Transcription & floating subtitles

Start and stop

1. In Recording Settings, confirm the input device, select a registered recognition model and set the source language. Initial loading can take time; wait for status feedback.
2. Click Start or use the configured hotkey. Speak a short test and inspect interim and final text. Interim text can change; use finalized passages for review and minutes.
3. Click Stop and wait for finalization before inspecting the meeting. Processing in progress does not mean saving is complete.

12

1. 1Interim recognition appears here.
2. 2Translations are grouped by language and segment.

Stop recording: wait for finalization, then check the record

These stop protections require Windows r514 or later. Check the version actually installed.

1. Before starting, check the input device and recognition language, then speak a short test to confirm that audio arrives. When changing the target text field during recording, confirm the new location first. Automatic typing for a selected voice remains unavailable; see the limits of the selected-voice trial.
2. Click Stop once and wait for saving and processing to finish. On a normal stop, audio already received and converted is finalized even when the last part is shorter than the usual segment. A pause in new subtitles is not a reason to press Start repeatedly.
3. If finalization fails or a device-error or audio-backlog notice appears, return to the main window and check the task status, then inspect saved recordings in Session browser. An unconfirmed recording is not a completed success. Listen to the original audio and check what was retained before explicitly retrying or importing saved audio for transcription. Do not treat audio on either side of a gap as a continuous recording, and do not press Start repeatedly.
4. With cloud recognition, check local recording saved and cloud recognition completed separately. A saved local recording does not mean the service returned the final words; a very short ending may have no cloud subtitle. Listen to the saved audio before deciding whether to transcribe it again.

Until stopping, saving and processing are confirmed, keep checking the current task notice. After an error, inspect saved content first to avoid duplicate recordings or a missed ending.

Source, target and interface languages

Requires the Windows r512 or later installer preview. Earlier previews do not include this source-language binding fix. Check the three choices separately, then try a short recording.

1. Under Recording settings, select a recognition model and check its source language. For SenseVoice, Auto uses the model's actual automatic recognition; it is no longer treated as Chinese. If you expect mainly Chinese speech, explicitly select Chinese. Review the choice for another language or task.
2. When you need translated text, select a separate Translation target. For Chinese speech translated into English, set “Chinese → English.” Record one sentence and check the original before reviewing the translation.
3. To inspect the translation request, expand Terminology and translation context in the main recording translation settings. Enter a sample sentence and select Check glossary and request (no model call). The check uses that page's current recognition language, backend and primary translation target. The sample remains source text; do not replace it with a language name.
4. After changing the source language, recognition backend, primary target or translation model, run the check again. The old preview is cleared, and a late reply cannot replace the new selection. The check does not recognize audio or run a translation model. Use a short recording to assess the actual result.

Source context is a separate optional setting: the default 0 sends no preceding text, and it applies only after you enable and save it. See interpretation languages and source context for segmented and saved-audio translation. For interpretation, also check the source and target selected in the workbench for that task.

Recognition backends and files

In Setup → Local model library, choose Speech recognition as the use, then choose a model family. Register a SenseVoice ONNX folder, the complete Fun-ASR-Nano-2512 folder, or one whisper.cpp GGML ggml-*.bin file. Give each size or quantization a distinct name, then select it from the library when using the app. FunASR Nano root model.pt contains audio and language-model weights; retain the matching Qwen3-0.6B configuration and tokenizer. Original Whisper safetensors, PyTorch or GGUF files cannot replace GGML. SenseVoice and FunASR Nano recognize recording segments; Whisper primarily provides file transcription, not a promise of native streaming or low latency.

After registration, choose its name in Model used for transcription at the top of the library, check its files, then expand Try transcribing a recording. To get files, use Prepare download beside the selected family or expand Download models and components below the library. Review the source and destination disk before explicitly starting a download. Registration changes the draft; finish and save Setup to keep it.

