{ "service": "ClipQuill", "description": "Public, read-only facts about ClipQuill, taken verbatim from https://clipquill.com/. No authentication, no write operations, no arbitrary-URL fetching.", "source_url": "https://clipquill.com/", "docs_url": "https://clipquill.com/ai/", "updated": "2026-11-01", "count": 19, "facts": [ { "id": "purpose", "topic": "What this site is", "keywords": [ "what", "purpose", "about", "clipquill", "tool", "does", "converter" ], "answer": "ClipQuill transcribes video or audio to text. It runs entirely in your own browser tab: the recognition model is downloaded once, then all recognition happens on your device. It is a video-to-text converter, not an upload service.", "source_url": "https://clipquill.com/", "updated": "2026-10-03" }, { "id": "privacy", "topic": "Upload and privacy", "keywords": [ "privacy", "upload", "uploaded", "server", "stored", "kept", "data", "safe", "private" ], "answer": "The file is processed in your browser and zero bytes are uploaded. It is never stored and nothing is sent to this site. It disappears when you close the tab. There is no copy on a server to ask us to delete and no retention period.", "source_url": "https://clipquill.com/", "updated": "2026-10-03" }, { "id": "accounts", "topic": "Accounts, signup and cost", "keywords": [ "account", "signup", "sign up", "login", "free", "cost", "price", "paywall", "quota", "cap", "credits" ], "answer": "None. No signup, no login, no daily quota, no credits, no paywall. The model is downloaded to your browser once and every transcript after that costs nothing on our side, because your own device is doing the work.", "source_url": "https://clipquill.com/", "updated": "2026-10-03" }, { "id": "input-formats", "topic": "Supported input formats", "keywords": [ "input", "format", "formats", "mp4", "mov", "webm", "m4a", "mp3", "wav", "aac", "ogg", "flac", "opus", "3gp", "m4b", "amr", "wma", "file type", "paste" ], "answer": "MP4, MOV, WEBM, M4A, MP3, WAV, AAC, OGG, FLAC, OPUS, 3GP and M4B. That list is what has been tested; a file outside it may still work, but it is not promised. You can also paste a file straight in with Ctrl+V.", "source_url": "https://clipquill.com/", "updated": "2026-10-03" }, { "id": "output-formats", "topic": "Supported output formats", "keywords": [ "output", "export", "txt", "srt", "vtt", "subtitle", "subtitles", "captions", "copy", "download", "transcript" ], "answer": "TXT, SRT and VTT, plus one-click copy. Export as SRT or VTT and you have a subtitle file ready to drop into a video editor. A transcript (TXT) and a subtitle file (SRT/VTT) come from the same run.", "source_url": "https://clipquill.com/", "updated": "2026-10-03" }, { "id": "srt-shape", "topic": "What an exported SRT actually contains", "keywords": [ "srt", "vtt", "subtitle file", "cues", "cue", "line length", "reading speed", "gap", "wav", "flac", "mp4", "mp3", "convert", "converter", "byte", "identical", "how many cues" ], "answer": "An exported SRT is built from the recogniser's own cue list, so it carries the shape of the speech and not of anything you set. Measured on 2026-11-01 on one 38.811-second recording with 89 reference words, built eight ways: the 48 kHz WAV master (3,725,838 bytes) and a FLAC made from the same samples (955,272 bytes) produced identical TXT, SRT and VTT byte for byte, and four MP4s that differ only in the captions already inside them produced one identical 902-byte SRT. All eight builds gave 13 cues and 16 text lines, with three of the cues wrapping to two lines; every cue boundary sits at 0 ms of gap, and no line exceeds the 42 display-width-unit budget, the widest being 38 to 40. The wrong words moved from 15 of 89 to 25 of 89 across the builds while the cue count did not move at all. The tool transcribes the audio stream: it does not read a subtitle track that is already in the file and it does not read text burned into the picture.", "source_url": "https://clipquill.com/audio-to-srt/", "updated": "2026-11-01" }, { "id": "limits", "topic": "Size and length limits", "keywords": [ "limit", "limits", "max", "maximum", "long", "length", "minutes", "60", "512", "mb", "hours", "size" ], "answer": "30 minutes of audio and 512 MiB (536,870,912 bytes) per file. These are the figures for the build that ships with this page: whisper-base. A file past either one is turned away before it starts, rather than run until the browser runs out of memory and the tab dies. Clips of one to three minutes are where it feels smoothest.", "source_url": "https://clipquill.com/", "updated": "2026-10-03" }, { "id": "links", "topic": "Fetching a video from a URL", "keywords": [ "url", "link", "paste a link", "youtube", "instagram", "reel", "tiktok", "facebook", "download from link", "fetch" ], "answer": "No. This page reads a file that is already on your device and it never sends anything anywhere, which also means it cannot fetch a video from a link for you. Save the video first, then