What's changed in SpokenWords 14

Reading in the Studio

Read the document with its formatting

The reading area now shows the document's paragraphs, headings, lists, tables, notes and other formatting together. A chapter the server has no imported document for still shows one sentence per line. The command Original document in a new window has been removed; you read and work with the document in the Studio.

• Pictures appear with the text. The words appear first, with pictures loading afterwards.
• Tables are centred when they fit, and wide tables have more room. Tables that do not fit within your chosen line length can use the wider reading area, and scroll within their own area when they are wider still. Moving to a sentence also brings it into view when it is off to the side.
• Small raised or lowered letters and numbers keep their appearance. This helps you read formulas and other text as written.
• Page breaks are visible. You can see where a printed page starts when the document includes page numbers. A page break inside a list item or table cell takes a line of its own. In a table, its rule crosses the full reading column rather than stopping at the cell edge. Its status dot stays at the start edge, where it works as the page break's keyboard handle.
• Text size scales the whole document. Changing Text size scales headings, lists, tables, captions and page numbers along with the main text. Changing reading settings keeps the same middle part of the manuscript in view instead of jumping to another row.
• Announcements appear beside the content they introduce. When a table, picture or other part of the document has an announcement, you can see the extra words to be read just before it, and the closing words just after it. Each end is now marked with a single icon rather than the words Announcement and Denouncement. A screen reader hears Announcement or End of announcement from the icon, and resting on the mark explains that the announcement comes from the book's structure and is read like any other sentence.

Status dots remain useful in every view

• Clicking a later sentence marks its dot. The visible ring belongs to the keyboard and menu handle rather than surrounding the whole sentence.
• A hidden dot can still be clicked. Clean and Word marks views leave a small invisible target where the dot belongs, so a mouse or touch user can still open the sentence menu.
• Lists keep their reading order. A list item's status handle follows the number announcement that is read before it. Arrow-key movement follows that same order and skips controls that are not currently shown.
• Status names explain more. The chapter tree and article rows now explain Draft, Recording, Review, Done, importing and exporting. Screen readers hear the same stage meaning.

Follow links, go to a page and search

Links in the document now work in the reading area. A link in the text is marked with a dotted underline. You can follow a footnote or a link to another part of the publication. Website links open in another tab, leaving the Studio open. With the keyboard, Up and Down move through the status dots and the links in the order they appear in the text, and Enter follows the link you are on.

Open Table of contents from the handle at the left edge of the text, then choose Go to page… and enter the printed page number. You can also type # followed by the number in the search box. This searches page numbers in the open chapter. If the open chapter prints no page numbers, Go to page… is switched off and gives the reason: This chapter has no page numbers.

If search finds words in source material that the open chapter cannot draw, the find bar says so and the chapter line carries the match. Search results therefore no longer appear to vanish simply because that source fragment has no visible manuscript element.

A chapter that will not open leaves you where you were

If the next chapter's content never arrives, because the connection is gone or the server refuses, the Studio used to claim you had moved while the manuscript, the take list and the overview stayed behind, and told you nothing. You now stay on the chapter you were reading, the chapter tree puts its highlight back, playback and your reading position are untouched, and the status bar says The chapter could not be opened. You are still on the one you were reading. The word chapter follows the publication: an article or a section is named as such.

Reading settings and Studio panels

Choose what you see while reading

Under Reading settings, the new View choices are:

• Clean: hide the status dots and recording marks for a quieter reading view. The current sentence, playback highlighting and the document's own formatting remain.
• Word marks: show the recording marks on the text without the status dots.
• Status dots: show the recording marks and each sentence's status dot. This is the starting choice.

A separate Spoken form choice controls the reading instructions shown with the text:

• Off: hide the added reading instructions.
• Authored: show instructions from the document and those added by people. This is the starting choice.
• Full: also show reading changes the system added on its own. Where their position is clear, the alternative reading appears beside the words it replaces. No text carries such marks yet, so the choice is unavailable today and says This text has no automatic marks.

Spoken form also applies to structural reading. Announcements and silence markers follow the Spoken form choice. Number announcements inside list items appear only with Full, while authored announcements appear with Authored as expected.

