medium (769M, WER 4.5% EN, 8% RU). Whisper mobil qurilmalarda Core ML (iOS, ANE) va TFLite (Android, GPU) orqali ishlaydi. Whisper tiny: CPU'da RTF 4–10, GPU'da RTF 0.5–2 (Android). Whisper small: GPU'da RTF 2–6. Whisper medium — RTF 8–15, faqat non-real-time uchun.

Whisper iOS'da (Core ML): Apple MLX Whisper — Whisper'ning Core ML va MLX (Apple Neural Engine) uchun implementatsiyasi. Whisper tiny Core ML iPhone 15 Pro'da: RTF 0.3–0.8 (real-time'dan tezroq!). Whisper small Core ML: RTF 1–3. Whisper Core ML A17 Pro'da ANE'dan foydalanadi (Neural Engine 35 TOPS). Hugging Face Transformers → Export to Core ML coremltools-whisper orqali. Streaming uchun WhisperStitcher (chunking + stitching) dan foydalaning. Whisper iOS'da — eng aniq on-device ASR (WER 6% EN, 10% RU).

// Whisper Core ML iOS'da
import MLXWhisper

let model = try WhisperModel(from: "whisper-tiny")
let audio = try AudioUtils.loadAudio(url: audioURL)
let segments = try model.transcribe(audio: Audio, options: TranscribeOptions(
    language: "ru",
    wordTimestamps: true,
    temperature: 0.0
))

for segment in segments {
    print(segment.text) // vaqt belgilari bilan yakuniy matn
}

Whisper vs Vosk: Whisper aniqroq (6% vs 9% WER EN), 99 tilni qo'llab-quvvatlaydi, language detection, punctuation, emotion recognition'ni o'z ichiga oladi. Whisper og'irroq (39M–769M vs 50M Vosk), real-time'da sekinroq (RTF 0.5–6 vs 0.3–0.8). Whisper tiny — tezligi bo'yicha Vosk bilan taqqoslanadi (RTF 0.5–2 GPU). Vosk yengilroq, tezroq, streaming real-time uchun yaxshiroq. Whisper — bir martalik transkripsiya uchun, bu yerda aniqlik tezlikdan ustun. Vosk — real-time ovozli kiritish uchun. Yakuniy matn uchun Whisper, interaktiv uchun — Vosk foydalaning.

Speech Recognition'ni mobil ilovalarda qo'llash

Ovozli matn kiritish — ASR'ning eng ommaviy qo'llanishi: xabarlar, qidiruv so'rovlari, eslatmalarni dikta qilish. Talab: RTF < 1 (real vaqt), partial results (matnni aytilayotganda ko'rish). Android Gboard va iOS Keyboard o'rnatilgan ASR'ga ega (on-device, WER 12–18%). Maxsus implementatsiya uchun Android SpeechRecognizer / SFSpeechRecognizer'dan foydalaning. Maksimal aniqlik uchun — Cloud ASR + Vosk fallback. Rus tilida 98% aniqlik bilan ovozli kiritish uchun — Vosk (offline) + Yandex SpeechKit (online) kombinatsiyasi.

Qo'ng'iroqlar va uchrashuvlar transkripsiyasi — stenogrammalarni avtomatik yaratish uchun ASR. Talablar: no latency requirement, WER < 10%, speaker diarization (kim gapirayotgani). Whisper Small (offline) + pyannote-audio (diarization) — transkripsiya uchun standart stack. iOS'da — Whisper Core ML (RTF 1–3, WER 6–10%). Android'da — Whisper TFLite (RTF 2–6, WER 8–12%) yoki Google Cloud ASR + diarization. Maxfiylik uchun (qo'ng'iroqlar serverga bormasligi kerak) — Whisper qurilmada. Diarization aniqligi: 80–90% DER.

Ovozli yordamchilar — Siri (SFSpeechRecognizer), Google Assistant (Android SpeechRecognizer), Alexa (on-device ASR). Maxsus yordamchi: ASR → NLU (Natural Language Understanding) → Action → TTS. ASR — birinchi bosqich — butun zanjirning aniqligini belgilaydi. NLU uchun RASA, Snips NLU (on-device) yoki Cloud NLU (Dialogflow, Amazon Lex) dan foydalaning. TTS uchun: Android TTS / iOS AVSpeechSynthesizer. On-device ovozli yordamchi (Vosk + RASA + TTS) — to'liq maxfiy, internetsiz ishlaydi, so'rovga latency < 2 soniya.

// Android SpeechRecognizer + TTS bilan ovozli yordamchi
recognizer.setRecognitionListener(object : RecognitionListener {
    override fun onResults(results: Bundle) {
        val query = results.getStringArrayList(
            SpeechRecognizer.RESULTS_RECOGNITION
        )?.firstOrNull() ?: return

        val intentResult = nluService.classify(query)
        textToSpeech.speak(intentResult.response)
        executeAction(intentResult.action)
    }
})

Motorikasi cheklangan odamlar uchun Accessibility — Speech Recognition ilovani ovoz bilan boshqarish imkonini beradi: kontaktlarni ochish, xabar yuborish, raqam terish. Android Voice Access va iOS Switch Control + VoiceOver — o'rnatilgan yechimlar. Maxsus implementatsiya uchun: fixed buyruqlar to'plami bilan GrammarRecogniser (Vosk) (50–100 ibora) — RTF 0.1–0.3, 99% aniqlik. Vosk Grammar grammatikani BNF formatida aniqlash imkonini beradi. Accessibility uchun tezlik muhim — Vosk Grammar erkin tanib olishdan 5x tezroq va kamroq batareya sarflaydi.