upload it. Where the file came from does not matter, only that you have it.", "source_url": "https://clipquill.com/", "updated": "2026-10-03" }, { "id": "language-selection", "topic": "Choosing the spoken language", "keywords": [ "language", "languages", "detect", "auto", "pick", "choose", "select", "english", "chinese", "spanish" ], "answer": "You have to pick the language that is spoken — this build of the model cannot work it out on its own. Pick the wrong one and the transcript comes back in the language you picked: choose English for a Chinese clip and you get English, not Chinese. The choice is remembered on your device.", "source_url": "https://clipquill.com/", "updated": "2026-10-03" }, { "id": "language-count", "topic": "How many languages", "keywords": [ "how many", "languages", "99", "tokens", "multilingual", "supported languages", "list" ], "answer": "The bundled model is multilingual and carries 99 language tokens, so files in those languages can be transcribed. Of those 99, English audio and one short Chinese clip have been run by us — that is the whole of what has been tested. English has a published word error rate. The Chinese run was one 23-second clip with 97 characters of reference; it came back in Traditional characters and, once converted to Simplified, 6 of those 97 characters were wrong. That is a count, not a rate.", "source_url": "https://clipquill.com/", "updated": "2026-10-03" }, { "id": "character-variants", "topic": "Traditional and Simplified characters, and what the model can write", "keywords": [ "simplified", "traditional", "chinese", "character", "characters", "script", "hanzi", "kanji", "japanese", "korean", "hangul", "full-width", "vocabulary", "variant", "cjk" ], "answer": "Measured 2026-09-24 on the saved raw output of a 2026-09-18 run: one 23.088 s Mandarin clip, 97 reference characters, language forced to zh, whisper-base quantised ONNX, on a machine with 2 logical CPUs and 3.94 GB of RAM, one run that was never repeated, so no jitter range. A raw character-for-character comparison counted 42 characters wrong, 43.3%. Classified one character at a time: 35 were the same word written in a different script, from 27 distinct script pairs; 1 was a number written as the ASCII digit 3 where the reference spelled it out; 6 were genuine substitutions. Deletions 0, insertions 0. After converting the script: 7 of 97 wrong, 7.2%. After normalising numerals as well: 6 of 97, 6.2%. Scanning the 50,258-entry vocabulary shared by whisper-base and whisper-tiny: 935 CJK ideographs have a token of their own, 360 Hangul syllables do and 566 distinct syllables appear somewhere in the inventory, 136 kana, 12 CJK punctuation marks, and 0 full-width forms. Of the 27 script pairs from this clip, 12 have both forms, 12 have only the Traditional form, and 3 have neither — the model produced all three anyway, because byte-level encoding lets any character be assembled from byte pieces. Japanese: all 7 Japanese-only characters sampled are absent from the inventory. No Japanese or Korean audio has ever been run, so nothing is claimed about either.", "source_url": "https://clipquill.com/video-to-text-character-variants/", "updated": "2026-09-24" }, { "id": "timings", "topic": "How long it takes", "keywords": [ "how long", "time", "slow", "fast", "speed", "seconds", "minutes", "benchmark", "timing", "wait" ], "answer": "Timed 2026-09-18 on clipquill.com itself, over the public internet, in a real window on this machine, whisper-base quantized, cross-origin-isolated, on a machine with 2 logical CPUs (an AMD EPYC 9754 host with two CPUs allocated to it) and 3.9 GiB of RAM. Cold means the browser cache was emptied first, so that run carried the first download of the model and the runtime and then the transcription itself; it does not include reading the file or decoding the audio; cached means the model was already in the cache, so only the transcription happened. 13 s of audio → 44.1 s cold, 23.1 s cached. 60 s → 37.2 s cold, 22.8 s cached. 277 s (4 min 37 s) → 96.3 s cold, 102.2 s cached. 1609 s (26 min 49 s) → 543.4 s (9 min 3 s) cold, 479.3 s (7 min 59 s) cached. Timed on a 92-second loop of mixed material (clean synthetic speech, real human speech, and real speech under background noise) run through whisper-base. These are the seconds the page reports from handing the file over to the finished transcript; wall-clock runs a few seconds longer because it also counts reading and decoding. It depends on your device, not on us.", "source_url": "https://clipquill.com/", "updated": "2026-10-03" }, { "id": "download-size", "topic": "First-visit download size", "keywords": [ "download", "size", "mb", "76", "97", "21", "model", "cache", "offline", "first run", "bandwidth" ], "answer": "A first visit transfers about 78.4 MiB (82.2 MB), counted file by file on clipquill.com on 2026-09-19 with brotli in transit (every file a first visit pulls fetched twice, once with