These choices are remembered between visits. They change what you see, not what the finished audio should say.

Find the panels at the right edge

The Studio's panels at the right edge now open from the icons there; the chapter list keeps its own handle on the left. Choose an icon to open its panel or switch to another one. You get the panels your work needs: Reading settings always, Bookmarks while you record or review, Annotations whenever your role's view allows editing, whatever the article's current stage, and Post-processing if you manage the production and are not in the editor's view. Annotations never appears beside Bookmarks or Post-processing, because it belongs to the editor's view alone.

The panel you last had open opens again the next time you come back to the Studio, as soon as your work offers that panel again. If you close the panel before you leave, it stays closed when you return. Panels stay hidden until the Studio is ready, so no empty panel frame appears while it loads.

Use the gear icon for Reading settings, or press V. Bookmarks, Post-processing and Reading settings are no longer listed in the ⋯ menu, and Annotations no longer has its own tab on the edge of the reading area. Keyboard shortcuts and the available production actions remain in the menu.

The panels at the right edge can also be pushed open or closed with a finger. Clicking the rail button and pushing the panel move it through the same open and closed states. A focus or reveal inside a closed panel no longer drags the whole Studio sideways.

The status bar floats over the manuscript

The status bar floats over the bottom of the manuscript. A new message no longer reserves a band or resizes the reading area. The manuscript has enough scroll clearance to move its final lines above the floating message.

Fewer controls in focus mode

Focus mode now hides the controls above the waveform as well as the surrounding panels and toolbars. The waveform and connection messages remain visible. Press V to open Reading settings even while the panel icons are hidden. Close settings to return to reading.

Reading instructions and editing

See how a sentence is read aloud

Choose How this is read aloud from a sentence's menu to see its Spoken form. This shows alternative words to read and lists its pronunciation, emphasis, pauses and other reading instructions.

When the information is available, instructions say whether they came from the source document, the editor or the Studio. A sentence with nothing to read now distinguishes Marked to be skipped during narration from No words to read aloud.

Words that did not match shows the expected words alongside what speech recognition heard. It also says when nothing was heard for a word.

Saved instructions come back correctly

• Saving edits works again. The problem that prevented editor changes from being saved has been fixed.
• Reading instructions return when you reopen the chapter. Saved Pause, Pronunciation, Reading form and Emphasis changes appear again, alongside Comment, Tone and Structure.
• Saving preserves the document's other instructions and formatting. Editing in the Studio no longer removes existing marks simply because they were added elsewhere.
• Instructions are clearer in the text. Narrators can see the editor's reading guidance on the relevant words. Instructions crossing a change in formatting fit around the document's text.
• Pause names and marks agree. A Short, Medium or Long pause now saves the grade you chose, not only its length in milliseconds, and pauses without a set duration receive a readable description.

Editing choices match what can be saved

Emphasis now offers Light emphasis, Normal emphasis and Strong emphasis. Extra stress has been removed.

Percent and Abbreviation have been removed from Reading form because those choices could not be saved correctly. Alternative reading, Spell out, Year and Currency remain.

Editing tools now explain when a sentence cannot accept an instruction. This includes text marked to be skipped, page breaks and announcements that cannot be edited here. The toolbar and sentence menu apply the same rules. If a save is refused, the message gives the reason instead of only saying that saving failed.

Recording and listening

Follow your reading as you record

During continuous recording, the Studio follows your reading and marks possible differences before you stop. It checks against the words you are meant to say, including alternative readings.

