What's new
Changes to Askwho Casts Pro, newest first.
Currently v3.4.0
October 2026
1 releaseInvite friends: 100 credits each
- Inviting a friend is now worth 100 credits, up from 50. You get 100 credits for every friend who subscribes, and they get 100 too, once their first plan payment clears. There's no limit on how many friends you invite.
- Your invite link now sits in your Library, and near the top of Plan & credits, ready to copy.
- If you connect the Cast this page extension before subscribing, it now shows you how to make your free first article on the website.
September 2026
20 releasesChoose your email updates
- Settings now includes an optional choice for feature-update emails, off until you opt in. You can change it at any time, including after your subscription ends. Essential account emails stay separate.
- Feature-update emails include an unsubscribe link that works without signing in. Your subscription, credits and podcast feed are unaffected.
Find Following from Studio
- A small notice in Studio introduces Following to subscribers. Open Following or dismiss the notice once to hide it across your account.
Follow publications in your podcast feed
- Subscribers can follow public Substack publications and RSS feeds from Following. New free posts are converted with your Askwho Casts voices and added to your podcast feed automatically, even when you are signed out.
- Set limits for each publication and post, daily and monthly allowances, and a balance reserved for manual conversions. Posts above your limits wait without charging. Following starts with new posts; older posts and paid previews are not imported automatically.
- Healthy feeds are checked about every five minutes; some publications need less frequent checks. The interval is shown beside each publication, and new posts are queued as soon as they are found. Posts can be converted even when the publisher already offers a voiceover.
More faithful footnotes, equations and article credits
- Footnotes keep linked and italicized words in their sentences and receive one introduction across their paragraphs. Population shorthand is distinguished more carefully from distances.
- More rendered equations survive article extraction, with clearer spoken fractions, probabilities and negation. Explicit guest introductions can credit the guest while keeping the host's introduction in a separate voice.
- More page controls, player instructions and repeated dates are removed from narration. These improvements apply to newly prepared episodes; existing recordings and saved in-progress plans remain unchanged.
A gentler opening and closing
- Newly rendered articles open with Quiet Margins, our original string quartet theme. The title, author and date sit over a quiet music bed, followed by a brief pause before the article.
- A separate closing phrase follows the final words. The article itself remains free of background music, with its full cast of voices.
- Stereo audio preserves the quartet's space and detail. An Article chapter lets you skip straight to the reading. Existing recordings and custom scripts retain their current presentation.
Cleaner article scripts and more consistent quote voices
- Inline WordPress equations stay in their sentences, with factorials read as mathematics. More publisher playback, advertisement, tag and Markdown controls are removed from narration.
- Newsletter dates use spoken punctuation instead of vertical bars, and short dialogue keeps its original punctuation without artificial padding.
- Quotations introduced through a named writer's titled work more reliably keep that writer's voice. These changes apply to newly prepared episodes; existing recordings are unchanged.
Clearer episode lists
- Removed automated review notices from episode lists. You can still report a problem from any completed episode.
Refine connection recovery
- Removed the extra session read introduced in 3.0.5 after live testing showed that database stalls could delay both reads together. Account batching, session joins, bounded connection recovery and private-response protections remain in place.
Recover stalled session lookups
- If the database stalls while checking your session, one backup read can restore the connection sooner. Both reads check the live session; sign-in, purchases and recording requests are never duplicated by this recovery.
Faster session validation and database routing
- Signed-in requests load the session and its user together, reducing authentication round trips while checking session expiry and revocation on each request.
- Requests that need account data run near its database, avoiding repeated long-distance calls. Page files continue to load from a server near you.
Faster account loading and reliable connection recovery
- Account details load together, reducing repeated database waits when opening the studio. A stalled connection gets one automatic retry, with an early retry button and clear recovery if the service remains unavailable.
- The page keeps a reload link available if its application files fail to load. Private account responses cannot be cached, and runtime dependencies include the latest compatible security fixes.
Retire the older full-episode sample
- Removed the older full-length recording from the homepage because it does not represent the current voice technology. The short cast demonstration remains available.
Seek through the homepage voice demonstration
- The homepage demonstration now supports seeking in production. Moving the playback slider or choosing a voice jumps to the intended moment instead of restarting the recording.
A clearer studio, from first visit to daily listening
- The studio keeps its dark brown, cream and amber design, with clearer headings, brighter secondary text and more room around controls. On phones, long episode titles keep their width while a recording progresses. Status, progress and playback controls sit below the title, with larger touch targets and more visible keyboard focus.
- Getting started is easier to follow. The mobile homepage puts the article form first, with direct routes to the voice demonstration, free account creation and plans. New answers explain the free article, supported sources, podcast apps, credits and cancellation. An empty Studio walks through review, recording and listening; loading and searches without matches have their own guidance.
