<?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 sumitava chakraborty on Medium]]></title>
        <description><![CDATA[Stories by sumitava chakraborty on Medium]]></description>
        <link>https://medium.com/@chakrabortysumitava?source=rss-949f6e266da2------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/0*Q3ZnE8xmSL36J8Fk</url>
            <title>Stories by sumitava chakraborty on Medium</title>
            <link>https://medium.com/@chakrabortysumitava?source=rss-949f6e266da2------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Fri, 20 Nov 2026 14:10:43 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@chakrabortysumitava/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[Hindi, Arabic and Thai Captions: Why AI Clippers Get Them Wrong]]></title>
            <link>https://medium.com/@chakrabortysumitava/hindi-arabic-and-thai-captions-why-ai-clippers-get-them-wrong-ee2924b8509f?source=rss-949f6e266da2------2</link>
            <guid isPermaLink="false">https://medium.com/p/ee2924b8509f</guid>
            <category><![CDATA[subtitles]]></category>
            <category><![CDATA[hindi]]></category>
            <category><![CDATA[speech-recognition]]></category>
            <category><![CDATA[thai]]></category>
            <category><![CDATA[localization]]></category>
            <dc:creator><![CDATA[sumitava chakraborty]]></dc:creator>
            <pubDate>Thu, 19 Nov 2026 19:39:59 GMT</pubDate>
            <atom:updated>2026-11-19T19:39:59.470Z</atom:updated>
            <content:encoded><![CDATA[<p>Hindi, Urdu, Arabic, Tamil and Thai break auto captions in six predictable ways. Here is what native-script captions need, and what we changed</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*F93TiRSRMXaeO5MKIVppgA.png" /><figcaption>“This is the biggest mistake” in four scripts. Two of them run right to left</figcaption></figure><p>Run a Hindi podcast through most AI clippers and one of three things comes back: captions in English, captions in Hindi spelled with Latin letters when you never asked for that, or captions in the right script drawn as empty boxes. For a creator in Delhi, Lahore, Cairo or Chennai, each of those means an unusable clip.</p><p>These failures are not random. They come from six specific places in the pipeline, and each has a known fix. I build the clipping pipeline at shitShorts, and this post walks through what we saw in production in November 2026, what we changed, and what still goes wrong. The checks at the end work on any tool.</p><h3>Why the transcript decides everything</h3><p>A clipper does almost all of its work on the transcript. It reads every word to find moments with a hook and a payoff, cuts on word boundaries so no clip starts mid-word, and builds the captions from the same word timings. If the transcript is in the wrong language, or in the right language but the wrong script, every later step inherits the mistake. The moments are picked from text that does not match the audio, and the captions drift away from what people hear.</p><p>So most of the fixes below are about the transcript, not about the captions. Here is the whole path a non-English video takes, in the order the steps run; the sections below go through it one failure at a time.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*cg8WDXET7mu1qh1CLTFI1A.png" /><figcaption>The caption pipeline for a non-English video. A transcript that comes back in the wrong script is sent back with the language forced</figcaption></figure><h3>Failure 1: the first ten seconds decide the language</h3><p>Speech recognisers such as Whisper detect the language from the opening audio. A long video often opens with an English intro, a song or a sponsor read. Judge the language from that, and a Hindi podcast with a 20-second English intro comes back with English captions for the whole hour.</p><p>The fix is to vote. We take three short samples spread across the video, near the start, the middle and the end, and go with the majority. A sample that is music or silence does not get a vote: a detector given music still returns a confident-looking guess, and that guess is noise.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*7bp6gRrKe--U3bIF-p_slw.gif" /><figcaption>Three samples spread across the video. The one that lands in the English intro gets one vote and loses to two Hindi samples; deciding from the opening alone would have captioned the whole video in English</figcaption></figure><p>Code-mixed speech needs one more rule. When English wins the vote but the odd sample is a language usually mixed with English, such as Hindi, Urdu, Bengali, Tamil, Telugu or Marathi, we force nothing. Each stretch of the transcript then detects its own language. That keeps Hinglish readable: Hindi stretches stay in Devanagari and English stretches stay in English. A tie is handled the same way.