• The reading mark moves between speech recognition updates. It follows your pace and pauses, rather than waiting still for every new result. A narrator the Studio has never heard before is marked a little behind their reading rather than ahead of it, because a mark that lags is corrected by the next answer while a mark that overshoots is not taken back.
• Possible differences have a different mark from confirmed differences. A dashed underline means the result may still change. A wavy underline marks a difference supported by the words heard so far. These are reading aids; the result can still be corrected as you continue.
• You can inspect what was heard once the recording has been checked. Point to a marked word, or open How this is read aloud for the sentence. While you are still reading, the marks do not show the heard words.
• A pause does not push the reading mark ahead. Following your place takes account of the time you actually spend speaking.
• Sentences are marked as heard while you are still recording. Once speech recognition has passed a whole sentence, that sentence's status dot reports Heard and explains that the take is saved when you stop. When you stop, each sentence's own recording status takes its place.
• A slower device gives you less while you read. The app measures what your device can manage and sets Speech feedback to match, as described under Speech recognition below. Only Realtime marks the word you are reading. At Sync on pauses there is no word mark: the reading moves on to the next sentence when you pause after the end of one, and the heard mark above is the only report you get while you read. At Sync after stop nothing is marked while you read, and the recording is checked once you stop.
• Without live speech recognition the mark is a guess. It still follows your pace and pauses, but no differences are marked while you read; they appear after the recording has been checked.
• Stopping gives speech recognition time to finish. The last words are no longer dropped just because you pressed Stop. If it cannot finish, it stops waiting after a limit instead of staying busy indefinitely.

Get a useful first answer soon after Stop

The realtime model now checks every finished recording first, even when this device uses Sync after stop. It quickly marks which sentences were covered and which words may differ while the final model continues in the background. Its preliminary marks are then replaced by the final answer rather than being saved as though they were final.

• A short take marks only the sentences it actually read. One word recorded against three short sentences used to mark all three, and kept them marked until the final transcript arrived over a minute later. As soon as something has read the take end to end, only the sentences it placed words in stay marked.
• Reading a sentence again replaces the older recording's contribution. A later take is no longer judged by words heard in the take it replaced.
• A nonsense answer no longer marks the whole manuscript as misread. A speech model that answered a 30-second take with one made-up word used to have that word filed as the transcript, which marked every word wrong and reported no fault. The final model is never handed a full 30-second field, and a word whose times cannot be real is rejected. The cost is stated plainly: a minute of unbroken speech with no pause to cut on is now read in three pieces instead of two. Recordings with natural pauses are cut at those pauses and never meet this path.

Hear the whole sentence when you play it back

Playback now starts where the sound begins, not where the word was marked. Speech recognition marks the moment a word was recognised, which on the reference recording was between 0.3 and 1.5 seconds after the narrator started speaking. Playing from that mark cut the opening off every sentence, and sometimes the end too. The Studio now listens to how quiet the room is in the recording and widens the playback window outward until the sound drops back to that room level, stopping halfway to the neighbouring sentence so that two sentences read in one breath cannot replay each other. Nothing is stored or sent. The word times and the delivered audio are unchanged. Trim is once again a tool for removing silence, not for getting your own words back.

Trimming and recording again

Trimming and punch-in recording produce the same result whether the take is still on this device, is being uploaded or already exists on the server.

• A cut saved before upload travels with the take. Final sync applies it when the take is filed, even if transcription was already queued or running.
• The chapter overview can trim a local recording. The scissors control opens over audio the device still holds instead of requiring a server copy.
• Punch-in cuts the old prefix on the device. Saving the new reading no longer leaves the replaced beginning in the take that later reaches the server.
• The words follow the surviving audio. Words completely removed by a trim are marked as not recorded. A word whose audio remains keeps its speech result, so cutting near a mismatch does not falsely clear it.
• Long recordings are handled in natural sections. Finished audio can be divided at the narrator's pauses and paragraph boundaries while it is transcribed, allowing completed sections to appear in the manuscript sooner.

If a trim cannot be saved, the message distinguishes a recording that is being filed, one that has left the device and an edit that did not reach the server. The explanation says whether to wait or open the sentence and save the trim again.

Small changes while working through a chapter

• Per sentence moves you forward after a first recording. Once the new take is saved on your device, the next sentence you can record is selected. Recording a sentence again keeps that sentence selected. If you have already moved elsewhere, the app keeps your chosen place.
• The microphone button sits beside the record button. It opens the microphone settings in Reading settings, where you can adjust the level and choose a microphone for the next recording. Long microphone and language choices no longer overflow their select controls.
• Post-processing waits until recording stops. If you manage the production, applying or previewing post-processing is unavailable during a take. Point at the control, or move to it with the keyboard, to see why.
• A reviewer opening a sentence with no audio gets an explanation. Its menu says No audio yet — nothing to review, rather than leaving you wondering why nothing happened.

