<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Apparao Mulpuri on Medium]]></title>
        <description><![CDATA[Stories by Apparao Mulpuri on Medium]]></description>
        <link>https://medium.com/@apparaomulpuri?source=rss-f772bca567f6------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/0*rsxIEZvH_unZuceF.jpg</url>
            <title>Stories by Apparao Mulpuri on Medium</title>
            <link>https://medium.com/@apparaomulpuri?source=rss-f772bca567f6------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sun, 22 Nov 2026 03:59:12 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@apparaomulpuri/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[How I Built a Smart Voice Detection System for My Android Streaming App]]></title>
            <link>https://medium.com/@apparaomulpuri/how-i-built-a-smart-voice-detection-system-for-my-android-streaming-app-6cde3847af70?source=rss-f772bca567f6------2</link>
            <guid isPermaLink="false">https://medium.com/p/6cde3847af70</guid>
            <category><![CDATA[android]]></category>
            <category><![CDATA[streaming]]></category>
            <category><![CDATA[voice-recognition]]></category>
            <category><![CDATA[audio-processing]]></category>
            <category><![CDATA[echo-cancellation]]></category>
            <dc:creator><![CDATA[Apparao Mulpuri]]></dc:creator>
            <pubDate>Wed, 16 Apr 2025 13:17:44 GMT</pubDate>
            <atom:updated>2025-04-16T13:17:44.536Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*u2-kiM2wDsSVudo9" /><figcaption>Photo by <a href="https://unsplash.com/@dylu?utm_source=medium&amp;utm_medium=referral">Jacek Dylag</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>When I started developing my Android streaming app, I hit a frustrating problem that many developers face: <strong>echo</strong>. You know that awful feedback loop when your microphone picks up sound from speakers in the same room? Yeah, that nightmare.</p><p>After countless late nights and way too much coffee, I finally cracked it. Here’s how I built a voice detection system that actually works in the real world</p><h3>The Problem I Was Solving</h3><p>I needed my streaming app to detect when a human was speaking while ignoring audio coming from a TV in the same room. Most solutions either required expensive hardware (not an option) or just didn’t work well enough.</p><p>My users were complaining: “Your app keeps picking up my TV and creating echo!” Something had to change.</p><h3><strong>My Approach: Smart Voice Detection</strong></h3><p>Instead of relying on a single technique, I built a multi-layered detection system. The magic happens in a class I created called `VoiceDetector` that analyzes audio in real-time.</p><p>The core idea is pretty straightforward: human speech has unique characteristics that distinguish it from TV audio or background noise. If I could identify those patterns, I could solve the problem.</p><h3>How My Algorithm Actually Works</h3><h4>1. Audio Fingerprinting</h4><p>I break down each audio chunk into frequency bands and look for patterns that match human speech:</p><pre>// Human voices have more energy in certain frequency ranges<br>if (energyBands[2] &gt; energyBands[6] * 1.2 &amp;&amp; energyBands[1] &gt; energyBands[7] * 1.5) {<br>    // This looks like human speech!<br>}kot</pre><h4>2. Spotting Those Sharp Consonants</h4><p>One thing I discovered in my testing: human speech has sharp spikes (like when you say words with ‘p’, ‘t’, or ‘k’ sounds) that most TV audio doesn’t have:</p><pre>// Calculate the difference between peaks and average levels<br>val peakAvgRatio = peakValue / avgValue<br><br>// Human speech typically has ratios above 3.5<br>if (peakAvgRatio &gt; 3.8) {<br>    humanVoiceConfidence += 0.3<br>}</pre><p>After running hundreds of tests with different voices, I found that human speech consistently shows peak-to-average ratios above 3.5, while TV audio usually stays below 3.0.</p><h4>3. Learning from the Environment</h4><p>One breakthrough came when I realized my system needed to adapt to each room. I added a calibration phase that measures the baseline noise level during the first second:</p><pre>// This runs when the app starts<br>if (frameCount &lt; 30) {<br>    totalEnergyBaseline += totalEnergy / 30f<br>    frameCount++<br>    // Once we have enough samples, we&#39;re initialized<br>}</pre><p>This simple addition made the system work much better across different environments.</p><h3>The Real-World Results</h3><p>After implementing my voice detector, I needed to actually handle the audio transmission. I ended up with this approach:</p><pre>if (isHumanVoice) {<br>    // Send the actual audio<br>    outputStream.write(buffer, 0, bytesRead)<br>} else {<br>    // Send near-silence to maintain the stream<br>    outputStream.write(silenceBuffer, 0, bytesRead)<br>}</pre><p>This effectively breaks the echo loop. When someone speaks, their voice goes through. When they stop talking, the system sends near-silence, preventing any TV sound from creating feedback.