</p><h3>Failure 2: the model translates instead of transcribing</h3><p>This is the one that surprised us. Asked to transcribe Hindi speech, a fast multilingual model will sometimes answer with a fluent English translation. Nothing errors. The output looks clean. It is simply not what was said, and the word timings no longer line up with the audio.</p><p>We now check the script of every chunk that comes back. For each language with its own script, we look at how much of the text is actually written in that script: Devanagari for Hindi and Marathi, Arabic letters for Arabic, Urdu and Persian, Tamil letters for Tamil, and so on. A chunk that comes back mostly in Latin letters is treated as a wrong-script answer and transcribed again on a larger model, with the language forced, and we keep whichever answer has more of the right script. Code-mixed speech still carries plenty of its own script, so Hinglish passes the check.</p><p>The redo is guarded so it cannot fail a run: if the larger model has a problem, the first answer stands.</p><h3>Failure 3: the fast model is worse on Indian languages</h3><p>The wrong-script check catches the worst answers. It does not catch a transcript in the right script that is simply full of mistakes. So we compared the fast speech model with a larger one on a public speech benchmark, language by language, counting character errors:</p><p>Tamil had about half the errors with the larger model (53% fewer). Hindi had 16% fewer and Bengali 12% fewer. The other languages we tested came out about the same on both.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*gu4J73ejapnhu0D7ZT2T_w.png" /><figcaption>Character errors with the larger model, with the fast model set to 100, measured November 2026. Shorter is better</figcaption></figure><p>On Tamil the fast model’s captions were not usable, and halving the errors is the difference between a caption people can read and one they scroll past. So languages where the larger model clearly wins are transcribed on it, and the rest stay on the faster one, where the two scored the same.</p><h3>Failure 4: Urdu and Hindi sound the same</h3><p>Spoken Urdu and Hindi are very close. A detector that only hears the audio picks between them almost at random, and the result decides the script: Urdu is written right to left in Arabic letters, Hindi left to right in Devanagari. Bengali, Gujarati and Marathi are also often heard as Hindi.</p><p>For YouTube links, the title and description help. If they are written in Arabic script, the video is very likely Urdu; in Devanagari, Hindi; in Bengali script, Bengali. We use that hint only to settle a vote among the Indian-family languages, so an Arabic or English video is never moved by its title. For uploads with no title, picking the language yourself is the reliable path. Picking a language forces it for the whole video.</p><h3>Failure 5: the captions cannot draw the script</h3><p>Now the transcript is right, and the captions can still fail. Caption themes use display fonts, and most display fonts cover Latin letters only. Hand them Devanagari or Tamil and you get empty boxes, often called tofu.</p><p>The fix is a font fallback per script. Our captions fall back to Noto fonts made for each script: Devanagari, Bengali, Tamil, Telugu, Kannada, Malayalam, Gujarati, Gurmukhi, Arabic, Hebrew, Thai, and Chinese, Japanese and Korean. Accented Latin letters and Cyrillic stay in the theme’s own font when it has them.</p><p>Right-to-left scripts need three more things:</p><p><strong>Direction per line.</strong> A caption line runs right to left only when most of its letters are in a right-to-left script. An Arabic line with one English brand name still reads right to left; an English line with one Arabic name stays left to right.</p><p><strong>The highlight moves the right way.</strong> Animated captions highlight each word as it is spoken. In Arabic, Urdu, Persian and Hebrew the highlight has to sweep right to left, or the animation contradicts the reading order.</p><p><strong>No letter spacing on joined scripts.</strong> Many caption styles add tracking for a bold look. In Arabic that pulls letters apart that should connect, so joined scripts get none.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ShRATgAh1tlR2QNBNd3Ttg.gif" /><figcaption>Word-by-word highlight in Arabic (right to left) and Hindi (left to right). Arabic letters stay joined</figcaption></figure><p>One honest limit: Urdu captions are drawn in a Naskh-style Noto face, not Nastaliq. Naskh stays readable at caption size on a phone, but Urdu readers will notice, and we have not found a Nastaliq face that holds up at that size yet.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Sq3iLMtfMRSmETOqi-1M2A.png" /><figcaption>Script, reading direction and caption font for eight of the 40 languages, with a sample caption in each</figcaption></figure><h3>Failure 6: Thai has no spaces between words</h3><p>Thai, like Lao, Khmer and Burmese, is written without spaces between words. A Thai creator told us our captions looked broken, and they were right. Three things went wrong at once.