Final sync and the recording queue

Continue while the accurate transcript is prepared

Stopping a recording no longer makes you wait for the most accurate speech model. SpokenWords saves the audio on your device, adds a Final sync job and lets you continue. Several recordings can wait in the queue and are handled one at a time.

• Final sync survives leaving the Studio. Another signed-in page can continue the work. If the browser is closed, signing in later resumes from the saved recording and completed checkpoints.
• Recording gets priority. Background transcription pauses before the microphone opens and while realtime speech feedback is active. It resumes automatically afterwards, so two speech models do not compete for the device. Pressing Record no longer waits for the previous take to finish: the transcription pauses at a checkpoint the moment you press Record, keeps its model loaded, and resumes from that checkpoint afterwards with nothing read twice. In the usual case the computer is yours within a moment. At worst, when a chunk is just starting, the pause takes a few seconds.
• The Jobs page lists every final sync. Each job has a readable name, can be retried through the shared queue and keeps its place if another page takes over.
• A failed job keeps the recording. The sentence says Final sync failed, and its explanation directs you to Try again.

See progress on each sentence

Every sentence covered by unfinished work shows whether final sync is queued, running, paused or failed. The background tint remains visible in Clean, Word marks and Status dots views. Running work moves across the sentence; people who prefer reduced motion see a still pattern instead.

A double underline means this device already holds the transcript for that sentence while the upload is still waiting. The underline disappears after the take is filed. Screen readers receive the same distinction in the sentence's status name.

The status bar reports what the queue is doing

The status bar at the bottom of the Studio summarises the recordings this device still holds. It used to remember that a sync had started and could keep spinning after the work had stopped, with no control beside it and nothing but a reload to correct it. It now asks the queue what is running every time it repaints.

• The spinner stops when the work stops. A page that was frozen or in the background when a sync ended no longer keeps promising that recordings are on their way. A queue where every recording has failed shows the warning-coloured 3 recordings waiting to sync line with Try again beside it, instead of an endless Syncing 3 recordings…
• You can watch a transcription happen. While the final model reads a finished take on this computer, the line reads Transcribing the recording, or Transcribing one of 3 recordings when more are waiting, with a progress bar and a time estimate such as about 40 seconds left. Progress follows the audio once the first part has been read, and the machine's own prediction before that. The bar never moves backwards and never fills before the words are actually there. This line appears with no connection too, because the transcription runs on your computer. A team whose takes are read by a network service sees no estimate, because nothing can price that wait.
• A line for recordings left behind. A recording whose tab was closed mid-upload used to look exactly like one that was never tried. It now reads 1 recording was interrupted on its way to the server, as information rather than a warning, because nothing failed. Try again puts it back in the queue and starts the rescue pass yourself instead of waiting for the next page to open.
• A recording remembers why it failed. A failure that happened before the upload step, or a job that ran out of attempts, used to leave the recording looking untried. The reason and the attempt count are now written on the recording, so the line above says what happened and Try again can reach it.

Try again always answers

Pressing Try again on a queue where nothing had failed used to do nothing, not even a flicker. Every press now produces a short message saying what it found:

• Trying 2 recordings again. Something had failed and has been restarted.
• Nothing needed restarting. 1 recording has failed, and it is being tried again automatically. The queue was already on it.
• 2 interrupted recordings are back in the queue.
• Nothing has failed, so there was nothing to try again. 3 recordings are still waiting to be sent.
• Nothing is waiting to be sent from this device.

A press reaches every recording it is about, including one waiting out a pause between attempts or parked behind the lock a recording holds, and a page you return to with the browser's Back button runs the same recovery as a fresh load.

Recordings keep all their edits

• Two changes arriving at once no longer erase each other. A saved trim, the words the realtime model heard and the outcome of an upload can all reach one recording at the same moment. Each used to be able to overwrite the others silently, and what was lost was the account of a take that had not reached the server: why it failed and how many times it had been tried. All of them now survive.
• A saved trim no longer queues behind a long transcription. Each kind of background job waits only behind its own kind. Only the moment a job is picked up is timed. The work itself, a decode or an upload over a slow connection, takes as long as it takes without being given up on.