</p><h4>Optimizing the Audio Processing</h4><p>Getting the FFmpeg audio filters right took a lot of trial and error. I started with complex filter chains that tried to do too much:</p><pre>// My first overly complicated attempt<br>&quot;bandreject=f=300:width_type=h:width=200,bandreject=f=500:width_type=h:width=200,agate=threshold=0.25...&quot;</pre><p>But I found that simpler is often better:</p><pre>// What actually worked well<br>&quot;volume=4.0,highpass=f=200:p=1,equalizer=f=2500:width_type=h:width=1000:g=3&quot;</pre><p>The key insights:<br>- A strong volume boost helps with quieter speakers<br>- A gentle high-pass filter removes room rumble<br>- Enhancing the 2–3kHz range makes speech clearer<br>- Compressing dynamic range makes voices more consistent</p><h3>Lessons Learned the Hard Way</h3><p><strong>1. More complex isn’t always better</strong>. My early attempts used 7+ audio filters in a chain, but they often dropped audio entirely. Simplifying improved reliability dramatically.</p><p>2.<strong> Streaming configuration matters</strong>. Increasing the SRT latency from 10ms to 200ms made a huge difference in audio quality:</p><pre>srtUrl + &quot;?mode=caller&amp;latency=50&amp;tlpktdrop=0&quot;</pre><p>3. <strong>Log everything</strong>. Adding detailed logging saved me countless hours of debugging:</p><pre>Log.d(TAG, &quot;Voice detection: $voiceFrames/$totalFrames frames (${String.format(&quot;%.1f&quot;, voiceRatio)}%)&quot;)<br>Log.d(TAG, &quot;Confidence: Human=${String.format(&quot;%.2f&quot;, humanConfidence)}, TV=${String.format(&quot;%.2f&quot;, tvConfidence)}&quot;)</pre><h3>What’s Next?</h3><p>I’m still improving the system. Some ideas I’m exploring:<br>- Using machine learning to better classify different audio types<br>- Implementing a formant tracker to identify specific voices<br>- Creating user-adjustable sensitivity controls</p><h3><strong>Final Thoughts</strong></h3><p>Building this voice detection system was harder than I expected, but the results were worth it. My app now reliably transmits human speech while ignoring TV audio, all without expensive hardware.</p><p>If you’re facing similar challenges in your streaming app, I hope my approach gives you some ideas. Sometimes the most elegant solutions come from combining simple techniques in clever ways.</p><p>What audio processing challenges have you faced in your projects? I’d love to hear your experiences in the comments!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=6cde3847af70" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Meta Unveils LLaMA 3.1: The Frontier of Open Source AI]]></title>
            <link>https://medium.com/@apparaomulpuri/meta-unveils-llama-3-1-the-frontier-of-open-source-ai-460ad36bf377?source=rss-f772bca567f6------2</link>
            <guid isPermaLink="false">https://medium.com/p/460ad36bf377</guid>
            <category><![CDATA[openai]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[llama-3]]></category>
            <category><![CDATA[llm]]></category>
            <dc:creator><![CDATA[Apparao Mulpuri]]></dc:creator>
            <pubDate>Wed, 24 Jul 2024 06:52:35 GMT</pubDate>
            <atom:updated>2024-07-24T06:52:35.424Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*qY9zCeW3BoZgKlEiuBhm-Q.jpeg" /></figure><p>Meta has introduced LLaMA 3.1, its most advanced AI model yet, setting a new benchmark in open-source AI. This release represents a significant leap in natural language understanding and generation, with far-reaching applications across various domains.</p><h3>Technical Innovations in LLaMA 3.1</h3><p>LLaMA 3.1 integrates several groundbreaking advancements:</p><p><strong>Extended Context Length:</strong></p><p>LLaMA 3.1 supports a context length of up to 128K, allowing for a more comprehensive understanding and generation of long-form content.</p><p><strong>Multilingual Capabilities:</strong></p><p>The model supports eight languages, enhancing its usability in diverse linguistic contexts and applications.</p><p><strong>Enhanced Architecture:</strong></p><p>Utilising a standard decoder-only transformer model, LLaMA 3.1 incorporates minor adaptations to maximise training stability and efficiency. The model was trained on over 15 trillion tokens using more than 16,000 H100 GPUs, a significant technical feat.</p><p><strong>Synthetic Data Generation and Model Distillation:</strong></p><p>This model introduces unprecedented capabilities in synthetic data generation, enabling the training of smaller models with high-quality synthetic data. Model distillation at this scale is now possible, providing a new dimension for AI development.</p><p><strong>Quantization:</strong></p><p>The models are quantized from 16-bit (BF16) to 8-bit (FP8) numerics, reducing computational requirements and enabling efficient deployment on single server nodes.</p><h3>Performance Metrics</h3><p>LLaMA 3.1 has been rigorously evaluated across over 150 benchmark datasets and extensive human evaluations. Key performance highlights include:</p><p><strong>Competitive Benchmarks:</strong></p><p>The 405B model rivals leading closed-source models like GPT-4 and Claude 3.5 Sonnet in general knowledge, tool use, and multilingual translation tasks.