</p><p>The speech model returns Thai as syllables with a space after each one, so one word such as โทรศัพท์ (phone) arrived as two, and the caption read like a stammer. The word highlight then split text by characters, and in Thai a vowel or tone mark is its own character that sits above or below a letter: a highlight that started on the mark left it floating with no letter under it. And because every syllable looked like a word, lines broke in the middle of words.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*1icR4LmQDMaUWvUG8KVKow.png" /><figcaption>The same Thai caption before and after. Before: syllables split by spaces and a highlight that cuts a letter in half. After: whole words, and the highlight covers a full word</figcaption></figure><p>The fix has three parts. We join the syllables back into words with a Thai word splitter, so only phrases are separated by spaces, which is how Thai is written. The highlight now works on whole letters, so a vowel or tone mark always moves with its letter. And Thai captions get a Thai font, more line height for the marks above and below, no letter spacing and no capitals, with lines breaking only between words. The titles, hooks and covers come back in Thai too, instead of English letters.</p><h3>Native script, English letters or a translation</h3><p>Getting the script right is only half of it, because not every creator wants it. Many Hindi creators post Reels in Hinglish written with English letters, because that is how their viewers text. So each run now has a caption choice: the video’s own script, English letters, or an English translation.</p><p>The default depends on the language. Thai and Arabic open in their own script, because almost nobody reads them in Latin letters. Hindi and the other scripts open in English letters, as before. The creator can switch before the run and again for each clip.</p><h3>What still goes wrong</h3><p>None of this makes multilingual captions perfect. English words inside a Hindi sentence can come out in either script. Strong Arabic dialect, slang and names may need a few words fixed by hand. Translation goes into English only. And automatic detection can still choose Hindi for an Urdu video without a title, which is why the language picker exists.</p><h3>A five-minute test for any clipper</h3><p>You can check any tool, including ours, with one real video before paying:</p><ol><li>Use a video with an English intro and at least one stretch of mixed speech.</li><li>Check that the captions come out in the script you asked for, word by word. Romanised Hindi you did not ask for is a fail.</li><li>Look for empty boxes in matras, conjuncts and Tamil vowel signs, and for Thai tone marks cut off from their letters.</li><li>For Arabic or Urdu, check that the line starts on the right, that the word highlight sweeps right to left, and that letters are joined.</li><li>Check that the hook line and post text come back in the same language as the video.</li></ol><p>If you post in one of these languages, the per-language pages for <a href="https://shitshorts.com/hindi-video-clipper?utm_source=medium&amp;utm_medium=article&amp;utm_campaign=native-script-captions">Hindi</a>, <a href="https://shitshorts.com/urdu-video-clipper?utm_source=medium&amp;utm_medium=article&amp;utm_campaign=native-script-captions">Urdu</a>, <a href="https://shitshorts.com/arabic-video-clipper?utm_source=medium&amp;utm_medium=article&amp;utm_campaign=native-script-captions">Arabic</a> and <a href="https://shitshorts.com/tamil-video-clipper?utm_source=medium&amp;utm_medium=article&amp;utm_campaign=native-script-captions">Tamil</a> show sample captions in each script. Every language costs the same on shitShorts, about 20 credits a clip.</p><p><em>I work on shitShorts. This post was drafted with AI assistance and checked against our production pipeline in November 2026.</em></p><p>Test it with your own Hindi, Urdu, Arabic or Tamil video: <a href="https://shitshorts.com/multilingual-video-clipper?utm_source=medium&amp;utm_medium=article&amp;utm_campaign=native-script-captions">shitShorts captions 40 languages in their own scripts, free for about 15 clips a month</a>.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=ee2924b8509f" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[OpusClip vs Klap vs Vizard vs Submagic: 2026 Comparison]]></title>
            <link>https://medium.com/@chakrabortysumitava/opusclip-vs-klap-vs-vizard-vs-submagic-2026-comparison-026d213ef18e?source=rss-949f6e266da2------2</link>
            <guid isPermaLink="false">https://medium.com/p/026d213ef18e</guid>
            <category><![CDATA[video-editing]]></category>
            <category><![CDATA[content-creation]]></category>
            <category><![CDATA[youtube]]></category>
            <category><![CDATA[creator-economy]]></category>
            <category><![CDATA[ai-video]]></category>
            <dc:creator><![CDATA[sumitava chakraborty]]></dc:creator>
            <pubDate>Tue, 17 Nov 2026 13:54:59 GMT</pubDate>
            <atom:updated>2026-11-17T13:54:59.632Z</atom:updated>