- Article reviews now move into view and receive keyboard focus. If a page may be a paywall preview, its opening and ending excerpts are expanded for inspection before generating. Script voice loading can be retried without losing the text you wrote.
- Settings show when values are loading, saving or saved, and a failed save restores the last confirmed values. Feed addresses are easier to select and copy, including when clipboard access is blocked. Billing and account errors offer retry, and connected-browser controls show whether a disconnection succeeded.
- The demonstration and episode player offer retry when audio loading stalls, keeping the selected demo voice or your episode position and playback speed. Slow library updates can finish without being overtaken by the next poll. If a recording request loses its reply, the studio refreshes the library and account without automatically submitting another recording.
- The updated browser extension fixes the Chrome popup's repeated resizing and keeps page review and recording progress when the popup closes. Reopening the same document reuses a recent quote and preserves the image-description choice. Interrupted submissions are checked against the exact selected episode, including after a browser restart; an uncertain response never triggers an automatic second recording. Finished episodes offer Review updated page for articles changed at the same address.
- Extension connection settings follow a connection completed in another tab and check a pasted replacement before saving it. Expired connections offer reconnect, while failed replacements preserve the previous connection. The install and connect pages have clearer recovery instructions, and the extension privacy notice explains page capture, local progress storage and completion of requested work after the popup closes.
Review the source, hear a more consistent cast
- Before generating in Studio, you can inspect the article's opening and ending, see counts of quotations, footnotes, images and code blocks, and review the detected speaking parts. Extraction warnings stay visible before purchase. Generate uses the saved source from that review, so a page edit cannot quietly change the article you approved.
- Automatic casting carries proposed voices forward from review, and only choices you change count as custom casting. Different identified speakers receive distinct voices, with better handling of aliases, multipart surnames and clear initials. Ambiguous initials and dialogue without reliable speaker evidence keep the shared quoted-text voice; complex fiction can still contain unattributed turns.
- More of the article survives extraction: richer footnotes, notes opened by hovering, nested images and code, and mathematical markup. A footnote containing several paragraphs or an image is counted once. Several publisher-specific fixes recover missing body text, improve interview titles and bylines, and keep unrelated page furniture out of the narration.
- Image descriptions now retain visible quotations and pay closer attention to chart units, uncertainty and what a number measures. A quotation inside an image gets a named speaking part only when the image supplies evidence for that name. Image and code descriptions also account for changes to the source and its context; undelivered descriptions retain their fallback notice and refund their description charge.
- New recordings use speech-level balancing and smooth peak limiting to reduce level swings and leave more headroom before MP3 encoding. Waveform, pitch and transcript checks run independently, so one problem cannot hide another. These automated checks help catch recording defects; they do not certify that every speaker attribution is correct.
- Search your episode library by title, author or publication, and report a wrong speaker, missing passage, description or audio problem at a playback timestamp. A failed report keeps your note for retry. Billing, cancellation and connected-browser controls are easier to find, and account errors offer a way to try again.
- Saved-page retries reuse the retained capture. Interrupted jobs recover or refund more reliably, with stronger protection against duplicate charges and an older run overwriting its replacement. These changes apply to new production; finished episodes keep their existing audio.
Code blocks, described aloud
- A code block used to be one line of narration: "Code block omitted." Now the narrator says what it does or shows, in one to three plain sentences: a Python function that retries a call up to three times, doubling the wait between attempts; a shell command that installs a library; a table comparing three model sizes with their scores and latencies. Terminal sessions, configuration files, diffs, database queries and data tables get the same treatment. Identifiers, versions and figures are spoken as words, never as symbols.
- Back-to-back blocks are described together, and a block seen before is described once and remembered. Description is on by default and shares the existing describe-aloud switch with images and charts; switching it off restores the one-line omission. Each described block is priced as a few seconds of narration in the quote, with no separate charge, and anything that could not be described is refunded the same way undescribed images already are.
- Preformatted text that is really prose, a letter or a poem still reads as speech, exactly as before.
A hundred new voices
- The narration library grows from 105 to 205 voices: 100 new fully synthetic voice designs, 50 in the male register and 50 in the female register, spanning American, Canadian, British, Scottish, Caribbean, South African and New Zealand accents. Every one is a professional narration design, no real person behind it.
- They were chosen by ear, at scale. Every candidate was rendered by the live narration engine on four article passages (an essay, a scene with quoted dialogue, a second take of the essay, and a paragraph of dates, money and percentages), then scored with a naturalness model trained on human listening ratings, checked word by word against the text, and measured for noise, dead air, level swings, pitch steadiness and pace. 546 candidates went in; the 100 that read cleanest and most naturally came out.
- Nothing already rendered changes. Finished episodes keep their audio, locked narrators keep their authors, and remembered characters keep their voices. The new voices join the rotation for new articles and appear in the advanced voice picker, each with its own audition.