</p><p><strong>Improved Reasoning:</strong></p><p>Enhanced reasoning capabilities make LLaMA 3.1 suitable for advanced applications such as long-form text summarization, conversational agents, and coding assistants.</p><h3>Real-World Applications</h3><p>The advancements in LLaMA 3.1 unlock numerous real-world applications:</p><p><strong>Conversational AI:</strong></p><p>LLaMA 3.1 provides contextually relevant and human-like interactions, significantly improving virtual assistants and chatbots.</p><p><strong>Content Generation:</strong></p><p>The model’s nuanced text generation capabilities are ideal for automated content creation in journalism, marketing, and creative writing.</p><p><strong>Educational Tools:</strong></p><p>Enhanced contextual understanding supports the development of interactive and personalised educational content.</p><p><strong>Research and Data Analysis:</strong></p><p>Advanced comprehension and generation capabilities facilitate data synthesis and hypothesis generation in research fields.</p><h3>The LLaMA System</h3><p>LLaMA models are designed to work within a comprehensive system, incorporating external tools and components. Meta’s vision extends beyond foundational models to provide developers with a flexible system for creating custom solutions. Key components include:</p><p><strong>LLaMA Guard 3:</strong></p><p>A multilingual safety model designed to enhance security and responsible AI use.</p><p><strong>Prompt Guard:</strong></p><p>A prompt injection filter to safeguard against malicious inputs.</p><h3>Ecosystem and Openness</h3><p>Meta has partnered with over 25 industry leaders, including AWS, NVIDIA, Databricks, and more, to provide robust support for LLaMA 3.1 from day one. The open-source nature of LLaMA 3.1 ensures that developers can fully customise the models for their specific needs, fostering innovation and broadening the accessibility of AI technology.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=460ad36bf377" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Apple WWDC 2024: Revolutionising Software and Hardware with AI]]></title>
            <link>https://medium.com/@apparaomulpuri/apple-wwdc-2024-revolutionising-software-and-hardware-with-ai-c2f47b009e10?source=rss-f772bca567f6------2</link>
            <guid isPermaLink="false">https://medium.com/p/c2f47b009e10</guid>
            <category><![CDATA[ipad]]></category>
            <category><![CDATA[apple]]></category>
            <category><![CDATA[apple-tv]]></category>
            <category><![CDATA[apple-watch]]></category>
            <category><![CDATA[wwdc]]></category>
            <dc:creator><![CDATA[Apparao Mulpuri]]></dc:creator>
            <pubDate>Tue, 11 Jun 2024 03:59:01 GMT</pubDate>
            <atom:updated>2024-06-11T04:00:05.537Z</atom:updated>
            <content:encoded><![CDATA[<p>Apple’s Worldwide Developers Conference (WWDC) 2024, held on June 10, unveiled a plethora of updates and new features that are set to redefine the user experience across its devices. From significant advancements in AI to enhanced customization options, here’s a comprehensive look at what Apple has in store.</p><figure><img alt="Apple WWDC 2024" src="https://cdn-images-1.medium.com/max/1024/1*wTPJpTPDg1IrC7IlkGQ5aw.jpeg" /><figcaption>Image source: AI Image generated with Stable diffusion</figcaption></figure><h3>iOS 18: A Giant Leap in AI and Customisation</h3><p>Apple’s iOS 18 is arguably the biggest update for iPhone users yet, featuring an array of new functionalities designed to enhance user experience and device performance.</p><h4>Key Features:</h4><ol><li><strong>Apple Intelligence</strong>:</li></ol><ul><li><strong>Custom Emoji</strong>: Generate emojis based on text descriptions.</li><li><strong>Photo Editing</strong>: Edit photos through written or spoken commands.</li><li><strong>Transcription</strong>: Convert audio recordings into text.</li><li><strong>Smart Summaries</strong>: Summarise web content in Safari.</li></ul><p><strong>2. Home Screen Customisation</strong>:</p><ul><li>Users can place icons and widgets anywhere on the home screen.</li><li>Change app icon colours, including a dark mode option.</li></ul><p><strong>3. Enhanced Messaging</strong>:</p><ul><li><strong>RCS Support</strong>: Rich Communication Services for improved messaging between iOS and Android.</li><li><strong>iMessage Effects</strong>: New effects and animations, with the potential for rich text formatting like bold and italics.</li></ul><p><strong>4. Security Enhancements</strong>:</p><ul><li>Lock individual apps using Face ID for an added layer of privacy and security.</li></ul><h3>macOS 15: AI-Powered Productivity</h3><p>macOS 15 brings powerful new features to the Mac, leveraging the Neural Engine for enhanced AI capabilities.Notable Updates:</p><p><strong>AI Integration</strong>:</p><ul><li>Enhanced AI features in productivity apps like Keynote and Pages.