Pustakam Library

Free Programming learning guide

Build And Deploy A Real-Time AI Audio Deepfake Detector Using Python And PyTorch

Build And Deploy A Real-Time AI Audio Deepfake Detector Using Python And PyTorch — a free advanced-level guide covering build and deploy a real-time ai...

79 min read8 chaptersadvanced

What you will learn

  1. Pehla Kaam: Machao Setup, Nahi Toh Aukaat Pe Aayega
  2. Data Daalo Toh Life Jhakaas, Nahi Toh Model Ghaati
  3. Spectrograms Aur MFCCs Ki Real Game, Chamka Kya?
  4. CNN Aur Transformer Ka Jugaad, Boss Level Architecture
  5. Train Karna Hai Aashiq Ki Tarah, Patience Rakh Guru
  6. Real-Time Ka Scene Hai, Latency Kam Kar Bhidu
  7. Deploy Karke Dikhao, FastAPI Aur Docker Ki Chakri
  8. Frontend Laga Ke Public Ko Batao, Asli Bheja Khol

1. Pehla Kaam: Machao Setup, Nahi Toh Aukaat Pe Aayega

Sun bhai, ek kaan pakad ke bol. Aaj raat 3 baje tera startup crash hua, investor ne phone uthaya nahi, aur tu screen ke aage baith ke soch raha hai—"Zindagi me kya karu?" Sahi hai. Pura Mumbai local train ki first class ka scene hai: sab bhag rahe hain, kisi ko pata nahi kahan jaana hai, par doorbell bajte hi andar ghusna hai. Tera idea hai—real-time AI audio deepfake detector. Bhai, dekha hai kya aaj kal ke scam? Aawaz clone karke kisi maa ko bol rahe hain "beta accident ho gaya, 2 lakh bhej de", ya phir CEO ki aawaz nakal karke finance department se crore transfer karwa rahe hain. Ye AI deepfake ki maa chd rakhi hai market mein. Tujhe isko detect karna hai. Asli hero ban-na hai? Toh pehla kaam setup machao, warna code likhne baithega aur environment hi ghaati pe aayega. Saale, ye chapter itna important hai ki agar tune ye skip kiya, toh tera pura project ekdum start hote hi muh ki khaayega. Jaise sangeet bajane se pehle tanpura tune nahi kiya, toh gaana kya khopcha niklega? Same logic. Asli Funda (Core Concepts) Dekh bhai, deepfake detection ka model train karna aur usko real-time chalana—ye koi "Hello World" Python tutorial nahi hai. Ye hardcore gym hai. Tereko weights uthaane padenge (tensors), stamina chahiye (GPU), aur ekdum sahi diet chahiye (libraries). Agar foundation hi bekaar hoga, toh baad me jab tu 10 GB ka audio dataset load karega, tera system RAM khaak ho jaayega aur laptop aise garam hoga jaise August ki dopahar Mumbai ki gali. Isliye, hum asli setup karenge. No jugaad, no "chalta hai" attitude. Professional gully don level pe kaam karenge. 1. Python Ka Base: Naya Version, Naya Rasta Tere paas jo purana wala Python 3.8 ka system hai, usko abhi dustbin mein daal de. Hum Python 3.11 ya 3.12 use karenge. Kyun? Kyunki bhai, speed! Naye Python versions me compiler optimization itni acchi hai ki code execution jaise local train fast track pe chala jaaye. Par ruk, Python seedha system pe install karega? Bc, tera dimaag kharab hai kya? Seedha system pe install kiya toh baad me dependency conflict hoga, ek library update ki toh doosri ki maa chd jayegi. Hum use karenge Conda (Miniconda specifically). Ye aisa tupperware dabba hai jisme tu apna saaman alag rakh sakta hai. Ghar ka saaman (system Python) alag, tera project ka saaman (virtual environment) alag. 💡 Pro Tip: Hamesha Miniconda use kar. Anaconda heavy hai, usme bohot saara bekaar saaman pehle se loaded aata hai jo tujhe zindagi me kaam nahi aayega. Miniconda lightweight hai, bas engine deta hai, baaki tu apni marzi se parts lagata hai. 2. PyTorch Aur Torchaudio: Asli Weapons Python toh bas …