Recurring characters keep their voices
- A named speaking part now keeps its voice across everything by the same author. Once a character or a quoted source has been heard in one article, every later article carrying the same byline casts them in the same voice, whoever requests it. Serial fiction is the obvious winner: a story's characters no longer change voice from chapter to chapter. Pages without a byline remember by site instead.
- The memory learns as it goes. A part first heard without any he-or-she evidence gets a provisional voice; the first later article that makes the character's gender clear re-casts them to match and fixes that voice from then on. Parts introduced with clear evidence keep their first voice permanently, exactly as narrators already do.
- Existing episodes are unchanged. Casts already recorded for authored articles seed the memory, so the next chapter of a serial you are following continues from the voices you heard most recently.
- In the browser player, lock-screen and headset controls now work: play, pause, skip back and forward 15 seconds, and scrubbing, with the episode's title, author and artwork shown on the lock screen.
- While an episode waits its turn for a recording lane, the studio says so instead of showing a stalled counter.
Hear who's speaking, then pick up where you left off
- The homepage now opens with a 32-second demonstration: a narrator, two quoted speakers and a footnote voice in one short scene. The active speaker and their line are shown as the audio plays, and a transcript lets you jump straight to any voice. It is an original, fictional example performed with the same synthetic voices that read episodes.
- Paste a link on the homepage and the article travels with you through sign-in, whether by Google or by email link, and survives a page reload for a day. Nothing is generated and nothing is spent until you have reviewed the price in the studio and pressed Generate.
- The player remembers where you stopped in every episode on this device for 90 days, so reopening the studio picks up from the same second. New −15s and +15s skip buttons sit beside the speed control, and a finished episode starts over from the beginning.
- Email sign-in explains itself: a clear "check your inbox" state, a resend option with a short cooldown, a way to correct the address, and a plain note when the security check still needs completing. Body text on the dark theme is a little brighter, and the sign-in page fits narrow phones without wrapping.
- Pricing states the currency and the ten-minute-per-article assumption behind each plan's article count, and the "Most popular" label gives way to a plainer recommendation. Prices and credit rules are unchanged.
- Three quote-attribution fixes: a name that appears only in a comparison ("like Michael Jordan saying…") can no longer become a speaker; a person introduced by occupation ("the orator Isocrates") keeps their own quotation; and names inside a biographical aside ("the handler of Ames and Hanssen") no longer displace the sentence's real subject.
Poems are not code
- Verse, letters and anything else an author typesets in a preformatted block — Substack's "Preformatted text" block especially — is now read aloud with a pause at each line break, instead of being skipped as "Code block omitted". Real code blocks are still skipped, and an equation that arrives as raw notation is now announced as omitted rather than read symbol by symbol.
- Bylines are more often right: authors credited in a WordPress post header, or listed one per line on a page, are picked up whole ("Andreas Lynge, John Bergschneider" rather than one glued name), and a sponsor label in the byline slot is never mistaken for the writer. Pages saved through the Wayback Machine credit their real author again.
- Less furniture in the audio: chapter-navigation rows ("Last Chapter Next Chapter"), classic-blog post footers and comment-form instructions, editorial pull quotes that repeat a body sentence, hidden citation locators, and image filenames masquerading as captions are all left out. A bare calendar date on a page can no longer roll back a day in the spoken intro.
- Small reading fixes: "eg," and "ie," are expanded like their dotted forms, a mid-sentence "etc." no longer forces a full stop, and "==" reads as "equals" once.
August 2026
29 releasesFewer false full stops
- Some page builders split a single sentence across separate blocks, and text pasted from a PDF carries hidden line breaks mid-sentence. Both used to be read with a false full stop at every seam — "Researchers. showed. that…" — sometimes dozens of times in one episode. The reader now recognises when a sentence carries on across a seam and keeps it whole, while genuinely separate lines still get their pause. A line of dialogue that ends inside quotation marks no longer collects an extra stop after the closing quote either.
- Fiction chapters on RoyalRoad now credit the actual author rather than the site's Follow Author button, and the hidden anti-copying sentences the site plants for scrapers — invisible to every real reader — are no longer read into the middle of a chapter.
- Newsletter tracking codes on some article links no longer make the same article count as two different ones.
Posts on X read like the posts they are
- A link to an article published on X now opens with the article's own title and author instead of "Name (@handle) on X" — and because the author is finally recognised, their articles keep a consistent narrator, the same way any other byline does. The cover image and the figures inside the article are found and described in place.
- A long post keeps its paragraph structure instead of arriving as one unbroken block, its photos are described, and names mentioned mid-sentence stay in the sentence. Only the post you linked is read — the conversation underneath it stays out of the episode.
- Posts embedded inside an article are acknowledged in passing rather than having their web address read out loud. And the => arrow now reads as "goes to" instead of half-dissolving into "equals".
