Default speaker detection moves from pyannoteAI to Alibaba Cloud Qwen Filetrans. The complete audio file is stored temporarily in private Alibaba Cloud object storage and processed into sentence-level anonymous speaker IDs. Qwen results are applied directly to speaker labels; no real-name or cross-video identity database is created. pyannoteAI remains only as an explicitly configured legacy fallback. The audio-processing notice is updated to 2026-11-07.
2026-11-05.2 · Reader-video associations expanded
Public reader associations now cover YouTube, Bilibili, Bilibili International, Douyin and Xiaohongshu. When a signed-in account reopens audio or video they transcribed from a local upload, the server saves the same-page association only if the current page and upload belong to that exact account. Administrators see only “Local upload”; no filename, reader link or transcript is shown or generated. The existing button-recording and account statistics controls apply, and historical pageviews are not backfilled.
2026-11-05.1 · Qwen base transcription
All base speech transcription moves from OpenRouter whisper-large-v3 to Alibaba Cloud Qwen Filetrans. Audio segments are written to private object storage in mainland China and deleted after the task; old Whisper caches are not reused for new Qwen requests. OpenAI remains only for optional GPT correction and timing verification, while pyannoteAI remains the speaker-detection processor. The audio-processing notice moves to 2026-11-05 and requires fresh confirmation before new transcription or speaker detection.
2026-11-05 · Audio processor update
The direct base-transcription recipient is now Alibaba Cloud; OpenAI audio use is now optional GPT correction and timing verification. This notice is stored separately from the Terms of Service, and saved content remains readable.
2026-11-04.1 · Text-model migration
Summaries, translation and reading AI punctuation now use OpenAI gpt-6-luna Fast instead of DeepSeek. The same subtitle or text-to-process content is sent without an account name, and free or purchased allowances are unchanged. Audio recipients and the separate audio-export consent are unchanged.
2026-09-25 · Public transcript pages pilot
Public transcript pages (/en/transcripts/) for a small number of YouTube videos whose creators released them under CC BY; at most 100 pages. They are selected and generated by the operator. User tasks, uploads, speaker labels and translations are unaffected and never appear on public pages. Creators and rights holders can request removal at copyright@vidleaf.app. The terms, privacy policy and content rules are unchanged.
2026-11-02 · Audio-recipient notice update
The audio-recipient notice is now version 2026-11-02 and includes OpenAI for transcription and subtitle-interval verification. Audio is sent only after the updated audio notice is accepted. Previously queued tasks with only the old consent retain their original route. This notice is separate from the Terms of Service; existing content remains readable.
A controlled diagnostic extends the anonymous purchase-funnel counts. When a signed-in user selects purchased S/L credits or a single pass and the quote cannot start because of a material duration limit, paused processing, unavailable paid processing, daily capacity, an unsupported speaker-detection source, an expired uploaded original or another fixed server limit, administrators can see the account, feature, specific reason and time to investigate the paid flow. The record excludes video, transcript, amount and order ID and is not sent to Google or other third-party analytics. Ordinary funnel counts still contain no account. Earlier “blocked for another reason” records did not save an account or reason and cannot be reconstructed. Diagnostics are kept long-term and removed with the account.
2026-10-21 · Pack details and first S/L purchase offer
Plans and terms now show the existing per-asset limits: 180 minutes for S and Single Pass, 300 minutes for L. Free allowances are unchanged. S/L show 10/50 transcription hours, separate first-purchase and regular prices, and 12-month validity. Single passes no longer reserve or consume the first S/L purchase offer. Unresolved S/L orders still reserve eligibility, and refunded S/L purchases do not restore it. Prices, allowances and existing order agreements remain unchanged.
2026-10-21.1 · Agreement and first-use statistics defaults
On first use with no previous choice, campaign attribution, submission links, account activity statistics, Google Analytics and configured Google Ads measurement are enabled by default and can be disabled separately. Defaults do not override previous refusals, deletion, expired or unverifiable choices. DNT/GPC and the global statistics opt-out still apply; past activity is not backfilled. Agreement checkboxes for signup, audio processing, purchases and refunds start selected and can be unchecked. Consent or an operation is submitted only after the final button is clicked; opening a page does not register, transcribe, pay or refund.