Each recognition backend needs its runtime components and model files. If a component is missing, follow the message to supply it, then check the model again. Choosing a backend does not install it.

iFLYTEK realtime ASR: configure once, then select

These iFLYTEK connection steps require the Windows r512 or later installer preview. The screenshots show the compiled app with example settings; they do not establish live-account recognition quality.

1. In the iFLYTEK console, enable Realtime ASR Large Model for your application. In Boundless Flow, open Setup → API → iFLYTEK and enable its card.
2. Enter wss://office-api-ast-dx.iflyaisol.com/ast/communicate/v1. Save App ID → API Key → API Secret from that product console, then choose a named recognition mode and finish Setup. The two keys are separate fields.
3. Reopen the card to check the saved state. App ID remains visible. Each saved key input stays empty and shows Configured · Leave blank to keep; use that field’s Clear key button only when you intend to remove it.

12345

1. 1Enter the complete endpoint for realtime ASR Large Model; classic RTASR/IAT endpoints are not interchangeable.
2. 2Save the public App ID from the same product console.
3. 3Enter API Key here. After saving, an empty input with Configured retains the saved key.
4. 4Save API Secret separately. Clear key affects only this field.
5. 5Select a named mode; no protocol ID needs to be typed. Multilingual requires separate account entitlement.

In Recording settings, choose Realtime API → iFLYTEK · Realtime ASR Large Model. The saved mode is shown by name and is read only. Authentication is configured only in Setup; use Manage service when it changes.

123

1. 1Select the saved iFLYTEK service. Connection credentials are not entered again.
2. 2The saved mode is shown by name and cannot be edited here. Use Manage service to change it in Setup.
3. 3In Chinese/English/dialect mode, leave this input empty and set the recording source to Auto. Gray zh, en is only a placeholder.

Classic RTASR/IAT endpoints and previous raw model values remain visible as a previous configuration. They are never replaced automatically. To migrate, deliberately change the endpoint, select the current mode and save, after checking access for the same application.

Start streams microphone audio to iFLYTEK and may incur charges. A source-language mismatch blocks capture before it starts. Stop sends the remaining audio and waits for the server’s final confirmation. During finalization, wait before starting again. On an error or disconnection, the app reports the failure without automatically reconnecting or resending audio. Confirmed text remains saved; interim text is not promoted to a confirmed result. With normal storage, the meeting session retains the original audio for retranscription, independently of the option to save another WAV copy. If a storage warning appears, check available disk space first; retention cannot be guaranteed when storage fails.

Account access, network, audio and language affect results. Test a short recording and review names and numbers; no fixed accuracy or latency is promised. iFLYTEK protocol, entitlement and language list →

Subtitle display and file transcription

Multilingual subtitles align by segment. Font size changes reading on this page without pausing translation. Focus Reading provides a clearer reading area. Browsing history pauses following; return to the latest content to watch new segments. A visible-row limit affects display, not the complete translation history; use translation recovery and meeting export for saved text.

With translation enabled, complete sentences can be translated first. An unfinished price, number or condition may wait for the following words, balancing meaning and timely updates. Held text is processed when the recording segment ends. If a short utterance has no translation yet, check the source transcript and whether recording is still in progress, then wait for the next words or the segment to end.

For local Fun-ASR-Nano transcription of a long file, audio is processed in non-overlapping clips of up to 16 seconds, including a shorter final clip. Each clip keeps its own text; check repeated speech against the recording. Follow clip progress and wait for the task to stop after cancelling. If model generation does not finish normally, the task reports an error. Cancellation or failure is not a complete transcript. Keep the original recording and review the message before retrying.

Use Audio transcription to select a file. Choose the recognition backend and language appropriate for it. Review names, numbers, terms and speaker boundaries after processing. Native offline transcription needs the matching native-stt component supplied with the app.

12

1. 1Copy the source transcript.
2. 2Export the transcript as a text file.