The launch month wraps up, as promised
- August's launch promotion ends with the month, exactly as advertised: the 500-credit grants close at midnight UTC on August 31, and unspent launch credits retire with them. Everything made with them is yours — every episode stays in your private feed and keeps playing.
- From September 1 the standing welcome returns for new accounts: your first article free, then plans from $9/month.
- If you joined during the launch month, the studio shows a one-time note about where things stand. Your free first article, if you haven't used it, is still yours — untouched, exactly as promised when the launch month began.
Named opinions keep their own voice
- A shared label qualified with a contributor's name — such as "Opinion (Nuño)" — is now treated as its own speaking part rather than folded into the general Opinion voice. The distinction also survives when one contributor hands over to another halfway through the same quoted passage.
- The printed label remains part of the recording. The studio changes voice at the hand-off, then says "Opinion (Nuño)" exactly as written.
Quote labels stay in the recording
- Labels such as "Opinion" or a person's name still select a separate voice for the passage, but they are now spoken as part of the text instead of being silently removed. The internal quote marker remains routing information only, so a recording says "Opinion" or "Jane Doe" — never "Quote: Opinion" or "Quote: Jane Doe".
- Credits at the end of quoted passages are preserved too. The studio no longer edits visible attribution text out of an article while preparing it for speech.
Substack links stay on the host that works
- Some Substack publications use their www address for the real article while the bare domain returns Not found or never answers at all. The studio was simplifying both spellings to the bare address before it tried to read the page, so an article that opened normally in your browser could fail here after a long wait. It now keeps the working address for reading while still recognising both spellings as the same article, and its browser backup also takes over when the first connection times out instead of returning an error.
Interview transcripts, fully cast
- Interview and podcast transcripts are now cast the way they read. A transcript labels each turn once, on its first paragraph, and then runs on for two or three more — and those follow-on paragraphs used to fall back to the narrator, so more than half of a long interview could end up in the wrong voice. Every paragraph of a turn now stays with its speaker until the next label or a section break, footnotes in the middle of a turn included.
- A speaker quoting someone inside their own turn — a scare quote, a "he told me" — used to have just the quoted words voiced and the rest of their sentence handed to the narrator. The whole turn is theirs now. Fiction written as a script keeps every rule it had.
- Far more first names are recognised when choosing a voice's register, so a Nate or a Clara is cast as one. Names that read either way are deliberately left for the prose to decide.
Launch credits, handed out fairly
- The 500 launch credits are now handed out a few accounts at a time each day. If today's are spoken for when you sign up, your first article is still free right away, and your credits are waiting the next time you visit. Accounts that look like duplicates of one already given credits get the free article instead of a second helping — which is what keeps the offer real for everyone else.
- Signing in now says plainly that it means agreeing to the Terms and Privacy Policy — they were only ever a footer link before.
Subscription status that matches your billing, always
- If you cancel your subscription from the billing portal, Settings now shows it ending on the right date instead of quietly saying it will renew. The portal reports a scheduled cancellation in a newer form than the studio was reading; both forms are understood now, and every subscription is re-checked against billing several times a day so a missed notification can never leave the studio out of date for weeks.
- When a renewal payment doesn't go through, we now know the moment it happens rather than when the subscription lapses, so a card problem can be sorted out while there is still time.
A slow night at the voice engine no longer costs you the episode
- Four long episodes were recorded at once during a stretch when the voice engine was answering slowly, and every one of them failed partway through — not because anything was wrong with the articles, but because each recording gave up after a few minutes of slow answers. That patience was sized for the occasional bad take, not for an engine having a slow evening. A recording now waits out a slow spell instead: it pauses, tries again a little later, and keeps stretching the pause for well over an hour before it considers the engine actually down.
- Recordings also take turns now. Only a few episodes are recorded at the same time; the rest wait their place in line, which is faster for everyone — a crowd of episodes competing for the same engine had each one spending more time waiting than speaking. And when a recording does fail for a passing reason, it gets one automatic second run about forty-five minutes later, picking up exactly where it left off. Only if that run fails too does the episode fail and your credits come back.
Sign-up gates can't pass as full articles anymore
- Some newsletters cut the article off for signed-out readers and ask for an email to see the rest — one live case words the gate, literally, "This is not a paywall." Every machine-readable marker on that page said fully open, so the reader took the preview at face value: the result was a ninety-second episode of one real paragraph, the sign-up pitch read aloud, and a sponsor ad. Two new checks close that class for good: gating language standing in the article text itself now trips the preview warning even when the page's metadata swears everything is fine, and when a page declares how long the full article is, an extraction far short of that number is flagged as a preview no matter what the gate looked like.