2026-10-21 · Default signed-in browsing links
While traffic statistics are enabled, new signed-in pageviews and explicitly linked tasks are automatically associated with the server-verified account, including existing sessions. Campaign statistics and a separate checkbox are no longer required. Traffic opt-out, DNT/GPC and record deletion remain available; account changes never relabel old pages, and missing history is not backfilled. Audio recipients and the separate audio-export consent version are unchanged.
2026-10-19 · Email signup and password reset protection
Email signup and password reset now use Cloudflare Turnstile before sending mail to prevent automated abuse. Reset notices are no longer emailed for nonexistent, disabled, deleted or passwordless accounts; the page uses a generic response. This update discloses Cloudflare browser verification and allows security measures necessary to address ongoing abuse to take effect upon publication, with notice in the affected flow. Other policy changes retain their existing notice periods. Audio providers and separate audio-export consent are unchanged.
2026-10-11 · Long-term record retention
Traffic, page history, clicks and action results, campaign attribution, account activity and sign-in security records are retained long-term without automatic age-based deletion, including existing and new records. Hourly campaign-arrival and purchase-funnel counts are also kept long-term. Existing account deletion, statistics deletion and link-withdrawal controls remain available. Deleted or uncollected history cannot be recovered. Cookie, sign-in session, advertising authorization, short-lived deduplication and original-upload lifetimes are unchanged.
2026-10-05 · Sign-in IP addresses and approximate locations
Sign-in security records now include the IP address and approximate country, region and city looked up in a local database, to identify unusual sign-ins and investigate account security. IP addresses are not sent to a geolocation provider; precise location and browser details are not collected. Records are visible to the account owner and administrators and retain the existing 90-day rolling cleanup. Older records are not backfilled, and proxy exit locations may differ from a person's actual location.
2026-10-04 · Optional identity, browsing and task links
Adds a separate, default-off choice to link account identity, browsing and tasks. After this independent choice, only future events record server-verified identity; optional account statistics are not enabled by it. Account changes require a new choice; older grants are not expanded and historical browsing is not backfilled or guessed from IP addresses or times. Optional links are kept for up to 90 days; withdrawal removes the links while necessary task records remain. Earlier video-submission and audio-export consent versions remain separate.
2026-10-04 · Optional visit and submission activity
Adds a separate, default-off option to link video submissions to visits. After explicit agreement, the administrator can group future submissions, current processing results and recorded page categories by pseudonymous browser and visit IDs, retained for up to 90 days. Signed-in accounts must also allow account activity statistics. Earlier submissions are not backfilled and older grants are not expanded; disabling the option removes the links. Existing traffic page categories can be grouped by browser and visit without displaying cookies or IP addresses. Known internal activity is excluded from this dashboard. Audio processing and its separate consent version are unchanged.
2026-09-30 · Video submissions in the admin dashboard
The admin dashboard gains “Video submissions”: daily counts of the videos submitted by guests and by signed-in users, and a list of each submitted video link with its title, duration and processing result, marking whether it was submitted while signed in. A submission made without signing in is shown only as “Guest”, with no identity; local uploads never show the file name. Submissions by administrator and test accounts are counted, not listed. From 9 October 2026 (Beijing time), a submission made while signed in also shows the account's username; this notice comes at least seven days before that applies. Privacy Policy §3 is updated accordingly. The kinds of data collected, recipients and retention are unchanged.
2026-09-29.1 · Purchase funnel counts
When a signed-in user opens a processing confirmation, opens “Plans & purchase” or submits a task that uses purchased credits, the server counts it per hour by fixed category (step, processing feature, reason or outcome), to judge whether the free allowances and packs fit. No account, video, transcript or amount is recorded, and the counts are not linked to traffic, campaign or account activity statistics. So that an account counts once a day, the server hashes the username one way with a random value for that day; both are deleted after 36 hours. The counts are kept for 90 days. The administrator dashboard also sums orders, payments and refunds by time, showing totals only. Other data collected, recipients and purposes are unchanged.