</li></ul><p><strong>Redesigned System Settings</strong>:</p><ul><li>Improved organisation for easier navigation and user experience.</li></ul><p><strong>Updated Siri</strong>:</p><ul><li>A new look for Siri’s menu bar icon and improved functionality for voice commands.</li></ul><h3>iPadOS 18: Bridging the Gap with iOS</h3><p>iPadOS 18 includes many features from iOS 18, enhancing the functionality and accessibility of iPads.</p><p><strong>Key Features:</strong></p><ol><li><strong>AI and Customisation</strong>:</li></ol><ul><li>Similar AI features as iOS 18, including transcription and smart summaries.</li><li>Enhanced home screen customisation.</li></ul><p><strong>2. Accessibility</strong>:</p><ul><li>New eye-tracking feature for hands-free control of the iPad.</li></ul><h3>visionOS 2: Expanding the Vision Pro Experience</h3><p>The first major update to VisionOS brings several improvements designed to enhance the Vision Pro headset experience.</p><p><strong>Major Enhancements:</strong></p><ol><li><strong>Native Apps</strong>:</li></ol><ul><li>More first-party apps built specifically for VisionOS.</li></ul><p><strong>2. Live Captions</strong>:</p><ul><li>Real-time captions for spoken dialogue in apps and in-person conversations.</li></ul><p><strong>3. International Rollout</strong>:</p><ul><li>Expansion of Vision Pro availability to more countries.</li></ul><h3>watchOS 11 and tvOS 18: Incremental Yet Significant Improvements</h3><h3>watchOS 11:</h3><ul><li>Focus on health and fitness, with new coaching features and improved tracking.</li></ul><h3>tvOS 18:</h3><ul><li>Enhanced Apple TV features include accessibility updates and potential new hardware like a revamped Apple TV box.</li></ul><h3>Hardware Announcements: Mac Studio with M3 Ultra</h3><p>While the keynote was predominantly software-focused, Apple hinted at an updated Mac Studio featuring the powerful M3 Ultra chip, promising substantial performance upgrades.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c2f47b009e10" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Decentralised Artificial Intelligence (DAI): Reshaping the AI Landscape]]></title>
            <link>https://medium.com/@apparaomulpuri/decentralised-artificial-intelligence-dai-reshaping-the-ai-landscape-37b3b756b28d?source=rss-f772bca567f6------2</link>
            <guid isPermaLink="false">https://medium.com/p/37b3b756b28d</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[decentralized-ai]]></category>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[stable-diffusion]]></category>
            <dc:creator><![CDATA[Apparao Mulpuri]]></dc:creator>
            <pubDate>Wed, 27 Mar 2024 09:44:38 GMT</pubDate>
            <atom:updated>2024-03-27T09:44:38.876Z</atom:updated>
            <content:encoded><![CDATA[<p>Artificial intelligence (AI) has become an integral force, transforming industries and redefining our relationship with technology. Yet, the current landscape is dominated by centralized AI systems. These systems, controlled by a handful of tech giants, bring concerns about data privacy, algorithmic bias, and the concentration of power. Enter the concept of Decentralised Artificial Intelligence (DAI), a paradigm shift promising to democratise AI and usher in a new era of transparency, fairness, and resilience.</p><p><strong>Understanding DAI: What Makes It Different</strong></p><p>Traditional AI models are often trained on massive datasets housed in centralised data centres. A few influential organizations control the creation and application of these models. In contrast, DAI utilises the principles of blockchain technology to distribute AI processes. Here’s how it works:</p><ul><li><strong>Distributed Data:</strong> In a DAI system, data is no longer confined to a single repository. Instead, data remains where it’s generated — on personal devices, at the network edge — contributing to training without compromising privacy.</li><li><strong>Network of Nodes:</strong> AI models are trained across a network of computers or nodes, rather than on a single, powerful server. This eliminates single points of failure and reduces the risk of manipulation.</li><li><strong>Consensus-Driven Decision Making:</strong> Decisions about model updates or predictions are made through consensus mechanisms among the nodes. This mitigates biases that may be present in a centralised system.</li></ul><p><strong>The Benefits of Decentralised AI</strong></p><ul><li><strong>Enhanced Privacy:</strong> DAI allows individuals and organisations to retain control over their data by keeping it local. Sensitive data doesn’t travel to centralised systems, drastically minimising privacy risks.</li><li><strong>Reduced Bias:</strong> With diverse datasets and contributions from a wider pool of participants, DAI models are less likely to reflect the biases of a limited group of developers or data sources.</li><li><strong>Increased Accessibility:</strong> The democratised nature of DAI removes the high entry barriers associated with centralised AI. Individuals and smaller organisations can participate in shaping AI development.