- The warning now also says what's actually in the way. A free sign-up gate is not a paywall, and the estimate no longer calls it one — it says the site wants a sign-in or an email first, how much of the article we could read (in the page's own word count when it declares one), and that the Cast this page extension can cast your full copy if the article opens in your browser.
- A captured page that still ends at the sign-up prompt gets the same honesty: the studio says your copy looks like the preview and suggests signing in on the site and capturing again — a free signup is often all it wants.
Bot walls get an honest answer — and a way through
- Some sites answer automated readers with a bot check instead of the article. One of those check pages recently slipped through and became a short episode narrating the checkpoint text itself. The reader now recognizes those pages and refuses them outright, before anything is priced or made.
- And when a page can't be fetched — a bot wall, a paywall, or a page with almost no readable text — the studio now points at the way through: if you can read it in your browser, the Cast this page extension can cast your copy straight from the page you already have open.
- Casting a LessWrong post through the extension no longer picks up the comment thread. Your browser renders comments below every post, so a captured page carried them along — one episode was two-thirds comment section, vote counts and all. Captures now keep exactly what a fetched copy of the article would: the post.
Cast the page you're reading — straight from your browser
- The new Cast this page extension puts your feed one click from any article you're reading — including pages only you can see, like subscriber posts you're signed into. It sends your browser's own copy of the page, shows the price in credits before anything is made, and the finished episode lands in your private feed like any other. Get it at app.askwhocasts.com/extension — for Chrome, Edge, Brave and Firefox.
- Connecting a browser to your account is one click, with nothing to copy or type. Every connected browser is listed in the studio's Settings, where any of them can be cut off instantly. And exactly what leaves your browser is spelled out at app.askwhocasts.com/extension-privacy: the page you chose to cast, and nothing else.
Every episode names its narrator
- Each episode in the studio and on the Feed page now carries its production credit: who narrates it, and how many supporting voices join in. Found a narrator you loved three articles back? The name is right on the row — ready to pick again for a future cast, instead of a mystery you'd have to solve by ear.
- Hover the credit to see the speaking parts — which voice reads which quoted speaker. And since casting happens before recording, the credit appears while an episode is still being made.
Quote-tweets, in the order an argument actually happens
- When an article embeds a post that quotes another post, the quoted post now reads first — in its own author's voice — and then the quoting author's commentary. A retort lands after the thing it's answering, the way you'd read the thread yourself; before, you heard "Cry me a river" and only then found out what it was a reply to.
- When the quoted post already appeared earlier in the article, the narrator says so up front — "A post on X by so-and-so, quoting so-and-so's post from earlier:" — instead of tacking the acknowledgment on after the commentary, where it drifted into the next paragraph.
Numbers that read like a narrator, not a spreadsheet
- Decades finally speak properly: "the 2010s" reads as "the twenty tens", "the 1990s" as "the nineteen nineties", "the '90s" as "the nineties" — across every article, including history pieces down to "the ten eighties". The voice model had started spelling out the plural letter by letter.
- Roman-numeral series speak as numbers: a series header like "Part I" or "(I, IIa, IIb, IIIa, IIIb)" used to make the narrator stumble badly enough to swallow the rest of the passage. "Parts one, two A, two B…" now reads straight through — with the sense to leave "Part L" of the Building Regulations and "(I, for one, dissented)" alone.
- A long list of number shapes now read the way you'd say them: review scores ("9.5/10" → "nine point five out of ten"), vote tallies ("voted 4-1" → "four to one"), date ranges ("October 8th-11th", "June 15-July 10", "the 2025/26 school year"), ISO dates, power units ("500MW"), follower counts ("363m followers" no longer reads as meters), and "P != NP" no longer asserts the exact opposite out loud.
- Math-adjacent typography reads sensibly too: exponents like 10²³, CO₂, §3, ≈, aleph numbers, and footnote stars riding a number — and stray marks like article-end diamonds and t.co link junk are no longer read at all.
- Pasting the same article with app junk on the link ("?open=false" and friends) no longer creates — and charges for — a second copy of an episode you already own.
LessWrong reads again
- LessWrong, the Alignment Forum and the EA Forum recently began answering automated readers with a browser challenge instead of the post — so those links failed with a rate-limit message no amount of waiting would clear. Posts from all three read normally again.
- Where a forum won't serve us the post directly, we now read the same post from GreaterWrong, the long-standing public mirror of those forums, and skip its comment thread. The episode is still the post: same text, same author, same link, and pasting a mirror address gets you the same episode as the original rather than a second one.
- Posts that open with a table of contents no longer start the audio by reading their own section titles aloud.
The script voice picker keeps your cast close
- When you're casting a script passage, the voice picker now pins everyone already reading in the script at the top — in the order they first appear, with how many passages each holds — so giving a recurring speaker their next line never means scrolling the full 105-voice list again.