2. Data Daalo Toh Life Jhakaas, Nahi Toh Model Ghaati

Abe sun, raat ko 3 baje jab tu code likh raha tha aur tere model ne "99% accuracy!" dikhaayi, fir tu jump karke apne aashiq wale status update kar diya. Subah ko jab real data pe test kiya, toh accuracy 12% pe gayi. Saala model tune train nahi kiya tha, tune usko ratta maarwa diya tha! Yehi hota hai jab tu data ko ignore karta hai. Sharma ji ka beta IIT crack karta hai kyunki uske paas 10 saal ke previous year papers hote hain. Tere paas ekdum kachra data hai, toh tera model bhi kachra hi banega. Bhai, machine learning mein ek rule hai—Garbage In, Garbage Out (GIGO). Tu apne model ko bina machli ke sikhayega ki machli kaise pakadni hai? Wo kya karega, bhookhe pet se code likhega? ASVspoof dataset teri machli hai, tera aadhaar hai. Isko nahi samjhega, toh tera pura project ghaati mein jaayega. Pehle chapter mein tune Conda environment banaya, PyTorch aur Torchaudio install kiya. Ab waqt hai asli khel shuru karne ka. Aaj hum data ka market karenge—sahi data uthayenge, usko saaf karenge, aur thoda tadka lagaake usko model-ready banayenge. Chalo, collar pakad, chalu karte hain. Asli Funda (Core Concepts) ASVspoof: Tere Project Ki Asli Aukaat Dekh bhai, deepfake detection ka jab bhi baat hoti hai, ASVspoof (Automatic Speaker Verification and Anti-Spoofing) ekdum Bhagavad Gita jaisa hai. Yeh ek international competition hai jo har 2-3 saal mein hota hai, jisme duniya bhar ke log fake audio banate hain aur usko pakadne wale algorithm banate hain. Tujhe mainly ASVspoof 2019 aur 2021 ka data chahiye. Isme do type ke hote hain: 1. Bonafide (Asli): Real insaan ki awaaz. 2. Spoofed (Nakli): AI se generate ki gayi awaaz (Text-to-Speech, Voice Conversion, wagrah). 💡 Pro Tip: ASVspoof ka data free mein available hai, but thoda bhaari hai (kai GBs ka). Isliye pehle hi apne hard drive mein 50-100 GB jagah khaali rakh. Cloud pe rakhna hai toh AWS S3 ya Google Drive use kar, local system mein ghoosa ke mat phasa. Data Imbalance: Teri Train Ki Engine Ka Accident Saale, dhyaan se sun. ASVspoof dataset mein asli audio clips kam hote hain aur nakli clips bohot zyada hote hain. Jaise Mumbai local mein general compartment mein log hota hai, aur AC compartment mein thode se. Agar tune model ko seedha train kar diya, toh wo banaega ki "Bhai, sab kuch nakli hi hai, kyunki maine zyada nakli dekha hai." Yeh Data Imbalance kehlata hai. Iska solution hai Oversampling ya Undersampling. Oversampling mein kam wale data ko duplicate karke zyada karte hain. Undersampling mein zyada wale data ko kaat ke kam karte hain. Tujhe apne model ko balanced diet dena hai, warna wo bias ho jaayega. ⚠️ Common …