Floating transcription

Floating Transcription opens compact controls and subtitles. It uses the same app input and hotkey configuration. Verify start/stop in the main workspace before testing the floating view. A floating display does not imply a separate recognition model has started.

The floating control has Start, Stop and Return to main window buttons. Watch the waveform while speaking and read the source and translation in the subtitle window. Check the control or main window for recording status.

12

1. 1Start transcription with the microphone button.
2. 2Return to the main window for the full transcript.

Local text cleanup preview

Local text preview is included in the r492 Windows installation preview. Enhanced live dictation delivery remains disabled; this section lets you try local text rules without typing into another application.

Open Recording settings → Output Mode → Text cleanup. Start with a short Chinese sample so you can inspect every change. You do not need a text model for local preview. To clean up raw text or a saved transcript segment, use the text cleanup workspace below.

Try one short sample

1. Choose Smart cleanup · local repair and format, then Plain text.
2. Expand Dictionary and protected terms. Add 齐码蓝 → ZimaBlueAI with scope All templates. Add 无界音流 → Boundless Flow with scope Email. Each Heard value must match the text exactly. The second rule applies only when the output template is Email.
3. Optionally put 一号项目 on its own line under Protected terms. Click Apply text rules. A warning about overlap or conflicting rules must be fixed before you continue.
4. Expand Local text preview · compare original and cleaned text. Type 预算三千不对三万，齐码蓝负责一号项目。, then click Preview local cleanup.
5. Compare the original with the preview. In this supported example, the explicit correction changes 三千不对三万 to 三万, and the first dictionary rule writes ZimaBlueAI. Check the displayed reasons. If protection or uncertainty notices appear, inspect the retained wording instead of forcing a change.
6. Only after checking, click Copy preview text and paste it where you need it. Preview does not type at the cursor or automatically save the sample.

1234

1. 1Choose Smart cleanup and an output template. Live dictation delivery remains disabled.
2. 2Enter exact dictionary rules and their scopes, then apply valid rules.
3. 3Enter a sample, then click Preview local cleanup.
4. 4Compare the original, preview and reasons before explicitly copying.

Use rules without losing the original meaning

Keep up to 64 dictionary rules and 32 protected terms, each 1 to 80 characters. Rules with overlapping Heard phrases must use separate, non-overlapping scopes or be combined. A protected term cannot overlap a Heard phrase that asks to replace it. Invalid settings are kept for repair and are not silently replaced with defaults. Correct the rule and apply it; then follow the main configuration save flow.

Uncertainty such as 不知道, quoted speech and conditional wording are not treated as a definite self-correction. Check key values and negative wording against the original, especially expressions such as “do not” or “not required” that change meaning.

Code scope keeps identifiers protected and does not call a review model. Smart-mode templates only arrange supplied text. Changing modes keeps your rule draft; during a speech task, settings and preview controls are locked until that task finishes.

Model review is a separate, optional setting

Smart mode offers Conservative text-model review in smart mode (optional). It reuses the configured text model for the delivery path; when a review exceeds one second or fails, that path keeps the local cleanup. A configured external service would receive the confirmed text. Live delivery is still disabled, and local preview never calls a model, including when this checkbox is selected. Code scope does not use online review in any mode.

Review dates, amounts and names together with their negation. “Do not pay today; pay tomorrow” must not become “Pay today; do not pay tomorrow.” If meaning changes, use the original or the local cleanup you have checked. Model review cannot guarantee semantic equivalence. Enhanced live delivery remains disabled; local preview does not call a model.

Local preview runs only when you click its button. It does not read your screen or clipboard and does not automatically persist sample text. Copying happens only after you click Copy preview text. These settings and examples do not establish live typing quality or superiority to another product.

Turn a transcript copy into usable text

The text cleanup workspace is included in the r492 Windows installation preview. It reuses your applied dictionary and protected terms; no additional model or API configuration is needed.