2026-09-29 · Campaign arrival counts
When a page is opened from a link with campaign labels, the server counts opens per campaign link per hour, for every visitor by default. No cookie is set and no IP address or visitor ID is stored; the counts are not linked to traffic statistics or optional campaign attribution, requests with DNT/GPC are not counted, and the counts are kept for 90 days. Other data collected, recipients and purposes are unchanged.
2026-09-28 · Audio formats and renewed consent
The audio notice now covers original audio files meeting format and size requirements as well as transcoded 16 kHz mono MP3 segments of up to 15 minutes each. Recipients and processing purposes are unchanged. Separate audio consent moves to 2026-09-28 and must be renewed before the next transcription or speaker detection.
2026-09-27.1 · Corrected processor names in the Privacy Policy
Privacy Policy §4 and Terms of Service §11 correct two processors: the direct recipient for speech transcription is OpenRouter (United States), which routes audio to a model provider (currently DeepInfra, Together AI or Groq, all in the United States) — Groq, named before, is only one of the possible providers; summaries and translation go to DeepSeek (mainland China), previously misstated as SiliconFlow. Since our call records begin (2026-09-13) these two have been the actual recipients, so this is a correction of the notice, not a change of provider. §5 adds their retention notes. Because the audio-recipient notice changed, the separate audio-export consent version moves to 2026-09-27: accounts that consented before will be asked once more before the next transcription or speaker detection. Data categories, purposes, retention periods and prices are unchanged; as with every version update, new cloud actions require reading and accepting the current terms.
2026-09-27 · English legal pages use the English interface labels
The English Terms of Service and Privacy Policy now name interface labels the way the English interface shows them (for example “AI-generated”, “AI translation”, “Human captions”, “Traffic statistics”) instead of quoting the Chinese labels, and links from the English legal pages back to Vidleaf open the English site. The meaning is unchanged: the kinds of data collected, recipients, purposes, retention and prices are the same, and the separate audio consent version is unchanged. As with every version change, please accept the current terms before starting new cloud work.
2026-09-25 · Speaker detection for uploaded audio
Once an uploaded audio file has been transcribed, speaker detection works on it the same way as on platform videos. To make that possible, a successfully transcribed original is no longer deleted as soon as its job ends; it is kept for at most around six hours after upload (a file never submitted was already kept for at most around six hours, so the maximum retention is unchanged). After that window, upload the same file again to run speaker detection; a file already transcribed is not charged again. Originals whose transcription failed or was cancelled are still deleted at once. §6 of the Privacy Policy and the Terms of Service sentence on speaker detection for local uploads are reworded accordingly; the kinds of data collected, recipients and purposes are unchanged, and the separate audio consent version is unchanged.
2026-09-24.2 · Extracted audio and AAC m4a follow the duration limit
AAC audio tracks extracted on your device from mp4 or mov videos, and AAC audio in m4a files, are no longer limited to 60 MiB: a track averaging at most 320 kbit/s can be uploaded as long as it fits the per-asset duration limit of the selected entitlement (120 minutes; 180 minutes with L). An m4a over 60 MiB is first checked on your device, and only its AAC audio track is uploaded. Other audio files remain limited to 60 MiB. For multichannel or higher-bitrate tracks over 60 MiB, you are asked to export a lower-bitrate audio file first. The Terms of Service are reworded accordingly; the kinds of data collected, recipients, purposes and retention are unchanged, and the separate audio consent version is unchanged.
2026-09-24.1 · More audio formats and on-device audio extraction
Uploads are no longer limited to mp3, m4a and wav: aac, flac, ogg, opus, amr and wma are also accepted, and are converted to 16 kHz mono mp3 before transcription. You can also choose an mp4 or mov video: the page extracts its AAC audio track on your device and uploads only the sound, so the video picture never leaves your device; if the track cannot be extracted, you are asked to export the audio first. Terms of Service §1 and Privacy Policy §1 are reworded accordingly. The kinds of data collected, recipients, purposes and retention are unchanged, and the separate audio consent version is unchanged.
2026-09-24 · Account sections, Pack L and first-order rules
The account area separates overview, plans, orders, allowances and security, with separate administrator entry points. Pack L and eligible S/L first-order prices are added. Single Pass is not discounted; paid orders of any product affect eligibility. Refund components are allocated in cents, and orders retain their accepted language and policy. Purchases open only when the page marks the product available; eligibility and price are checked before ordering. Existing orders keep their terms; the separate audio consent version is unchanged.