            <content:encoded><![CDATA[<p>Price per clip, free plans, long-video limits and export quality, read from each pricing page as of November 2026</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*e_bh256o0Yw63B29Sbwwng.png" /><figcaption>Each tool counts usage in a different unit, which is why headline prices are hard to compare</figcaption></figure><p>Most comparisons of AI clippers rank tools by a feature list. That tells you little, because the four best-known tools in this space do not even sell the same thing. OpusClip sells credits, Vizard sells minutes of uploaded video, Klap sells finished clips, and Submagic sells short videos with a length cap. Before you can say which is cheaper, you have to convert all four to the same unit.</p><p>This post does that conversion. Every number below comes from the vendors’ own pricing pages as of November 2026, and where a page does not state something, I say so instead of guessing. Plans change often, so treat this as a snapshot and check the pricing page before you pay. I also add a fifth tool at the end, which we make, with the same rules applied to it.</p><h3>The short answer</h3><p><strong>OpusClip</strong> is the safe first test: a monthly free plan, a $15 entry plan and the widest set of extras on Pro.</p><p><strong>Klap</strong> is the easiest to budget if you pay yearly, because it counts clips, not minutes.</p><p><strong>Vizard</strong> gives the most source minutes for the money on its Creator plan and lists 4K export there.</p><p><strong>Submagic</strong> is a caption tool for videos you have already cut. Its plans cap the length of each input video, so it is the wrong pick for clipping an hour-long podcast.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*FS36xbRYvjYmrj_SBqNf0A.png" /><figcaption>Free plan, entry price, yearly price, unit and longest input per tool. Source: each vendor’s pricing page, as of November 2026</figcaption></figure><h3>How each tool counts usage</h3><p><strong>OpusClip</strong> gives 60 credits a month on its free plan. Starter is $15 a month for 150 credits and is billed monthly only. Pro is $29 a month, or $14.50 a month billed yearly, for 300 credits per pack. The pricing page does not say how many minutes of video a credit covers. Many reviews assume one credit per minute, but I could not confirm that on the page itself, so the per-hour figures for OpusClip below carry that assumption.</p><p><strong>Klap</strong> has no free plan. It offers a trial on one video without a card, and you can preview and edit those clips, but downloading or posting needs a paid plan. Basic is $29 a month, or $14.50 billed yearly, for 100 clips a month. Pro is $79 for 300 clips and Pro+ $189 for 1,000. Klap bills per generated clip, accepts source videos up to 3 hours, and gives yearly customers the full year of clips up front.</p><p><strong>Vizard</strong> has a free plan of 60 credits a month, where one credit is one minute of uploaded video. Creator is $29 a month, or $14.50 billed yearly, for 600 credits a month on the base tier. Business is $39, or $19.50 billed yearly, and adds a shared workspace and a brand kit.</p><p><strong>Submagic</strong> has no free plan described on its pricing page, only a free trial. Starter is $19 a month, or $12 billed yearly, for 15 videos of up to 2 minutes each. Pro is $39 for 40 videos up to 5 minutes, and Business is $69 for 100 videos up to 30 minutes. Submagic does have a long-to-short feature, but its pricing page does not say what it costs per plan.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*DURUKOBrh1t9WemRPUL1XQ.gif" /><figcaption>The same job, one 60-minute episode cut into 10 clips, on each tool’s meter. OpusClip assumes 1 credit per minute, which its pricing page does not confirm</figcaption></figure><h3>Price per clip</h3><p>To compare fairly, take one realistic job: a 60-minute episode cut into 10 clips, four episodes a month. That is 240 minutes of source video and 40 clips.</p><p><strong>Klap Basic</strong> covers it at $29 a month: 40 of its 100 clips. Per clip that is about 73 cents at this volume, or 29 cents if you use all 100. Billed yearly, the full-use figure drops to about 15 cents.</p><p><strong>Vizard Creator</strong> covers it at $29 a month: 240 of its 600 minutes. If you clip more, the unit cost falls, because Vizard does not count clips at all. At full use, 600 minutes for $29 is about $2.90 per hour of source.</p><p><strong>OpusClip</strong> needs Pro at $29 for 240 minutes, because Starter’s 150 credits would run out after two and a half episodes, assuming one credit per minute.</p><p><strong>Submagic</strong> cannot take the job as described. Even Business caps each input at 30 minutes, so you would have to split every episode first, and the long-to-short price is not listed.</p><p>The pattern: per-clip pricing (Klap) is cheapest when you post a few clips from many videos; per-minute pricing (Vizard, OpusClip) is cheapest when you pull many clips from each long video. Work out which kind of creator you are before comparing headline prices.