Start with the text you are reading

1. In Raw text, click Clean up text to bring the current text into the workspace. To process only part of it, select that passage in the text box before clicking.
2. Alternatively, open Session browser → Transcript preview and click Clean up text on a transcript segment. A selection takes priority; otherwise the full segment is used. Short segments start compact. Scroll inside a long text box, or drag its lower-right corner to make it taller.
3. In the Text cleanup workspace, check the source name and the left-hand Original copy · editable. You can edit this copy while the original transcript and recording remain unchanged.
4. Check Mode for this copy and Template for this copy, then click Clean up text inside the workspace. Loading a copy does not start processing.
5. Compare the right-hand Cleaned draft · review before use with the original. Expand Individual changes to see each before/after pair and reason. Pay particular attention to names, numbers, negative wording and code.
6. When satisfied, click Copy cleaned draft and paste it yourself where needed. The result does not automatically replace a session, type into another application or save as a report.

Temporary options leave central settings unchanged

Template for this copy offers Plain text, Chat, Email, List and Code. Mode and template apply only to this copy. Your dictionary and protected terms stay in one place; click Manage central text rules to return to them. Follow the rule setup steps above, apply valid rules, then return to process the copy.

Undo a cleanup or switch to another passage

• Undo cleanup restores the cleaned draft to the current original copy and clears the changes. It does not undo your edits to the copy or alter the source transcript. Clear cleaned draft clears only the current result and changes.
• If a copy is already open, bringing text from elsewhere first asks you to choose. Only Use new text in this workspace replaces it. Choose Keep current copy to continue your current work.
• Editing the copy, temporary options or central rules invalidates the previous result. Click Clean up text again. Invalid central settings must be repaired; the workspace does not silently switch to original-wording mode and proceed.
• Collapsing or closing the workspace retains the copy during the current app run. Late results from an earlier task cannot overwrite it. Drafts are not automatically saved to disk; copy any text you need before quitting the app.

Each request supports up to 65,536 Unicode characters; an emoji counts as one. Over-limit text is not processed or truncated. Return to the source and select a smaller passage. Long change lists have page controls. If the workspace reports busy, timeout or capacity limits, your copy is kept; adjust as directed and retry.

This path uses local text rules only, with zero cloud API calls and zero API fees. It does not upload text or call the centrally configured text-review model. Review the recording, verify facts and decide which changes to use yourself.

Read the floating transcription start, level and stop states

These floating-status improvements are included in the r492 Windows installation preview. This section covers the ordinary transcription Mini opened with Floating Transcription in the top bar. Enhanced dictation delivery remains disabled.

1. Open Floating Transcription and click the microphone to start. Read the status before deciding whether recording has begun. Repeated button clicks or hotkeys do not queue multiple starts.
2. While speaking, watch the audio level and subtitles. The level comes from received audio frames, with no simulated pulse. A silent frame can have a level of zero.
3. During preparation, use Cancel preparation to cancel that preparation only. During recording or file transcription, click Stop transcription and wait for the state to change. The Stop control remains available for file transcription.
4. Return to main window changes the view only; the current background task continues. To end a task, click Stop and wait for finalization.

If a start or stop remains unconfirmed for more than 12 seconds, Refresh recording status appears. Refresh first, then act on the latest state; the app does not automatically restart. Even after a refresh shows idle, Start stays disabled while an earlier operation request is still awaiting its reply, preventing duplicate recording. Retry Refresh if it fails. Errors use fixed explanations rather than exposing sensitive details from service responses.

If no text appears

Check for speech on the selected input, microphone permission, model inspection feedback and recording state. For cloud recognition, also check account permissions and ASR configuration. If auxiliary speaker analysis fails, inspect subtitle results separately rather than assuming transcription failed.

Keep your edits while a file is being transcribed

From r496, manual edits in Original text are preserved while file transcription continues. Later recognition results appear separately under Compare recognition draft. Voice Studio shares this protection when using original text.