- The narrator sits pinned up there too, so handing a passage back is always one click. The full alphabetical roster stays right below, unchanged, with every voice still auditionable in place.
The private feed, done properly
- The feed now carries every channel tag Apple's podcast spec calls for — category, explicit rating, show type — so strict apps and feed validators accept it without complaint. (It already passed the W3C validator; now it passes with zero warnings.)
- Scan to subscribe: on a computer, the Feed page shows a QR code. Point your phone's camera at it and you get a one-tap page that opens Apple Podcasts, Pocket Casts, Overcast, or AntennaPod with your show ready to follow.
- The subscribe buttons now only offer apps that can exist on your device — no more buttons that silently do nothing on a desktop — and every button copies the address as it opens the app, so a paste is always ready as a fallback.
- The Feed page is now honest about app support: Spotify, YouTube Music, Podbean, Castbox and other catalogue-only apps can't add any private feed. The page says so, instead of leaving you to wonder why a paste failed.
- Feeds, chapters, artwork and audio now answer cross-origin requests properly, so web-based podcast players and in-browser validators can read them too.
The front door says it
- The landing page and the sign-in card now lead with the launch offer: sign in and get 500 free credits to use throughout the month of August — no card needed. The demo card's price line prices against those credits too.
- All of it reverts to the standing free-article pitch by itself when the launch month ends.
The launch month: 500 free credits, no subscription needed
- Sign up during August and get 500 free credits — no subscription, no card. Everything they buy is the real studio: full-length articles, spoken image descriptions, chapters, and your own private podcast feed.
- Already signed up but never subscribed — or lapsed? The same 500 credits land on your account the next time you open the app. (Active subscribers received their own 1,000-credit thank-you with 2.0.)
- For August, credits are the only gate: while you have them, the studio works without a plan. The launch credits are good through August 31, and they're spent before any credits you already had.
- The free first article pauses during the launch month — 500 credits is simply the better offer — and returns in September exactly as it was, unspent entitlements intact.
AskwhoCasts Pro 2.0 — a new narration engine, and a thank-you
- The narration engine has been rebuilt from the ground up. Every one of the 105 narrators was recreated from longer, cleaner reference recordings — the same voices you know, and every author keeps the narrator they've been locked to.
- Episodes arrive much faster. The narration stage — most of an episode's wait until now — renders many times quicker than before; a typical article's narration finishes in seconds rather than minutes.
- Nothing about quality control changed: every take still passes the same waveform, pitch, and word-accuracy gates, with automatic retakes and best-take selection, before it reaches your feed.
- To say thank you for being here early: every active subscriber has been given 1,000 free credits for the 2.0 launch month. They're yours to use through August 31, they're always spent before your regular credits, and any credits you already had roll over exactly as before.
Fresher private podcast feeds
- Private feeds now identify their exact address and require podcast readers to revalidate before reusing an older copy, while the existing content fingerprint keeps unchanged checks lightweight.
- The Feed page now explains how to request an immediate episode-list refresh in Pocket Casts while an episode remains playable in Askwho the moment it is finished.
- A narrowly isolated push-refresh trial can be enabled for one disposable test feed without changing any customer feed. It stays off by default while private-feed compatibility and the privacy boundary are verified.
Every author gets a voice — and keeps it
- Authors now keep their narrator. The first time an article by a writer is produced, they're assigned a voice — matched to their name's conventional gender where that's clear — and from then on every article by that byline, for every listener, is read by the same narrator. Writers you follow start to sound like themselves.
- The narration library grew from 79 to 105 voices (53 male-register, 52 female-register) — all fully synthetic voice designs, selected from ~650 candidates by measured audio quality (noise floor, clipping, dead air, pitch steadiness) and word-perfect articulation, checked by transcription.
- Nothing already made changes: existing episodes keep their exact casts, and voice assignment applies from now. In advanced casting mode your own narrator pick still wins for that episode, and the picker now opens on an author's established voice.
Try it from the landing page — and cancelling keeps your library
- The landing page's paste box is now real. Paste any article link and see its actual title, listening time, and credit price in the demo card — the same reader the studio uses — before creating an account. Running it still takes a sign-in; looking is free.
- Cancelling keeps your library in the app, not just in your podcast player. Lapsed accounts keep their Feed and Credits tabs: episodes play, the feed address can be recovered (or rotated), and the credit history stays readable. Only making new episodes asks for a plan again — which was always the promise.
- Upgrading your plan mid-month now pays for itself immediately: the price difference is invoiced at the moment you switch, and the matching share of the bigger tier's monthly credits arrives with it — no more waiting for renewal. Downgrades still settle as balance toward your future renewals.
- The free first article's 30-minute cap is now stated up front on the landing page — before signup, not after — and the landing-page reader says honestly when an article is longer than the free one covers.