</li><li><strong>Robustness and Security:</strong> Decentralisation eliminates single points of failure, making DAI systems inherently more resilient against cyberattacks or manipulation attempts.</li></ul><p><strong>DAI in Action: Real-World Applications</strong></p><ul><li><strong>Federated Learning:</strong> This DAI technique enables models to be trained on decentralised datasets, ensuring data privacy. It has applications in healthcare, where training AI models on sensitive patient data can be done without compromising individual information.</li><li><strong>Decentralised Marketplaces:</strong> DAI platforms are establishing marketplaces where AI models and datasets can be bought, sold, or shared. This fosters collaboration and drives innovation in a transparent environment.</li><li><strong>Personalised AI Experiences:</strong> DAI paves the way for truly personalised AI services. Individuals can fine-tune models based on their own data, leading to AI that understands their unique needs and preferences.</li></ul><p><strong>The Future of Decentralised AI</strong></p><p>Decentralised AI is a rapidly evolving field brimming with potential. While challenges like scalability and the need for robust governance mechanisms persist, the promise is undeniable:</p><ul><li><strong>The Rise of Ethical AI:</strong> DAI can usher in an era of more trustworthy and transparent AI, where biases are reduced and responsible use is prioritised.</li><li><strong>Unlocking New Possibilities:</strong> DAI applications will likely emerge in sectors like supply chain management, autonomous vehicle networks, and more.</li><li><strong>Human-AI Collaboration:</strong> DAI will foster new ways for humans and AI to work in tandem, building trust and enabling AI to be used with greater effectiveness.</li></ul><p>Decentralised Artificial Intelligence is not a distant dream but a transformative force set to reshape the AI landscape. As DAI evolves, we can anticipate a more equitable distribution of power, enhanced privacy, and a new wave of innovative AI applications that benefit all.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=37b3b756b28d" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Understanding iOS App Store Statuses]]></title>
            <link>https://medium.com/@apparaomulpuri/understanding-ios-app-store-statuses-7ef1c6dc01c4?source=rss-f772bca567f6------2</link>
            <guid isPermaLink="false">https://medium.com/p/7ef1c6dc01c4</guid>
            <category><![CDATA[app-state]]></category>
            <category><![CDATA[ios]]></category>
            <category><![CDATA[android]]></category>
            <category><![CDATA[mobile-app-development]]></category>
            <dc:creator><![CDATA[Apparao Mulpuri]]></dc:creator>
            <pubDate>Thu, 11 Jan 2024 09:22:12 GMT</pubDate>
            <atom:updated>2024-01-11T09:22:12.553Z</atom:updated>
            <content:encoded><![CDATA[<p>Navigating the iOS App Store’s application process can be complex. Each status your app encounters represents a specific stage in its journey from development to being available to users. Understanding these statuses is crucial for developers to manage their app effectively.</p><ol><li>readyForSale: The app is approved and available on the App Store.</li><li>prepareForUpload: The app is registered but not yet uploaded.</li><li>waitingForUpload: The app is ready for binary file upload.</li><li>uploadReceived: The uploaded app binary is received by the App Store.</li><li>waitingForExportCompliance: Awaiting confirmation of export laws compliance.</li><li>waitingForReview: The app is in the queue for Apple’s review.</li><li>pendingContract: Waiting for financial or legal agreement completion.</li><li>parking: Status not commonly documented or used.</li><li>processing: The app is being processed after upload.</li><li>inReview: The app is currently under review by Apple.</li><li>pendingDeveloperRelease: Approved and waiting for developer to release.</li><li>rejected: The app is not approved by Apple.</li><li>metadataRejected: Only metadata issue, no need to resubmit binary.</li><li>removedFromSale: The app is removed from sale on the App Store.</li><li>devRejected: Developer has rejected the binary after upload.</li><li>developerRemovedFromSale: Removed from sale by the developer.</li><li>invalidBinary: Uploaded binary is not valid.</li><li>missingScreenshot: Screenshots required for the app are missing.</li><li>deleted: The app record is deleted.</li></ol><p>Each status is a critical part of the app’s lifecycle on the App Store. Developers should regularly check their app’s status and understand what action, if any, is required from their end to move to the next stage. This awareness ensures a smooth transition through the review process and helps in planning and updating app release strategies.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=7ef1c6dc01c4" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Architecture changes in Microsoft New Teams app]]></title>