Work through a lost connection

• Speech recognition keeps running after the connection is lost. A narrator gets local word timings, playback highlighting and mismatch marks before the recording uploads.
• Opening a chapter offline shows the recordings held by this device. If the app cannot load the server's list, it says that recordings made elsewhere may be missing and offers Try again.
• Final sync does not spend failed attempts while the network is absent. The local transcription can finish, and the upload waits for an actual connection.
• Edits made offline are retained. A saved trim is queued with its recording. If deleting a take cannot reach the server, SpokenWords keeps its local audio rather than discarding the only copy.

Speech recognition

Recognition can run on your device

SpokenWords can now recognise speech on your device when your team has that service available. It prepares the required download after you sign in, even if you have not opened the Studio yet. This happens for anyone who could end up reading: a Narrator, a Producer, a Lead or an Administrator. An Editor or a Reviewer downloads nothing.

Reading settings shows whether speech recognition runs On this device or your recording is Sent to a network service. This describes where speech is recognised; recordings still need to be saved to the production.

• Download progress is visible across signed-in pages. You can see how much has arrived and an estimated time remaining when available.
• The app checks what works on your device. It chooses suitable speech recognition for following the reading and for checking the finished take. Successful checks are remembered.
• Recording explains what it is waiting for. If installation is still in progress, the Record control says so. If your team has no speech recognition service, it tells you to ask your team administrator to add one.
• Another tab cannot take over speech recognition silently. If a tab is already using it on this device, the app tells you to close that tab or record there.
• An available network service can be used if recognition on the device cannot start. The app tells you when it makes that change.
• Recognition on your device also runs in Firefox.
• Final transcription runs away from the Studio page. Its own background worker favours steady throughput and leaves the live reading tracker responsive.

Checking recordings also does more to reject words that the audio does not support, including words invented where nothing was said, and a word the recogniser repeated over and over instead of following your reading.

Two models, two names

Your computer runs two speech models. The realtime model follows your reading while you record and marks where you have got to. The final model reads the finished take after Stop and produces the transcript of record, replacing whatever the realtime model heard. Which published model fills each role is decided by measuring the machine, so the Speech recognition page names the model itself where its size matters, such as KB-Whisper tiny or KB-Whisper medium. Everything a reader sees uses these names, including the Final model load row on that page.

The machine decides, and prices before it downloads

• Quality picks the model. Speed only vetoes it. The realtime model is the most accurate one that can read 20 seconds of speech in under a second on the computer in front of you. A machine with power to spare spends it on fewer wrong marks. Before, the smallest model always did the marking, however fast the machine was.
• A final model is priced before a byte of it is downloaded. The install times the medium model over 20 seconds of real narration and estimates from that reading what the larger model would cost here. A model expected to read a take slower than four times its length is never fetched. A machine that used to download 1.17 GB of KB-Whisper large on the strength of its memory alone, and then wait over three minutes after a 30-second take, now spends no bandwidth on it.
• Finished takes are read with KB-Whisper medium for now. Until each model's accuracy has been scored, medium is the one model a finished take is read with. No new machine downloads the large model for that job, and a machine that already holds it keeps using it. The speech worker also hands its models back to the machine when it has had nothing to read for a while, and at once when you press Record.
• A model is downloaded only where it will be used. Devices that rely on a network service do not fetch an unnecessary local model. Model files are versioned, and a replaced file is fetched again instead of being mistaken for the old one.
• Every machine is asked whether it has a graphics adapter. A desktop with a strong processor and a graphics card used to be exactly the machine that was never asked. Each installed model is now timed on the graphics adapter and on the processor, and the faster reading is kept for each model.