Full-product review: security, money, and delivery
- A ground-up review of every part of the product — how pages are read, how episodes are made and delivered, how billing works, and how the app behaves when something goes wrong. This release is what came out of it.
- Pages that use older text encodings are read properly again. Previously every accented letter, curly quote and dash on such a page was silently dropped, so "café" was narrated as "caf" for a whole episode; the page's own declared encoding is now honoured.
- Your podcast feed can no longer be broken by an odd character in an article's title. One stray control character used to make the whole feed unreadable to strict podcast apps — which quietly hid every episode you had.
- Resuming a part-finished download is now safe: if an episode was re-recorded in the meantime, the app is handed the complete new file instead of a spliced mix of both. Unsupported range requests answer properly rather than erroring.
- Episodes now declare the language they were narrated in, so podcast apps file non-English episodes correctly.
- Money handling was hardened end to end. A retried step can no longer grant the same refund twice or reclaim credits twice, a downgraded subscription now receives the credits for months already paid for, and purchases made by slower payment methods (such as bank debit) now deliver their credits instead of silently arriving empty.
- If a free article's render fails, the entitlement is now guaranteed to come back — a new sweep catches any that slipped through, so nobody loses their free article to our own failure.
- Sign-in and the free render no longer dead-end when the human check can't load: the reason is explained and can be retried. Failed renders explain themselves in plain language, and a failed free article no longer claims credits were refunded.
- Audio integrity gates: chunks are verified to match the expected audio format before they're used, corrupt audio is rejected rather than spliced into an episode, and episode runtime is computed exactly.
- Security: stricter checks on every address we fetch (including redirects and article images), constant-time comparison for signed links, a tighter content policy on our own responses, and firmer limits on the signed-out preview so it can't be turned into a free crawling service.
- Speaker attribution and read-aloud fixes: money ranges, shorthand millions, decimal commas, fractions before units, compound emoji, and mixed currencies all read correctly now.
Live-fire audit: conversion quality batch
- Twenty recent episodes were re-run against their live pages and reviewed line by line; this release fixes everything found in how articles become scripts.
- Embedded tweets wrapped in WordPress's modern embed block now voice as attributed quotes instead of silently vanishing — one such drop had stitched an author's own words onto someone else's introduction.
- Footnotes now wait for the sentence to finish before speaking, instead of cutting in mid-clause; a very short sentence between two footnotes reads through rather than being sandwiched between voice switches. Footnotes inside image captions are no longer dropped.
- Email addresses hidden by a common anti-spam obfuscation are decoded and read out — listeners no longer hear the literal words "email protected".
- Chunk boundaries no longer land inside classical citations or between an initial and a surname ("D. Hoyos", "Xen. Lac. 2.2"), and words fused across formatting boundaries ("resultsHow") are separated again.
- Slashes read by sense: "2/3rds" as two-thirds, "450m/s" as meters per second, "he/she" as he or she, word pairs with a natural pause — spoken "slash" is reserved for paths and URLs.
- One percent convention per article: every percentage is spoken in words, so "0.07%" and "46%" no longer read in two different styles mid-breath.
- A batch of history-prose fixes: era ranges ("202BC-220AD", "90-31 BC"), circa before any number shape, classical author abbreviations (Polyb. reads as Polybius), "--" dashes, glued en-dashes, multipliers ("10x"), "25H1", "Jan '24", metric units, and Latin shorthand (e.g., i.e., vs, etc, aka, WWII) all read naturally now.
- "$84,000 per treatment course" keeps its plural, and malformed WordPress footnote shortcodes voice as the footnote they were meant to be instead of being read as markup.
The front page says it plainly
- The landing page now leads with the offer: the hero button reads "Hear your first article free", and the sign-in card and how-it-works steps say so too.
Your first article is free
- Hear it before you subscribe: create an account — no card — paste one article of up to 30 minutes, and a fully produced episode lands in your private podcast feed. The wall returns for article two.
- The free episode is yours to keep: private feeds keep playing regardless of subscription state, exactly as they always have.
- The honest-failure promise applies: if a free render fails, the free article comes straight back.
- Fair-use guards around the offer — an invisible human check at sign-in and at the free render, plus service-wide daily limits. Subscribers see no change anywhere.
- Signed up before today but never subscribed? The free article is on your account too, starting now.
July 2026
14 releasesLive-fire audit: reading quality batch
- Every article rendered since the last audit was re-run through the reader and reviewed line by line; this release fixes everything found.
- Paywall honesty, any length: a paid post serving only its free preview now warns before purchase even when that preview is long — previously a long teaser could slip past the check.
- Fiction dialogue: characters who are merely mentioned inside dialogue can no longer be mistaken for the person speaking, and two-sided exchanges keep both sides distinct. Explicitly tagged speech still gets its own voice.
- Sponsor reads, book-plug image captions, gallery credits, and slide-number labels are no longer read aloud; repeated slide captions are announced once.