3. Spectrograms Aur MFCCs Ki Real Game, Chamka Kya?

Abe sun, kal raat 3 baje ek bhandar local train ki general compartment mein chadhha. Ek bhai saab apne phone pe kisi "influencer" ki video dekh raha tha—bhai ka voice itna smooth, itna perfect, ki laga sachhi mein koi God bol raha hai. Phir ekdum se side wale uncle ne bola, "Ye toh fake hai saale, AI ne banaya hai." Bhandar bhai ka bharosa toot gaya, dil tuta, aur usne phone zameen pe patak diya. Tu soch raha hoga, "Bhai, iska mera deepfake detector se kya lena-dena?" Saala lena-dena tab hai jab tera model us fake voice ko pakad ke uske muh pe maar na sake. Chapter 2 tak humne ASVspoof jaisa solid dataset collect kiya, Bonafide aur Spoofed labels separate kiye, aur Data Augmentation jaise jugaad se Data Imbalance theek kiya. Ab raw audio files tere paas hai. Par seedha raw audio ko neural network mein daalega toh wo utna hi samjhega jitna tera dost Hindi mein Spanish movie samajhta hai—zero percent, BC. Tujhe is audio ko aise roop mein convert karna padega jo tera model "dekh" sake. Neural networks aankhon wale hote hain, kaanon wale nahi. Hum audio ko image bana denge. Yahan se shuru hoti hai Spectrograms aur MFCCs ki asli game. Collar pakad ke baith, aaj tera dimag thoda aayega, par dimag aayega toh hi model banega. Asli Funda (Core Concepts) 1. Spectrogram Kya Hai? Chai Ki Tapri Wali Example Dekh bhai, audio ek 1D signal hai. Time vs Amplitude ka graph. Jaise ECG machine pe line upar-niche hoti hai na? Waisa. Par is 1D signal mein kaunsa frequency (sur) kab aayi, ye dekhna mushkil hai. Aise samajh, jaise tu chai ki tapri pe 5 log bol rahe hain. Ek uncle politics pe, ek ladki apne boyfriend pe, ek bhai cricket pe. Tu ek sath sab sun raha hai. Terre dimag mein ek mixed noise jaa raha hai. Par agar tujhe pata karna ho ki "cricket ki baat kis time pe hui aur kitni zor se", toh tu us audio ko chhote-chhote hisson mein kaatega (frames), aur har hisse mein kaunsa frequency dominant hai ye dekhega. Isko technical bhasha mein Short-Time Fourier Transform (STFT) bolte hain. Hum audio ko chhote windows mein kaat-te hain, aur har window ka frequency spectrum nikalte hain. Jab ye saare frequency slices ko ek sath 2D grid pe rakh de—X-axis pe Time, Y-axis pe Frequency, aur pixel ki intensity (color) pe Amplitude—toh ban jata hai ek Spectrogram. Ek image! Saale, ye image itni powerful hai ki isme audio ka pura DNA chhupa hota hai. Real voice aur fake voice ke spectrograms mein aisi jagah hote hain jo human ear se nahi pakde jaate, par CNN (Convolutional Neural Network) jaise model in …

4. CNN Aur Transformer Ka Jugaad, Boss Level Architecture

Abe sun, raat ke 3 baj rahe hain. Tu chat pe dekh raha hai tera dost startup mein funding le gaya, aur tu abhi tak model.fit() pe atka hua hai. Sharma ji ka beta already 99% accuracy pe pahunch gaya, aur tu abhi soch raha hai ki "Bhai, CNN kya hota hai? Ye Transformer kya maachod raha hai internet pe?" Saale, dimaag ki batti jalaa. Pichle teen chapters mein tu ne Conda ka setup kiya, ASVspoof ka data utara, aur Librosa se spectrograms aur MFCCs nikal liye. Tune audio ki image bana li. Ab kya? Ab uss image ko dekh ke model ko bolna hai, "Bhai, ye aawaz asli hai ya nakli?" Ye chapter nahi, tera actual boss fight hai. Agar isme architecture gadbad kiya, toh tera model local train ki 6:00 PM ki Borivali local jaisa behave karega—bhagda, overfitting, aur kahin pahunchne wala nahi. Pencil kaata, dimaag lagaa, aur padh. Asli Funda (Core Concepts) Dekh bhai, deep learning ka model banane mein koi jadoo nahi hai. Ye exactly waisa hi hai jaise tu Mumbai mein ek ghar dhoond raha hai. Tujhe location chahiye (Data), map chahiye (Architecture), aur dimaag chahiye (Training). Humara data ASVspoof ka spectrogram hai. Ab is spectrogram ko samajhne ke liye do badmaash hire karenge: ek hai CNN, doosra hai Transformer. 1. CNN: The Local Gunda (1D vs 2D) Saale, sab pehle aata hai Convolutional Neural Network (CNN). Ye model ka local gunda hai. Ye dekhta hai ki image ke ek chhote se patch mein kya hai. Tune spectrogram banaya hai na? Wo ek 2D matrix hai—x-axis pe time, y-axis pe frequency, aur uske andar color se energy dikh rahi hai. Ab ye 2D image hai, toh hum use karenge 2D CNN. Agar tu MFCCs jaisa 1D sequence nikal raha hai, toh 1D CNN use karega. Humara main focus 2D CNN hoga kyunki visual features (peaks, valleys, patterns) pakadne mein 2D CNN ki aukaat alag level ki hai. CNN kaise kaam karta hai? Tu ek chhota sa window (filter/kernel) spectrogram pe ghuma raha hai. Jaise autorickshaw wala meter check karta hai—har chhote area ka feature nikalta hai. "Yahan frequency high hai, yahan drop hai, yahan noise hai." Ye local patterns pakadta hai. 💡 Pro Tip: 2D CNN ke filters tujhe fake voice mein aane wale 'glitches' ya synthetic artifacts dhoondhne mein madad karenge. Fake audio generate karne wale AI models chup jaate hain, par local frequency patterns mein unka paaon phisal jaata hai. CNN usi ko pakadta hai. 2. Transformer Encoder: The Big Boss Ab CNN ne local patterns dhoondh liye. Par ek aawaz asli hai ya nakli, ye sirf local patterns se nahi pata chalta. Context chahiye. "Ye shabd kaise bol gaya? Iske pehle kya …