The model check shows how far it has got

After the download, the install times each model on this machine. The status bar entry reads Testing speech models on this device — 1 of 2 tested, about 20 seconds left, with a progress bar. The bar shows the share of the work behind you rather than a count, because one model takes most of the time and a count would stand still and then jump. The cheapest model is checked first, so the first reading lands within about a second and prices the rest. The time estimate appears only after that first reading, and if a check runs past its estimate the estimate is withdrawn rather than counted past zero. Download, model check and transcription all word their remaining time the same way: whole minutes for a long wait, five-second steps for a short one.

A computer that cannot run speech recognition says so

A machine that could not build its speech model used to tell nobody. The download bar simply vanished, exactly as it does after a successful install, and the Record button then sent a narrator whose team had a perfectly good network service to ask an administrator for one. Three sentences replace that silence:

• Without a network service, the status bar keeps a standing error: This computer could not install speech recognition, so it cannot be used for recording. Try a computer with more memory or more free disk space, and tell your team administrator that this one failed.
• With a network service: This computer could not install speech recognition. Your team has a network service, so recording still works here and finished takes are transcribed there instead.
• The Record button's own explanation now separates the two causes. Where the team has a service but this machine cannot run it, it says Speech recognition is set up for your team, but this computer cannot run it, so recordings cannot be transcribed here. Try a computer with more memory or more free disk space.

See how much feedback your device can give

Following a reading word by word asks more of a computer than some can manage, so the app measures this device and sets Speech feedback in Reading settings to what it found:

• Realtime: the mark follows the word you are reading, and speech recognition corrects it as its answers arrive.
• Sync on pauses: no word mark. The reading moves to the next sentence when you pause after the end of one, and each sentence's status dot reports what was heard shortly after you finish it, with its misread words marked.
• Sync after stop: no feedback while you read. The recording is checked once you stop.

The installation only estimates the level, from the check it runs while preparing. Your own reading measures it properly: once you have recorded for a while, the app times its work on your speech and sets the level from that, so the level shown — and which levels are switched off — can change from what the installation estimated. You can choose a setting below the measured one, and the levels above it are switched off with the reason: This device was measured as too slow for this level. The choice applies to your next recording. It changes nothing about what was installed, because the installation follows the measurement.

Beside it, Cloud sync decides where the finished recording is checked: off, on this device; on, at your team's network service. It appears only when your team has such a service, and it is switched on and fixed there when speech recognition cannot run on this device.

Saving, connection problems and signing in

Know whether a recording produced a saved take

The app no longer reports that everything was saved when a recording produced no take. It distinguishes a recording with no speech from one that could not be matched to the manuscript. When needed, it asks you to read the passage again.

A sentence no longer stays marked as saving for the rest of the visit after such a recording. If the app cannot check whether recordings are waiting to sync, it says so instead of claiming that syncing has finished.

Keep a recording with the version you read

If someone imports an article again while a recording is waiting to be saved, the app does not attach that recording to the new sentences. It can keep the audio on the server, with a copy on your device, and tells you to record the passage again. This prevents an older recording from being mistaken for work on the new import.

Stay signed in

A connection failure no longer signs you out simply because the app could not check your sign-in. It keeps your sign-in information and tries again. Connection messages now appear on every page instead of only in the Studio, and tell you when the server cannot be reached and when the connection is back.

A tab left open in the background no longer signs you out of the tab you are working in. When such a tab's own session ran out, it used to discard the sign-in information for this browser, including the sign-in the tab you were using had just renewed, so your changes stopped being saved there. It now discards only the sign-in it was still using itself.

Pages open on the sign-in this device already holds. Every page used to spend a round trip to the server confirming your session before anything else could start. A session that is stored whole, matches and has more than a minute left is applied at once. Anything doubtful is still checked with the server, and the server keeps the last word: a session it has ended still sends you to sign in on the first real request.

When a request never reaches the server at all, the message now reads The server could not be reached. Check your connection and try again. It used to name what failed and then end with the browser's own error text, such as “Failed to fetch”. This applies wherever the app reports a failure, not only in the Studio.

Clearer messages when a change is refused

Saving messages are clearer about changes that were refused or could not be stored, including reading position and sentence edits. These arrive over a working connection: they are the server's answer when a change could not be written. The English and Swedish messages now describe these outcomes more consistently.