2026-09-23.2 · Launch packs and refund policy (in preparation)
S lasts 12 calendar months. A single pass must start within 30 days of purchase, with remaining uses valid for 30 days after starting. Ordinary refunds must be requested within 14 days of successful payment: unused orders receive a full refund; used orders receive the unused component share. One successful refund closes every remaining entitlement in the order. Each purchased task requires source selection and a use estimate. All processing has a 120-minute limit; uploads remain capped at 60 MiB. These prepared product rules apply when published; live sales are still closed. Contact support for non-delivery and other after-sales issues. Payment records are separate from optional analytics; separate audio consent is unchanged.
2026-09-23.1 · Local upload privacy and cleanup
Transcripts, translations and summaries from local uploads are restricted to their account and are not shared with public-video readers. Original files are removed after jobs where possible, while periodic scans clear unsubmitted files; an outage or storage fault can delay cleanup, followed by a recovery scan. The privacy policy clarifies that hiding a task is not deletion of its result. Please accept the current terms before starting new cloud work.
2026-09-22.1 · Local audio upload
Adds a path for uploading local audio (mp3, m4a or wav) and having it transcribed. Privacy policy section 1 now names audio files you upload; section 6 states their retention (deleted as soon as transcription ends; an upload never submitted is kept at most 6 more hours); section 4 states that uploaded audio also leaves mainland China. The content rules state that you may upload only audio you recorded yourself or are authorised to use. The list of processors, the purposes and the retention periods are unchanged, so the separate audio-export consent keeps its version.
2026-09-20.3 · Optional account statistics available
The initial rollout no longer has a seven-day wait. After accepting the current policy, you can opt in to account statistics. Collection is off by default, past activity is not backfilled, and you can opt out or delete the events.
2026-09-20.2 · Policy update notice
New-transcription duration limits are now labeled explicitly, and optional account activity statistics were introduced. The initial waiting period has been updated by version 2026-09-20.3. Saved content remains readable when you decline updated terms.
Feature changes, legal text versions and the people or workflows they affect.
Open a video in one step while watching it
Replace youtube.com or bilibili.com in a video's address with vidleaf.app and press Enter: the video opens in Vidleaf, just as if you had pasted its link in the reader. Videos already read open directly, new ones are read under the usual rules (guest daily reads still count), and transcription never starts on its own when a video has no captions. On a computer you can also drag the “Open in Vidleaf” button from the transcript tool page to your bookmarks bar and click it on a YouTube or Bilibili video page. These addresses only open videos and are not indexed by search engines.
Local audio upload, chapter exports and mobile reading
After signing in, you can upload MP3, M4A or WAV audio. Uploading does not start transcription; review the file duration, allowance and processing notice before confirming. The new audio-to-text guide explains limits, privacy and TXT/SRT exports. Results from local uploads are account restricted, and periodic cleanup sweeps originals never submitted. Summary chapters can be selected for TXT or Markdown export. The mobile reading toolbar now collapses selected-line actions when nothing is selected, while full-track export remains available.
Steadier reading pages and clearer public navigation
Homepage examples and reading links now have a more stable first display. Tool pages explain export choices, source checks and failed requests, with links between English and Chinese tools, guides and allowances. Shared links use a consistent site image.
English home page and transcript entry pages
An independent English home page and Bilibili and English YouTube transcript pages explain sources and limitations before handing the entered video to the reader. This affects people starting a video from a public page.
Cancel a running task
Reading, transcription, summary, translation and speaker tasks can be cancelled from their progress area or task history. Existing results stay available; processing already submitted to a provider may still finish and incur a charge. This affects people starting these tasks.
Sign-in records and their retention
The privacy policy now covers the time, method and outcome of sign-in attempts, a 90-day retention period and administrator access to these records. These records do not contain IP addresses or browser details. This affects people signing in to an account.
The privacy policy now covers the email address you voluntarily submit for a top-up availability notification and explains its purpose. This affects people registering for that notification on the home page.