1. When the preserved-edits notice appears, expand Compare recognition draft. Your editable draft stays above; the recognition draft below is read-only. Scroll long results or resize the box.
2. Keep current draft folds the comparison without replacing text. You can reopen it to inspect later recognition updates.
3. Choose Use recognition draft only when you want to replace your current draft. Undo apply restores the previous draft. If you type again, undo is disabled to protect those newer edits; the earlier text remains available under View previous draft.
4. Copy or export Original text to retain your current draft. These actions do not change saved meetings, transcripts or recordings. Edit and save the meeting transcript separately in the session browser.

The current, recognition and previous drafts belong to this interface session. Starting another recording or file transcription, or clearing text, resets these comparison drafts; copy or export anything you need first. Without manual edits, recognized text continues to update normally.

Read a transcript and follow original audio

In Session browser → Transcript preview, each passage is displayed as readable text. Time, speaker, Clean up text and Edit sit together; short sentences do not reserve a large editor. These changes are included from r492.

1. Click Edit to change the passage or its speaker name. Use Save text or the appropriate rename button to save. Closing the editor keeps the draft; Restore discards the local changes.
2. Click a passage time to listen from that passage. Pause, Stop and Continue passages control playback. Missing or unreadable recordings leave the transcript available.
3. Follow audio highlights each timestamped passage. Manual scrolling or filter/page navigation pauses following; editing does not get interrupted by auto-scroll. Turn Follow audio back on when ready.
4. Starting recording or transcription stops original-audio playback. Changes to the recording or transcript require a refresh before listening again.

Highlighting follows passage timestamps, not individual words. Continue passages moves to the next saved passage start, omitting unselected gaps. Live, unsaved recording does not provide this saved-audio playback. Audio stays local and is loaded only when requested.

Recover recent dictation results

If recognition finishes but the target field has no text, return to the main window and expand Recent dictation results. This retains accepted final segments from recordings with typing into the focused window enabled, for this app session. It does not run recognition again. For ordinary meetings or file transcription, use the original transcript and Session browser.

1. Select a timestamped segment and check its read-only Recognition draft. When translation is enabled, Translation separately shows the primary target language. You can copy recognition even if translation has not finished. Other target languages remain in the existing translation area.
2. Check the original target field before choosing Copy recognition or Copy translation. Return there and paste manually. Some text may already have been typed; avoid duplicates. Copied confirms the copy action, not successful typing.
3. New results do not replace your current selection. Use Refresh after a read error. If copying fails, the text stays selectable for manual copying.
4. Clear this list only clears these results. Original transcripts, saved sessions and recordings are unchanged. If a new result arrives during clearing, it is retained and you are asked to review first.

This temporary recovery list holds up to 64 segments, with additional text-capacity limits. A notice identifies omitted earlier or oversized content; check the original transcript or saved sessions. Save or export anything you need to retain. The list does not survive quitting the app. It only displays text and copies on request; it does not type again, restart recording or send Enter automatically.

Correct a term, then confirm whether to remember it

Open Clean up text from Original text or a transcript passage. In the original copy, select a misrecognized short term and replace it with the correct spelling. A Remember suggestion appears. Ordinary passage edits are not learned automatically. You can also choose Remember a term and enter the pair yourself.

1. Check Recognized as and Write as. Enter two different terms of 1–24 characters using letters, numbers or _ . / -. Do not enter sentences, spaces or purely numeric corrections.
2. A project transcript defaults to that project; see Project terms below. Unassigned text defaults to the workspace template. Central template rules are shared across projects, and Central · All templates requires an additional acknowledgement.
3. Only Confirm and remember saves the rule. Cancel does not save. Dismiss does not learn the pair and suppresses its repeated suggestion for this workspace lifetime. No suggestion is made during unfinished IME composition.
4. After a successful addition, Undo remembered term removes that rule only. It does not restore an old passage or clear other rules. If central settings have changed, review the dictionary first. Identical rules are not added twice.
5. Saving does not rewrite this draft or enable automatic typing. Original wording applies no rules. Explicitly choose Conservative rules or Explicit repair and format, then run Clean up text to review the effect. Resolve protected-term, overlap, capacity or unapplied-draft conflicts in Manage dictionary.