5. Train Karna Hai Aashiq Ki Tarah, Patience Rakh Guru

Abe sun na, tera Model 4 wala architecture toh bana liya na? CNN aur Transformer ka jugaad lagaya, layers thoki, sab kuch mast hai. Par ekdum raw biryani hai abhi ye. Masala daala, sabzi kaati, lekin chulha pe nahi pakaya. Dekh bhai, raat ke 3 baje code likhte hue jab tu 5th cup coffee pee raha hota hai, tab dimaag ekdum slow ho jaata hai. Aise hi tera neural network bhi bina training ke ekdum ullu bantha hua hai. Architecture bana liya matlab gaadi ka body ready hai, engine fit kiya, par petrol daala nahi, chala nahi toh kya karega? Kachra hi rahega na? Saale, Chapter 4 mein tune wahi kiya. Ab Chapter 5 mein petrol daalenge, engine garam karenge, aur isko aashiq ki tarah train karenge. Ek dum patience se. Nahi toh model overfit hoke tera test data pe aukaat dikhayega, aur tu client ke aage sharma ji ke bete ki tarah sharmaaega. Chal, collars theek kar, chai ki tapri pe baith, aur training shuru kar! Asli Funda (Core Concepts) Training ka funda simple hai—Loss Function (galti batata hai), Optimizer (galti theek karta hai), aur Scheduler (speed control karta hai). Teeno ka jugaad chahiye. Ek bhi gadbad hua toh BC tera model train hote hue zindagi nikal jaayegi. 1. Loss Function: Galti Ka Hisaab Bhai ye aise samajh, jaise tu Mumbai local mein Virar fast pakadhta hai. Bogie mein itni bheed hai ki tu andar ghusa, par tere pet pe jo chot lagi, wo teri "Loss" hai. Model bhi aise hi prediction karta hai. Agar model bolta hai "Bhai ye audio asli hai" aur reality mein wo nakli (Spoofed) hai, toh model ko thappad lagana padega. Isi thappad ko mathematical form mein dikhane wale ko Loss Function bolte hain. Kyunki humara problem Classification wala hai (Asli vs Nakli), hum use karenge Cross-Entropy Loss. Kyun? Kyunki ye classification ka baap hai. ⚠️ Common Mistake: Abe oye, tera data ASVspoof 2019 aur 2021 se aaya hai. Tune Chapter 2 mein Data Imbalance dekha tha na? Bonafide (Asli) aur Spoofed (Nakli) ka ratio ekdum off hai. Agar tune simple Cross-Entropy lagaya, toh model chup-chap sab kuch "Spoofed" bol dega kyunki usse wahi easy lagega. Accuracy 90% dikhega, par asli audio aate hi model ki maa-been ho jaayegi. Isliye Class Weights use kar! Nakli wale ko kam weight do, asli wale ko zyada, taaki model dono pe equal dhyaan de. PyTorch mein iska code dekh, ekdum simple hai: Samjha? Agar ye weight ka chakar nahi rakha, toh tera model biased ho jaayega. Jaise arranged marriage mein ladke walo ki demand zyada hoti hai, model bhi waise hi majority class ki sunega. 2. Optimizer: Mechanic Jo Galti Theek Kare Loss function ne batadiya ki …