Productions and teams

Most of this section is for the people who run a production. If your part is recording, editing or reviewing, you will not see the import and export controls described here.

See why an issue cannot be imported

The issue list now shows when an upload is waiting, running or has failed. Issues with unfinished uploads cannot be selected for import. The list also explains when nothing has been uploaded or there is no article content to import.

If an import is refused, the message now names the real reason. A refusal used to say the issue was already being produced even when speech was still being generated. An issue whose upload has not finished is now refused with that reason as well; before, such an issue could be imported with only the articles that had arrived so far. The reason also stays in the notification list after the message disappears.

See when speech settings have changed

If you manage the production's articles, a small gear marker now appears on the article's row before you export it. Its name is Uses earlier speech settings, and resting on it explains what has changed. The same information is reported when an export uses the earlier settings.

This means the product's speech settings have changed since the article was imported. Existing recordings are unaffected. Sentences read by text-to-speech still use the settings saved at import.

To use the current settings, choose Import again. This deletes the affected article's recordings, reviews and bookmarks. The message now explains this consequence.

Recording controls match the assignment

Only a narrator or producer assignment can own a recording. The Studio no longer offers Record to someone whose other role cannot support the take, and its explanation says that recording needs a narrator or producer assignment.

When an eligible lead or supplier administrator reaches an unstaffed chapter, the status bar can offer Assign me as producer. A producer can record and can staff the chapter, so one assignment unlocks both jobs. An assignment added by somebody else reaches an already open Studio without requiring a reload, and device speech recognition starts when the new right to narrate arrives.

Supplier staffing follows the granted work

• A supplier with a lead grant can staff every role from its own organisation for the production.
• A supplier with a producer grant can staff every role for the chapters covered by that grant.
• People from the client organisation are not exposed to the supplier. Each organisation chooses from its own roster.
• Remove appears only when the server says the row can be removed. A supplier cannot remove its own organisation grant, while a person's own assignment row remains separately manageable when permitted.

Production information explains itself

Article rows now explain states such as no work, recording, rejected and approved. Production stage symbols describe what Draft, Recording, Review and Done mean rather than repeating the short label.

Content type is now a read-only field in Production details because it comes from the source product. The help text names that product when it can and directs a manager to correct the product if the value is wrong. The ambiguous clock date has left production cards; Last updated is named in the details view instead. A declined assignment is also visible on the production card without opening it.

Resolve problems without searching for the right screen

• Production problems can offer a direct action. If you are allowed to change the production's details, a message about a missing or invalid setting can take you to the place to correct it, including from the Studio.
• An open Studio catches up with production changes. Changes such as a renamed production or a changed publication type update its title and the words used for chapters, sections, features or articles.
• Deadline controls match your permissions. If your only part in a production is your own assignment, every deadline in Production roles now reads as plain text. Narrators, reviewers and editors are no longer offered a deadline change that could not take effect: on their own row it was accepted and never saved, and on another person's row it was refused.
• Adding a team member gives a clearer refusal. If an account belongs to another organisation, the message explains that it cannot be added here and says who can help.
• An invitation error no longer stands in the status bar. The message appears over the window you are working in and is kept in the notification list, instead of remaining in the status bar after you close the window.
• Three choices on a production have new names. Crew roles is now Production roles, Edit metadata is now Edit production details, and Revive production is now Reopen production. They do the same thing as before.

Production messages use the publication's familiar terms more consistently: chapters for a book, sections for a document, features for a magazine and articles for a newspaper.