- Spoken publish dates now match the date on the page (late-evening posts no longer shift a day forward), and WordPress "Posted on" lines stopped being read as an author's name.
- Emoji riding a social-media display name are dropped from the spoken introduction instead of being described aloud.
- A batch of number fixes: bare decimals (".05"), ratios ("6:1"), slash dates, "$1.1 billion USD", ampersands, and "0s and 1s" all read naturally now.
- Substack share links from custom-domain publications no longer create a duplicate copy of an article that already has audio.
- Pages that refuse ordinary server fetches are retried through a real browser, so more articles ingest on the first try.
Quote attribution: recall
- Blockquotes introduced with an explicit lead-in ("…he marveled:", "As Berlin interprets Tolstoy:") now inherit the introduced speaker's voice instead of falling to the generic quote voice.
- A follow-up quote from a speaker the article just named continues in their voice — with a hard rule that any ambiguity (two candidates, credit lines, mere mentions) stays generic.
Quote attribution: precision
- A quote can no longer be pinned on the nearest name from a previous paragraph — attribution now demands real evidence next to the quote, and joint or unnamed authorship falls back to the neutral quote voice instead of guessing.
- Titles of talks, papers, and sections introduced in prose are read as part of the narration, not performed as if someone said them.
- When a document speaks ("the report says:"), the quote keeps the neutral quote voice — we never invent a speaker for it.
- Social-embed author names are no longer truncated when long (organization accounts keep their full name).
Quote-attribution scoreboard
- Groundwork for smarter quote attribution: each episode now records how many quoted passages were matched to a named speaker versus the general quote voice.
- A hand-labeled scoring corpus of real articles now gates every attribution change — improvements have to prove they never make a correct attribution worse before they ship.
Language awareness
- Articles are language-detected at read time (page markup first, text profiling as fallback).
- Non-English articles now render in their own language: the voice model is told the target language instead of being forced to English, and cross-lingual cloning keeps the narrator's timbre.
- Quality control transcribes takes in the article's language, so non-English episodes stop failing spurious word checks.
- Text normalization no longer injects English words ("dollars", "slash") into non-English scripts — digits and symbols pass through for the voice model to read natively.
- The estimate card warns when an article isn't in English: our narrators are English voices, so expect an accent.
- And this page: the version stamp in the footer now opens the technical details and changelog.
Paywall honesty + cleaner intros
- Pre-purchase paywall detection: when a page only serves its teaser, the estimate says so before any credits are spent and points to the Script tab instead.
- Intro metadata hardening: byline and title extraction survive a batch of hostile page shapes (glued markup, placeholder site names, site-constant titles).
- A round of text-normalization fixes for money, slashes, markdown residue, and elongated shout-words.
Social embeds get voices
- Embedded social posts (tweets and friends) are now read as attributed quotes in their own voice instead of being dropped.
Origin patience + fewer false alarms
- Transient origin failures (rate limits, hiccups) retry on a minutes-long ladder instead of failing a paid render; permanent refusals still refund immediately.
- Word-check QC stopped flagging number-dense passages the narrator read perfectly (spoken numbers vs digits is transcription ambiguity, not an audio defect).
Script tab + scholarly pages
- Direct text input: paste or write your own script, hand passages to chosen voices by highlighting, and render it like any article.
- Scholarly rendering support: arXiv (native HTML and ar5iv), Distill-family pages, MathJax-era blogs — math folds to speakable text instead of symbol soup.
- Interview Q&A and transcript formats get proper per-speaker voicing.
- QC alerts consolidated into one per-episode digest.
Reliability audit
- Atomic episode finalization and durable refunds — money and audio can no longer disagree after a crash mid-pipeline.
- Best-of-takes doctrine: an episode always ships if any audio exists; quality gates pick the best take rather than blocking.
Advanced casting + referrals
- Choose Voices mode: the estimate detects the article's speakers and lets you cast each one before generating.
- Referral program: share a code, you both get credits when their subscription starts.
- Image alt-text feeds spoken descriptions when captions are missing.
Footnotes + voice curation
- Footnote and endnote lifting across blog engines: notes are spoken in place, in their own voice, at their anchor.
- Voice pool curation with a strike system — a voice that misbehaves twice leaves the rotation.
- Chapter timing hardened against player-side drift.
Storefront + audio QC
- Public storefront: sample shelf and a live demo estimator, no account needed.
- Per-chunk audio quality gate: waveform sanity, pitch profile against each voice's reference, and word-error checks against a transcription, with automatic retakes.
Public launch
- Signups open. Paste an article link, see a priced estimate, get a multi-voice audio episode in your private podcast feed.
- Voice-cloned narration, automatic speaker attribution for quotes, spoken image descriptions, chapters at section breaks.
- Subscription billing with rolling credit lots.