6. Real-Time Ka Scene Hai, Latency Kam Kar Bhidu

Abe sun, tera model train ho gaya, 99% accuracy aa gayi, tu khud ko AI ka baap samajh raha hai? "Bhai maine toh kar liya!" BC, ekdum waise reh gaya jaise Sharma ji ke bete ne IIT crack kiya par zebra crossing pe nahi utar sakta. Gully ka asli don model ko paper pe nahi, sadak pe chalata hai. Tera model agar 2 second mein bole "Bhai ye audio nakli hai" jab tak takke wada pav thanda ho gaya, toh uska kya faayda? User gaali deke chhod dega. Real-time ka scene alag hi hai, bhidu. Yahan koi traindataloader nahi rookega, koi epoch nahi chalayega. Mic se aawaz aayi, model ko turant bolna hoga "Asli hai!" ya "Nakli hai!". Agar latency (deri) zyada hui, toh tera deepfake detector ekdum bekaar, ekdum ghaati. Saale, ye itna important hai ki isko skip kiya toh tera career ka barbaad ho jaayega, pakka. Chal, collar thik kar, aur dhyaan se sun. Asli Funda (Core Concepts) Dekh bhai, real-time audio stream aise samajh, jaise Mumbai local train ki bheed. Log (audio samples) lagaatar aa rahe hain. Agar tu poori bheed ko ek saath andar aane dega (poora audio ek saath process karega), toh platform foot pe aag lagegi, system hang ho jayega. Isliye hum logon ko batches mein, yaani chunks mein andar bhejte hain. 1. Audio Chunking: Katne Ka Asli Jugaad Tere paas mic se continuous audio aa raha hai. Tujhe isko chhote-chhote tukdon mein todna hai. Jaise biryani ke dabbe mein ek saath 5 kilo nahi khaata, thoda-thoda khaata hai, waise hi. Par yahan ek catch hai. Agar tu audio ko seedha kaat dega—matlab 1 second ka piece, phir agle 1 second ka piece—toh jodhne par aawaz mein "click" aayega. Aawaz toot-ti dikhegi. Ye hota hai kyunki tukdon ke beech ka connection toot jaata hai. Iska solution hai Overlap. 💡 Pro Tip: Overlap ka matlab hai agla chunk pichle chunk ke thoda pehle se shuru ho. Jaise tu 1 second ka audio le, aur agla chunk 0.5 second pehle se shuru kar. Isse aawaz mein koi "cut" nahi aayega aur model ko context bhi milega. Sharma ji ka beta ye nahi batayega, ye gully ka asli funda hai. Window Size aur Hop Length: - Window Size: Ek baar mein kitna audio process kar raha hai (e.g., 1 second ya 2 seconds). Deepfake detection ke liye 2-3 seconds ka window best hai, kyoshi nakli aawaz pakadne ko thoda context chahiye. - Hop Length: Agla window kahan se shuru hoga. Agar window 2 second ka hai aur hop 1 second ka hai, toh 50% overlap hai. 2. Buffering: Rasta Nahi Rokna Bhai Ab soch, tera model 2 second ka audio chunk process karne mein 0.5 second leta …

7. Deploy Karke Dikhao, FastAPI Aur Docker Ki Chakri

Abe sun, raat ke 3 baj rahe hain. Tu chai peeta hai, code karta hai, aur sochta hai "BC mera model toh 99% accuracy de raha hai, main toh AI ka baap hoon." Lekin sach bolun? Tera model abhi tak ek 'kunwaara kunwara' hai. Apne room ke andar, Jupyter notebook ke andar, kuch dikh raha hai. Bahar laya toh darr ke maare jaan pe aa jaayega. Jab tak tera deepfake detector kisi server pe deploy nahi ho jata, kisi unknown user ki audio file receive nahi kar raha, aur 0.5 second mein "Ye nakli hai, saale!" ka verdict nahi de raha—tab tak tu kuch nahi hai. Tu woh Sharma ji ka beta hai jo exam mein 100% laata hai par zindagi mein auto chalana nahi aata. Chapter 6 mein humne real-time latency ka scene sambhala tha. Ab waqt aagaya hai is model ko zindagi mein utaarne ka. Is chapter mein hum FastAPI ka engine banayenge, model ko ONNX ki gadi mein baithayenge, aur Docker ke container mein band karke duniya bhar mein bhejenge. Kasam se, agar ye chapter skip kiya toh tera 6 mahine ki mehnat ka barbaad ho jaayega, pakka. Asli Funda (Core Concepts) Dekh bhai, deployment ka funda simple hai lekin log isko overcomplicate karte hain. Tu biryani order karta hai Zomato se. Zomato app tera order receive karti hai (API), kitchen mein chef biryani banata hai (Model inference), aur delivery boy tujhe pakha deta hai (Response). Humko exactly yahi karna hai, bas biryani ki jagah audio file hogi. 1. FastAPI: The Zomato Delivery Boy Python mein API banane ke 100 tareeke hain. Flask tha, Django tha, par sab slow hain. Humko speed chahiye, kyunki real-time audio detection mein latency matlab ijjat ka faluda. FastAPI aise hai jaise local train ka fast passenger—chalta hi engine start, rukta nahi. FastAPI asynchronous hai. Matlab agar 10 log ek saath audio bhej rahe hain, toh ek ka wait karke dusre ko block nahi karega. Saala sabko ek saath handle karega. 💡 Pro Tip: FastAPI automatically Swagger UI generate karta hai. Yaani tera API banate hi /docs endpoint pe ek fancy interface mil jaayega jisme tu apna audio upload karke test kar sakta hai. Postman ki zaroorat hi nahi padti, bhai! 2. ONNX Conversion: Sharma Ji Ka Beta Banke Model Ko Fast Bana Abhi tera model PyTorch mein hai. PyTorch research ke liye best hai, training ke liye best hai, par production ke liye thoda heavy hai. Jaise woh ghar ka khana—swaad hai par packing aur delivery mein time lagta hai. Yahan aata hai ONNX (Open Neural Network Exchange). Ye ek format hai jo tera PyTorch model ko ek lightweight, optimized file mein convert kar deta hai. ONNX Runtime C++ ke base …