            <link>https://medium.com/@apparaomulpuri/architecture-changes-in-microsoft-new-teams-app-c9265150e7e9?source=rss-f772bca567f6------2</link>
            <guid isPermaLink="false">https://medium.com/p/c9265150e7e9</guid>
            <category><![CDATA[architecture]]></category>
            <category><![CDATA[team]]></category>
            <category><![CDATA[jiomeet]]></category>
            <category><![CDATA[microsoft]]></category>
            <dc:creator><![CDATA[Apparao Mulpuri]]></dc:creator>
            <pubDate>Thu, 11 Jan 2024 09:15:13 GMT</pubDate>
            <atom:updated>2024-01-11T09:15:13.987Z</atom:updated>
            <content:encoded><![CDATA[<p>This new approach is a significant shift from the original architecture of Teams launched in 2017, focusing on bringing communication and collaboration tools into one interface​​.</p><p>The new Microsoft Teams architecture has made several significant changes to its tech stack:</p><ol><li><strong>From Electron to Edge WebView2</strong>: Teams has moved away from the Electron framework to Edge WebView2, which allows web technologies to be embedded within native apps and is expected to use half the memory previously required​​.</li><li><strong>Adoption of ReactJS</strong>: Angular is no longer used; instead, Teams is now built entirely on ReactJS, which is a JavaScript library for building user interfaces​​.</li><li><strong>Leveraging Apollo GraphQL</strong>: Apollo GraphQL is being used in Teams, which is a data querying and manipulation language for APIs, and a runtime for fulfilling those queries with your existing data​​.</li><li><strong>AI Integration</strong>: The infrastructure has been laid out to support advanced AI features, including meeting recaps and AI-powered assistance with Copilot for Teams, signaling a move towards a more intelligent and context-aware application</li></ol><p>These changes represent a shift towards a more performance-optimized and modern web technology stack, aiming to improve the overall efficiency and user experience of Microsoft Teams.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c9265150e7e9" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Speaker Diarization in Python: A Step-by-Step Guide]]></title>
            <link>https://medium.com/@apparaomulpuri/speaker-diarization-in-python-a-step-by-step-guide-351a094237f2?source=rss-f772bca567f6------2</link>
            <guid isPermaLink="false">https://medium.com/p/351a094237f2</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[audio-processing]]></category>
            <category><![CDATA[speaker-diarization]]></category>
            <dc:creator><![CDATA[Apparao Mulpuri]]></dc:creator>
            <pubDate>Tue, 28 Nov 2023 04:55:59 GMT</pubDate>
            <atom:updated>2023-11-28T04:57:26.992Z</atom:updated>
            <content:encoded><![CDATA[<h3>Introduction</h3><p>In the era of burgeoning audio and video content, speaker diarization — the task of partitioning audio streams into homogeneous segments according to the speaker identity — is becoming increasingly significant. Whether it’s for transcribing interviews, analyzing meetings, or enhancing voice recognition systems, the ability to accurately identify who is speaking when is invaluable. In this article, I’ll explore the basics of speaker diarization and walk through a Python implementation to get you started on this fascinating journey.</p><h3>What is Speaker Diarization?</h3><p>Speaker diarization is a process in audio processing that involves identifying and segmenting speech by the speaker. It answers the question, “Who spoke when?” This is particularly useful in scenarios where multiple speakers are involved, such as meetings, court proceedings, or phone calls.</p><h3>The Process of Speaker Diarization</h3><p>The typical workflow for speaker diarization involves several steps:</p><ol><li>Voice Activity Detection (VAD): This step identifies whether a segment of audio contains speech or not.</li><li>Feature Extraction: Here, I extracted meaningful features from the audio data, which can be used to differentiate speakers.</li><li>Speaker Clustering: In this step, similar speech features are grouped together, usually by employing a clustering algorithm. Each cluster ideally represents a unique speaker.</li></ol><h3>Python Implementation</h3><p>Let’s dive into a simple Python implementation. I’ll use librosa for audio processing and sklearn for clustering.</p><h3>Setting Up Your Environment</h3><p>First, ensure you have the necessary libraries installed:</p><pre>pip install librosa scikit-learn</pre><h3>Code</h3><p>Here’s a basic script for speaker diarization:</p><pre>import librosa<br>from sklearn.cluster import KMeans<br>from sklearn.preprocessing import StandardScaler</pre><pre>audio_path = &#39;path_to_your_audio_file.wav&#39;<br>audio, sr = librosa.load(audio_path, sr=None)</pre><pre>mfccs = librosa.feature.mfcc(audio, sr=sr)</pre><pre><br>scaler = StandardScaler()<br>mfccs_scaled = scaler.fit_transform(mfccs.T)</pre><pre>kmeans = KMeans(n_clusters=2)  # Adjust based on the expected number of speakers<br>speaker_labels = kmeans.fit_predict(mfccs_scaled)</pre><pre><br>for i, label in enumerate(speaker_labels):<br>    print(f&quot;Time Segment {i}: Speaker {label}&quot;)</pre><p>This script loads an audio file, extracts MFCC features, scales them, and then uses KMeans clustering to identify different speakers.