Save and undo contain the selected term pair, text settings and necessary local project identifiers, not this passage. They do not read another application or the clipboard. Central template rules are shared across projects; project-specific rules are described below. Quick undo retains only the most recent addition; manage other rules in the dictionary.

Remember project terms for explicit cleanup in that project

Assign a saved recording to a project in the library, then open Transcript → Clean up text for that recording. Project dictionary shows its actual project and term count. Ownership is checked by recording identity, not by a matching project name or the currently selected card.

1. Select and correct a short misrecognized term, or choose Remember a term. The default scope is Only project …. Confirm the project name and pair before saving. Central · Plain text and other central scopes are shared across projects; Central · All templates needs a separate acknowledgement.
2. Rules do not rewrite original transcripts, recordings or old reports. Explicitly choose Conservative rules or Explicit repair and format, then run cleanup and review it. Original wording still preserves the source exactly. This is a text-cleanup dictionary, not evidence that an acoustic model has learned the term.
3. Expand View and manage project terms to inspect or remove rules. Removal affects this project only. Restore removed term restores the latest removal; the term-confirmation area also offers undo for the most recent addition.
4. After moving an entry, archiving a project, or changing rules, refresh and review. Late output from an earlier project cannot replace this draft. Copying a cleaned draft also rechecks the project; a change retains your text and asks you to run cleanup again.

Unassigned recordings are explicitly identified and use central rules. Independent original text is not assigned a project by guesswork. A read failure or archived project does not silently substitute another project’s dictionary.

Each project supports up to 64 rules; the combined effective rule set is bounded too. Identical mappings are processed once. Conflicting central meanings, protected terms or overlapping rules require review. The current entry is a transcript already assigned to a project; bulk dictionary import and standalone management of an empty project are not available yet.

Import terms in batches and review change history

This dictionary interface requires the Windows r512 or later installer preview. The Mac r492 source ZIP does not include it.

1. For shared terms, open text settings and choose Dictionary import and history. For project-only terms, select the project in the library and choose Project dictionary and history. Empty projects can be prepared before recording. Apply or discard an existing text-rule draft first.
2. Paste terms or explicitly select a UTF-8 JSON, TSV or CSV file. Choose the matching format, then Preview import. A file may contain up to 256 candidate rows / 64 KiB. The saved dictionary still holds at most 64 rules.
3. Review New, Duplicate, Replace, Conflict and Invalid rows. Replacements are not selected automatically: check each replacement only after reviewing its previous spelling. Reduce your selection if it exceeds capacity or contains conflicting rules. Choose Adopt selected terms to save.
4. Open Change history to review the time and added/removed terms. Choose Revert this change…, review the prompt, then confirm. Later unrelated terms remain; if affected terms changed, refresh and review instead of overwriting them.
5. If a write cannot be confirmed, keep the pasted input and choose Refresh dictionary and history. Check the saved result before trying again. Closing this panel during saving does not cancel the existing write; reopen the same dictionary to review it.

Tabular columns are from, to, and optional scope; omit scope to use all. Other shared scopes are plain, chat, email, list and code. Project dictionaries use all within that project only. CSV quoted commas and UTF-8 BOM are supported.

from	to	scope
齐马蓝	齐码蓝	all

JSON may be a rule array, for example [{"from":"齐马蓝","to":"齐码蓝","scope":"all"}]. Import does not add protected terms; continue editing those in text settings. The original transcript and recording remain unchanged. These are cleanup rules, not ASR training.

History keeps up to 64 recent changes / 256 KiB. Earlier changes may become a baseline; actions before history was enabled are not reconstructed. This is local rule history, not a complete authenticated activity log or cross-device sync. Avoid overwriting new project dictionaries with older software.