Speak an instruction or listen to an answer¶

Speech offers two separate ways to use HiveAI: turn a recording into words you can review, and generate a voice reading of an eligible saved answer or proposal. Recording alone does not submit an instruction. Listening to a proposal does not approve it.

Availability: This guide follows the intended 06 Final Screens journey. The current account chooser still marks HiveAI Not ready yet; a working browser microphone and playback journey has not been verified. Speech requires Advanced access, organization permission, available credits and a ready speech service. The service determines the languages and voices actually available. Figma's example language, voice and prices are illustrations.

For typing, sources, choices and saved-request recovery, use the conversation guide. Start with HiveAI access if you cannot enter the workspace.

Before starting¶

• Confirm the organization and conversation. A recording belongs to that context and to the person who starts it.
• Choose a quiet place where you can speak without exposing confidential information to others.
• Read the displayed recording or spoken-answer limit before starting paid processing. These have separate usage from the conversation's own work.
• Keep the keyboard available. You can type a new instruction or correct a transcript when speech is unsuitable.
• If an existing recording is still active, finish or cancel it before starting another in the same signed-in session.

The current recording contract accepts up to two minutes of audio. The client must prepare the supported format; it should not require an ordinary user to choose codecs or upload technical fragments. Browser permission prompts, microphone selection and denied-permission recovery still require client verification.

Record and review an instruction¶

1. In the intended conversation, use the microphone control to open Speak your next question.
2. Review the recording disclosure and available limit, then choose Start speaking. Use keyboard returns to typed input.
3. Speak the instruction clearly. Include names, quantities and dates explicitly.
4. Choose Finish recording. The Preparing your transcript state means the recording is being processed and its language detected. No instruction has been submitted yet.
5. When Review your voice instruction appears, read every word. Check names, numbers, dates, units, technical terms and negation such as “do not”.
6. Correct any words that change your meaning. The reviewed words are the instruction that will be submitted.
7. Check the recording cost. A usable transcript requires its processing charge to be confirmed. A cost still being checked must not be presented as zero.
8. Choose Send instruction once. Look for the saved-request confirmation and then follow the request's progress in the conversation.

Your instruction is saved confirms acceptance. It does not mean that the requested business change has finished. An answer may still need a source check, clarification, proposal review or a protected step.

If the words need correction¶

Correction needed means the service could not confidently resolve a usable instruction. Enter the intended words and use Review corrected words before sending. The correction becomes the submitted text while the original detection evidence remains distinct.

Language detection is automatic. Choosing a voice for an answer does not change how the original recording was detected. Do not keep resubmitting uncertain words in the hope that the system will infer a different person or action.

Corrected instructions have a 16,000-character limit. If the instruction is longer, shorten it and separate the work into clear requests. Changing the words after submission requires a new deliberate instruction; it does not rewrite the saved request.

Reply to a current choice or proposal¶

Speech can be bound to the current pending request. Always read the current choices or proposal before activating its listening control.

For a numbered choice, the implemented resolver accepts a single supported ordinal, such as 2, option 3 or second. Vague replies such as “the other one”, “yes” or a person's name are not established selection commands. Use the visible choice control if the intended option cannot be expressed unambiguously.

For a proposal, the implemented voice decision requires a fresh listening action on that exact proposal and the reviewed word approve or reject. A general “yes” is not enough. Verify the proposal's scope before deciding. A recording from an older proposal cannot approve its replacement.

If the choices, proposal, workspace access or conversation context changed, refresh the current request and review it again. Refreshing does not renew an expired proposal. A protected step still needs its own secure completion; saying “done” does not complete it.

These continuation rules are source-backed behavior for client implementation. Their complete browser journey, including accessible controls and spoken confirmation, is still awaiting verification.

Listen to a written answer¶

1. Read the saved written answer or current proposal, including its sources and any limits.
2. Use the intended spoken-answer option when it is available.
3. Review Will speak, the available language and voice, the speaking speed and the separate Voice limit.
4. Choose Prepare spoken answer to request generation.
5. Wait for Ready to play. The written answer remains available while speech is prepared.
6. Choose Play spoken answer where the delivered client offers it.
7. Use Stop playback to silence the device when needed.

The voice is generated by AI. Check the written content for precise names, figures and instructions. A voice language selection does not promise translation of the saved answer.

An eligible answer must still be available and match the current saved response. The current service supports suitable completed or partly completed answers and a current proposal; it does not speak arbitrary text from the microphone or every intermediate progress message. Long or unavailable responses may not be eligible.

The mobile Ready to play example needs a visible-play-control review before that exact mobile journey can be accepted. This guide does not treat the existence of a design state as a verified working player.

Four different stop actions¶

Once the reviewed instruction has been submitted, manage its business request through the conversation's stop and recovery controls. Cancelling its old recording is no longer the applicable action. Stopping a ready audio player also does not refund the speech generation already performed.

Checking whether voice preparation stopped means the result is not yet confirmed. Read the written answer and return to the same voice request for its final status.

Understand speech costs¶

The three stages have distinct accounting:

The displayed maximum is the approved ceiling for that speech job. Used credits and released credits become known after processing is reconciled. Released credit is unused reserved credit made available again; it does not promise a cash refund.

A changed rate or limit requires a new disclosure before processing proceeds. Do not rely on the example amounts in the designs as a tariff.

A failed or unavailable audio result can still have measured processing usage. Conversely, work confirmed as never dispatched settles at zero. When usage is unknown, the system keeps that uncertainty and attempts recovery against the original job. It does not automatically repeat recognition or speech generation to obtain another result.

A later cost receipt may resolve the charge without recovering the transcript or audio. Check both the result and its cost; one becoming known does not prove the other is available.

Recover without duplicating work¶

The backend makes raw recordings and generated audio short-lived: their ten-minute lifetime starts when each recording or output record is created. That is not ten minutes from first playback. Cleanup is asynchronous, so an expiry is an access limit rather than proof that physical deletion happened at that instant.

Audio expiry does not by itself delete the submitted words, conversation history or accounting records. Those follow their own retention and recovery rules. The data lifecycle guide explains exports, member removal, hold minimization and cleanup evidence. Do not use speech as an archive of evidence.

For managers, administrators and support¶

Managers should review the written result and source record before relying on a spoken summary. Executives evaluating adoption should distinguish useful completed work from recordings, generated audio and unresolved usage; this guide makes no measured accuracy or productivity claim.

Administrators should confirm Advanced access, permissions, credit availability and current service capabilities before telling staff speech is available. Recognition languages and output voices are different inventories. The current browser rollout still requires verification.

For support, collect the organization, approximate time, visible status, conversation or request reference, and recording or voice-job reference if shown. Explain which stage failed and whether an instruction was saved. Share confidential transcripts or recordings only through the organization's approved support process.

Return to conversations or use the conversation recovery guide.