</p><h3>Understanding the Code</h3><ul><li>MFCCs: These are coefficients that collectively make up an MFC. They are derived from the Fourier transform of a log power spectrum of a sound.</li><li>StandardScaler: This is used to standardize features by removing the mean and scaling to unit variance, which is a common requirement for many machine learning estimators.</li><li>KMeans Clustering: An unsupervised machine learning algorithm that groups data into k number of clusters.</li></ul><h3>Challenges and Considerations</h3><p>Speaker diarization comes with its challenges, such as dealing with overlapping speech, varying audio quality, and differentiating speakers with similar voice characteristics. Accurately determining the number of speakers (n_clusters in KMeans) is also crucial and can be complex in real-world scenarios.</p><h3>Conclusion</h3><p>While this guide provides a basic introduction to speaker diarization, the field is vast and continuously evolving. Advanced techniques involving deep learning and neural networks are being actively researched and developed. As audio content continues to grow in volume and importance, the role of speaker diarization will only become more prominent.</p><p>I encourage you to experiment with this script, adjust it to different audio samples, and explore the expansive world of audio processing and machine learning. Happy coding!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=351a094237f2" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Harnessing Diverse Computing Environments: From CPU to TPU]]></title>
            <link>https://medium.com/@apparaomulpuri/harnessing-diverse-computing-environments-from-cpu-to-tpu-0ea3d28f82b8?source=rss-f772bca567f6------2</link>
            <guid isPermaLink="false">https://medium.com/p/0ea3d28f82b8</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[processing]]></category>
            <category><![CDATA[environment]]></category>
            <category><![CDATA[nvidia]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Apparao Mulpuri]]></dc:creator>
            <pubDate>Mon, 27 Nov 2023 01:36:25 GMT</pubDate>
            <atom:updated>2023-11-27T01:36:25.300Z</atom:updated>
            <content:encoded><![CDATA[<p>In the rapidly evolving world of computing, the efficiency and effectiveness of data processing, machine learning, and deep learning tasks are heavily influenced by the choice of hardware. Different environments, from traditional CPUs to more advanced TPUs, offer unique advantages. This blog will explore the distinct features and applications of various environments including CPUs, and NVIDIA’s A100, V100, and T4 GPUs, as well as Google’s TPUs.</p><h3>CPU: The Versatile Workhorse</h3><p>CPUs (Central Processing Units) are the general-purpose processors in most computers. They are incredibly versatile, capable of handling a wide range of tasks from basic computing to complex algorithms. CPUs are excellent for tasks that require sequential processing and are an ideal choice for applications that don’t demand intense parallel computing power. They are the backbone of general computing, suitable for a broad array of applications, from running operating systems to web browsing and traditional application workloads.</p><h3>NVIDIA A100 GPU: The AI Powerhouse</h3><p>The NVIDIA A100 GPU, based on the NVIDIA Ampere architecture, is designed specifically for accelerating AI and high-performance computing (HPC). It provides a massive leap in performance and capabilities compared to its predecessors. The A100 GPU excels in parallel processing, making it perfect for deep learning and large-scale AI workloads. It offers substantial improvements in throughput, energy efficiency, and the ability to handle diverse compute workloads simultaneously.</p><h3>NVIDIA V100 GPU: The Versatile Accelerator</h3><p>The NVIDIA V100 GPU is another powerful tool in the realm of AI and machine learning. Based on the previous generation Volta architecture, the V100 provides excellent performance for both deep learning training and inference. It’s a versatile choice for researchers and data scientists requiring significant computing power for neural network training, AI model development, and HPC tasks.</p><h3>NVIDIA T4 GPU: The Efficient Performer</h3><p>The NVIDIA T4 GPU, powered by the Turing architecture, is a more energy-efficient option. It’s designed to accelerate diverse cloud workloads, including computing, graphics, and AI inference. The T4 is a popular choice for businesses looking for cost-effective performance. It’s well-suited for applications like AI inference at the edge, video analytics, and image processing tasks.