brotli and once uncompressed), and about 97.5 MiB (102.3 MB) once decompressed: 76.0 MiB of model files plus 21.5 MiB of runtime that runs them. The weights are already quantised and barely compress, so 73.3324 MiB of that 78.4 MiB is model weights crossing the wire at full size. Your browser caches all of it, so from the second file on nothing is downloaded again.", "source_url": "https://clipquill.com/", "updated": "2026-10-03" }, { "id": "accuracy", "topic": "Accuracy", "keywords": [ "accuracy", "accurate", "correct", "wrong", "wer", "quality", "percentage", "noisy", "accent" ], "answer": "Measured 2026-10-03 on the models that ship with this page. On typical files — clean or light-noise real speech, the audio most people actually have — the English default (Moonshine base) is 7.1% wrong and the optional small tier (whisper-small) is 8.2% wrong (85 reference words across four clips). We also test deliberately harsh audio: a clip pushed through a telephone band with echo and hiss, and two clips in heavy background noise; there the English default is 18.2% and 21.3%. English, single speaker, one machine — a sample, not a benchmark. Music under speech, several people talking at once, and strong accents have not been measured, so nothing is claimed for them.", "source_url": "https://clipquill.com/", "updated": "2026-10-03" }, { "id": "mobile", "topic": "Phones", "keywords": [ "phone", "mobile", "android", "ios", "iphone", "375", "responsive", "small screen" ], "answer": "On a phone-sized window (375 px wide, same machine) the layout holds — no horizontal overflow, no broken layout. That is a narrow window on a desktop, not a phone. On a real phone — a Samsung Galaxy Z Fold5 — we ran two audio files: an English clip of 31 s (0.5 MB) and a Chinese clip of 28 s (0.38 MB). Both finished and gave text we could copy. We did not put a stopwatch on them, so no phone timings are listed and none are claimed. On a phone this will occupy the whole device, longer files take longer, and the phone gets warm — on those two runs it was warm to the touch, though the phone stayed responsive and did not stutter. We have not tested it while charging. Practical advice: start with a short clip and be on Wi-Fi, because the first run downloads about 78.4 MiB.", "source_url": "https://clipquill.com/", "updated": "2026-10-03" }, { "id": "sources", "topic": "Who made this, and where the numbers come from", "keywords": [ "who made", "author", "about", "source", "raw data", "github", "zenodo", "doi", "dataset", "provenance", "verify" ], "answer": "There is no About page and no author byline on this site, so there is no name to trust instead of the data. Every figure on the page was measured by whoever runs the page, and the raw data, the measurement scripts and the reference transcripts are published: the repository is at https://github.com/supersophia8888-cloud/clipquill-asr-benchmark , an archived copy with a DOI is at https://doi.org/10.5281/zenodo.22826968 (the concept DOI, which always resolves to the newest version of the dataset, all versions), and the same data as a dataset is at https://huggingface.co/datasets/sophia8888/clipquill-asr-benchmark . If a figure on the page disagrees with the published data, the published data is right and the page is wrong.", "source_url": "https://clipquill.com/", "updated": "2026-10-03" }, { "id": "how-to-use", "topic": "How to use it", "keywords": [ "how", "use", "steps", "guide", "instructions", "start", "upload", "drag" ], "answer": "1. Have the video or audio file on your device — a phone recording, a Zoom or Teams capture, a Reel, or a file saved from YouTube, TikTok or Facebook. 2. Drop it on the page or use Choose a file; Ctrl+V paste works too. 3. Pick the language that is spoken. 4. Wait — the first run also downloads about 78.4 MiB. 5. Copy the text, or export TXT, SRT or VTT.", "source_url": "https://clipquill.com/", "updated": "2026-10-03" }, { "id": "save-reel", "topic": "Saving an Instagram Reel to your device", "keywords": [ "instagram", "reel", "reels", "save", "camera roll", "gallery", "copy link", "screen record" ], "answer": "Your own Reel: open Instagram, go to your profile, open the Reel, tap the three dots, then Save (Android) or Save to Camera Roll (iPhone). Someone else’s Reel: tap the three dots, Copy link, then use Instagram’s own Save feature if offered, ask the creator for the file, or record your screen while it plays. The same goes for YouTube, TikTok and Facebook: whatever you can save to your own device, this page can read.", "source_url": "https://clipquill.com/", "updated": "2026-10-03" }, { "id": "install", "topic": "Installation", "keywords": [ "install", "download", "app", "plugin", "plugin", "extension", "desktop", "nothing to install" ], "answer": "Nothing to install — no download, no plug-in, no app, no signup. It runs in the tab you already have open.", "source_url": "https://clipquill.com/", "updated": "2026-10-03" } ] }