</p><h3>Long videos and streams</h3><p>Long input is where these tools differ most, and it matters if you clip podcasts or live streams.</p><p>Klap accepts source videos up to 3 hours.</p><p>Vizard’s free plan takes uploads up to 60 minutes; Creator raises that to 600 minutes or 30 GB.</p><p>OpusClip’s free plan accepts local uploads up to 10 GB; its page does not state a length limit.</p><p>Submagic caps input length per plan at 2, 5 and 30 minutes.</p><p>None of the four pricing pages say how the vertical frame handles a stream layout, where your face sits in a small box over the game. Test that on your own VOD. A plain centre crop of a gaming stream shows a slice of the game and none of you.</p><h3>Languages</h3><p>This is the part I could not compare from the pricing pages: none of the four list which languages they caption, or in which script. If you post in Hindi, Arabic, Urdu or Tamil, test with a real video and check two things before paying. First, whether the captions come back in the language’s own script rather than in Latin letters or as an English translation. Second, whether the caption font draws the script at all; a theme that only covers Latin letters exports empty boxes.</p><h3>Export quality</h3><p><strong>OpusClip:</strong> free clips render up to 1080p with a watermark and can be exported for 3 days. Business lists 4K.</p><p><strong>Klap:</strong> export resolution is not stated on the pricing page; check it during the trial.</p><p><strong>Vizard:</strong> free exports are 720p; Creator lists 4K and no watermark.</p><p><strong>Submagic:</strong> Starter is 1080p at 30 fps, Business 4K at 60 fps, no watermark on paid plans.</p><h3>What each tool does well</h3><p>OpusClip has the broadest paid plan: a scheduler, AI B-roll, 9:16, 1:1 and 16:9 output, and export to Premiere and DaVinci. If you finish clips in a desktop editor, that export alone can decide it. Klap is the most predictable bill, and its yearly plans are half the monthly price. Vizard gives the most source minutes per dollar and the cheapest team plan of the four, with extra seats at $5 a month. Submagic is priced for styled captions on short edits, shows prices in euros and Brazilian reais, and is a good second tool if you already cut your own clips.</p><h3>Where a fifth option fits</h3><p>We make <a href="https://shitshorts.com/?utm_source=medium&amp;utm_medium=article&amp;utm_campaign=opusclip-klap-vizard-submagic-2026">shitShorts</a>, so weigh this section accordingly. It counts credits, where one credit is one cent and a clip costs 20 credits plus 10 for each run’s transcript. The free plan gives 300 credits when you sign up and again every 30 days, about 15 clips, at 720p with a small watermark; the first 2 clip exports each month are free, then 20 credits each. Paid plans start at $10 a month for about 50 clips at 1080p without a watermark, and there is no yearly plan.</p><p>It was built for the cases the table above cannot show: YouTube links up to 8 hours, stream layouts with the facecam stacked over the gameplay, and captions in 40 languages, each in its own script. File uploads are limited to 50 MB, so long recordings go in as a YouTube link. The weak spots are real too: no yearly discount, a small upload limit, and far fewer published reviews than OpusClip.</p><h3>How to choose in one afternoon</h3><p>Pick one video that looks like your channel, ideally long and with more than one person on screen. Run it through the free plan or trial of two or three tools on the same day. Count the clips you would post without changes, then try one fix in each: a misheard word, a trimmed ending, a moved crop. The tool that gives the most postable clips for the least fixing wins, whatever this comparison says.</p><p>For a longer list with every plan we tracked, see the <a href="https://shitshorts.com/guides/ai-clipper-pricing?utm_source=medium&amp;utm_medium=article&amp;utm_campaign=opusclip-klap-vizard-submagic-2026">AI clipper pricing guide</a> and the head-to-head pages for <a href="https://shitshorts.com/guides/opusclip-vs-klap?utm_source=medium&amp;utm_medium=article&amp;utm_campaign=opusclip-klap-vizard-submagic-2026">OpusClip vs Klap</a> and <a href="https://shitshorts.com/guides/opusclip-vs-vizard?utm_source=medium&amp;utm_medium=article&amp;utm_campaign=opusclip-klap-vizard-submagic-2026">OpusClip vs Vizard</a>.</p><p><em>All prices from the vendors’ public pricing pages as of November 2026. All trademarks belong to their owners. I work on shitShorts. Drafted with AI assistance; every number was checked against its source.</em></p><p>Try the same test on your own footage: <a href="https://shitshorts.com/?utm_source=medium&amp;utm_medium=article&amp;utm_campaign=opusclip-klap-vizard-submagic-2026">shitShorts is free for about 15 clips a month, no card</a>.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=026d213ef18e" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>