</p><h3>TPU: Specialized Hardware for Machine Learning</h3><p>Google’s TPUs (Tensor Processing Units) are custom-developed ASICs (Application-Specific Integrated Circuits) designed to accelerate machine learning workloads. They are particularly effective for training and running TensorFlow models. TPUs provide extremely high throughput for specific tasks, especially large-scale and complex neural network models. They are optimized for matrix operations and data-intensive tasks, which are central to deep learning algorithms.</p><h3>Conclusion: Choosing the Right Tool for the Task</h3><p>Each of these computing environments has its niche, and the choice depends on the specific requirements of the task at hand. CPUs offer general-purpose versatility, A100 and V100 GPUs deliver powerful AI and HPC capabilities, the T4 GPU provides efficient performance for a range of cloud workloads, and TPUs offer specialized acceleration for TensorFlow-based machine learning tasks.</p><p>As technology continues to advance, understanding the strengths and optimal use cases for each of these environments becomes crucial for businesses, researchers, and developers aiming to leverage the power of modern computing to its fullest potential. Whether it’s through a CPU’s flexibility, a GPU’s parallel processing prowess, or a TPU’s machine learning specialization, the right hardware choice can significantly enhance the efficiency and outcomes of computational tasks.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=0ea3d28f82b8" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[List of LLMs available]]></title>
            <link>https://medium.com/@apparaomulpuri/list-of-llms-available-bbad76477de8?source=rss-f772bca567f6------2</link>
            <guid isPermaLink="false">https://medium.com/p/bbad76477de8</guid>
            <category><![CDATA[openai]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Apparao Mulpuri]]></dc:creator>
            <pubDate>Mon, 13 Nov 2023 01:50:27 GMT</pubDate>
            <atom:updated>2023-11-13T01:50:27.525Z</atom:updated>
            <content:encoded><![CDATA[<h3>As of 2023, several significant Large Language Models (LLMs) are available, each with unique features and capabilities:</h3><ol><li><strong>GPT-4:</strong> Developed by OpenAI, GPT-4 is a multimodal model capable of processing both text and images. It’s known for its complex reasoning understanding, advanced coding capability, and improved factuality and alignment with human values​​.</li><li><strong>GPT-3.5:</strong> Also from OpenAI, GPT-3.5 is a general-purpose LLM, praised for its speed and versatility in tasks like essay writing and creative content generation​​.</li><li><strong>PaLM 2 (Bison-001):</strong> Google’s PaLM 2 model excels in commonsense reasoning, formal logic, mathematics, and advanced coding in multiple languages. It has been trained on 540 billion parameters​​.</li><li><strong>Claude v1:</strong> Developed by Anthropic, Claude v1 focuses on creating AI assistants that are helpful, honest, and harmless. It scores well in benchmark tests and offers a large context window of 100k tokens​​.</li><li><strong>Cohere:</strong> Founded by former Google Brain team members, Cohere is designed for enterprise applications and is known for its accuracy and robustness in generative AI use cases​​.</li><li><strong>Falcon:</strong> An open-source LLM developed by the Technology Innovation Institute (TII), UAE, Falcon is notable for being available under the Apache 2.0 license, allowing commercial use​​.</li><li><strong>LLaMA:</strong> Released by Meta, LLaMA models range from 7 billion to 65 billion parameters and are designed for research purposes​​.</li><li><strong>Guanaco-65B:</strong> An open-source LLM derived from LLaMA, Guanaco-65B is fine-tuned on the OASST1 dataset and performs well in various benchmarks​​.</li><li><strong>Vicuna 33B:</strong> Another LLaMA-derived open-source model, Vicuna 33B has been fine-tuned using supervised instruction and performs impressively in benchmark tests​​.</li><li><strong>MPT-30B:</strong> Developed by Mosaic ML, this open-source model is fine-tuned on a large corpus of data from various sources and is suitable for local use​​.</li><li><strong>30B-Lazarus:</strong> Based on LLaMA, the 30B-Lazarus model by CalderaAI uses LoRA-tuned datasets from multiple models and excels in text generation​​.</li><li><strong>WizardLM:</strong> Built to follow complex instructions, WizardLM is fine-tuned on the LLaMA model and is known for its efficient performance in benchmarks​​.</li><li><strong>GPT4All:</strong> A project by Nomic AI, GPT4All allows running local LLMs on computers without dedicated GPUs or internet connectivity. It includes a 13B Snoozy model and houses 12 open-source models from different organizations​​.</li></ol><p>Each of these models has its strengths and is suited for different applications, ranging from general-purpose use and creative tasks to specific enterprise solutions and research-oriented tasks.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=bbad76477de8" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>