Keyboard controls and messages

• Escape closes the control you are using first. Closing a menu or a choice inside a panel no longer also closes the surrounding panel or clears your Studio selection with the same press.
• Clicking a sentence points at its status dot. Outside the editor's view, a click on the text moves the keyboard to that sentence's dot and makes it pulse, instead of opening the sentence's menu over the text. Open the menu from the dot itself, or press Enter while it has focus. Nothing changes in the editor: clicking or dragging over words selects them and opens the menu.
• The selected keyboard position is easier to see. Menu choices now show an outline around the choice the keyboard is on, including when using high-contrast colours, where they showed no mark at all before.
• Help messages are easier to read and less likely to get in the way. They fit better near screen edges, and a sentence's help message no longer covers the handle that opens and closes the chapter list at the left edge of the text. Explanations for unavailable controls return correctly when those controls change.
• Studio messages leave the manuscript in place. They no longer push the text down when they arrive, and they appear in a corner away from the record button. A long message wraps onto several lines instead of running across the button.
• Download messages can show a progress bar. The same progress is available in notifications and the status bar.
• Messages across the app have been rewritten. More than two hundred English and Swedish messages now use plainer, more direct wording. The two languages were reviewed separately and then checked against each other, so a message says the same thing in both.
• Two more controls have clearer names for the people who run a production. In the Studio's ⋯ menu, under Article, Set stage is now Change stage. In the Export check, Clear allowlist is now Clear approved word differences.

Advanced and help

Check speech recognition

The account menu's Advanced choices now include Speech recognition. It shows download and checking progress, what is available on this device, and why speech recognition may be unavailable.

The page opens with the answer: a few tinted sentences saying what this machine will do, such as Marking the reading: KB-Whisper tiny, on the processor and This machine has no graphics adapter, followed by Expected wait after Stop with the estimate for each candidate. The expected wait itself was corrected from about 3.4 times the recording to about 1.2 times, after the two figures behind it were measured on the same recording. The tables sit in sections you can open: Models, Engines, and what they report, Deployment checks, What this device remembers and Try the realtime recogniser. Server checks that passed are folded away with the browser details, so the ones that failed are easier to find, and a failed deployment check opens its section by itself, so it is never hidden behind a closed heading.

Once the installation has finished, choose Init model to start speech recognition on this device. The recording controls then appear: choose Record, or Play a file, set the Language — it starts on the language the app is shown in — or choose Detect, and watch the words appear. After stopping, Decode the take shows what the finished recording was heard as, and Save the take downloads the recording to your computer as an audio file. These controls let you try speech recognition without recording into a chapter.

If installation checks failed, Retry verification lets you try again after the cause has been corrected. It reloads the page and clears the saved failures. It keeps the downloaded files and successful checks.

See files kept on the device

Advanced also includes Private file system. It lists stored files, their sizes and whether the expected download is complete. You can refresh the list, delete a file or choose Delete everything, with a confirmation before deletion.

Deleting speech recognition files means they will need to be downloaded again. This page explains that signing out or clearing the usual browser cache does not remove these files.

Reproduce queue states on demand

Administrators and testers get two rows in the account menu under Advanced. They exist to reproduce the queue states described under Final sync and the recording queue on demand, so a problem report can be checked against a known state. They are troubleshooting instruments, not part of the recording workflow.

• Queue three unsent recordings puts three silent test recordings on this device as though they had been recorded with no connection. The message reads Queued 3 unsent recordings on this device. Open a studio to see what the footer says about them, and remove them again under Advanced → Local recordings. The server refuses them if they ever arrive.
• Hold the final sync parks every queued transcription on every page of this site until you choose Release the final sync or close the tab. While held, the status bar shows what it says about recordings that are waiting without any of them having failed.

Clearer guides

The guides have been corrected against the controls this release moved or renamed, and use simpler wording. They explain the production card's Production roles and Production details choices, the panel rail that now opens Reading settings, the floating status bar, the current recording controls, the read-only content type, and which editing instructions return after reopening a chapter.

Read the Overview, Editor's guide, Narrator's guide, Reviewer's guide, Lead's guide, What a producer manages or Administrator's guide for help with your role.

Behind the scenes

You do not need to learn new controls for these changes.

• A reference recording ships with the app. A 20-second recording of continuous Swedish narration lets an install time its own speech recognition. It is kept out of the offline snapshot so that teams without local speech recognition do not carry it.
• Background jobs support several kinds of persistent work. They coordinate between tabs with a browser lock, hand unfinished work to another page and cancel their active worker when a job is removed.
• The Studio reader was divided into smaller focused modules. Document rendering, reading position, search highlights, sentence menus, annotations, word selection and status painting have separate owners. This reduces the chance that a change to one reading tool disturbs another.