8. Frontend Laga Ke Public Ko Batao, Asli Bheja Khol

Abe sun, tera startup idea leak ho gaya kya? Pichle 7 chapters mein tune Conda set kiya, ASVspoof ka data download kiya, Librosa se Spectrograms banaye, CNN-Transformer ka model banaya, train kiya, real-time latency kam ki, aur FastAPI + Docker se API bhi deploy kar di. Bhai, tu abhi tak wo Sharma ji ka beta hai jo raat ko 3 baje library mein padhta hai par college fest mein kabhi nahi jaata. Sab kuch hai, par duniya ko dikhta nahi! Model backend pe so raha hai, par frontend nahi hai. Aaj is chapter mein, tu us model ko public ke saamne laaunga. Asli bheja kholne ka time hai. Asli Funda (Core Concepts) Dekh bhai, ye aise samajh. Tu Mumbai local train mein hai, peak hour. Virar se Churchgate. Tu andar ghussa, par darwaza band hai. Log bahar latke hain. Tera model wo andar wala insaan hai jo poora system chala raha hai, par bahar wale log (users) tujhe nahi dekh pa rahe. Frontend wo darwaza hai, wo window hai jahan se log tujhe dekhenge, tujhse interact karenge. Agar frontend bakwaas hua, toh andar ka kitna bhi gold ho, log "bhai ye toh ghaati hai" bolke utar jaayenge. Garbage Out wala scene ho jaayega tera, chahe input kitna bhi Bonafide (Asli) kyu na ho. 1. Streamlit vs React: Autorickshaw vs Mercedes Tu soch raha hoga, "Bhai, frontend kaise karu? Main toh Python ka aadmi hoon, JavaScript mera dushman hai." Chal, solution do. Streamlit (Autorickshaw): Bhai, agar tu jaldi ghar pohanchna hai, aur budget kam hai, toh autorickshaw best hai. Streamlit exactly wahi hai. Python mein 10 lines likho, aur UI ready. Data scientists ka dream tool. Par iski limit hai. Real-time microphone input handle karna Streamlit mein thoda tricky hota hai, kyunki ye server-client model mein data flow ko thoda awkwardly handle karta hai. Par MVP (Minimum Viable Product) ke liye? Best. React (Mercedes): Agar tu dikhawa karna chahta hai, smooth UX chahiye, aur real-time audio stream ko pixel-perfect control karna hai, toh React seekho. Par tera dimag ka dahi ban jaayega agar tune pehle JS nahi kiya. Hum is chapter mein Streamlit ka jugaad karenge, kyunki asli backend tera PyTorch model hai, frontend sirf mukaam dekhta hai. 2. Live Microphone Input: The Jugaad Real-time mic input handle karna is the real boss-level task. Browser security itni strict hai ki jo teri GF ka attitude, mic ko directly access nahi kar sakta. Browser tujhe permission maangni padti hai, aur tab audio stream process hoti hai. Streamlit pe mic input ke liye streamlit-webrtc ka use karna padta hai. Ye package browser ke WebRTC protocol ko Streamlit ke server ke saath jodta hai. WebRTC ka funda simple hai: Browser mic se audio uthata …

Continue learning