Understand staff voice and spoken responses¶
Use this guide to understand what staff voice can do, how to review a spoken instruction and what to do when processing is interrupted. The backend behavior has been inspected. The complete staff voice screen journey and environment delivery remain unverified, so the steps below describe the workflow rather than a verified sequence of interface controls.
Before using voice¶
You need an authorized Platform staff identity and your own staff assistant conversation. Your current staff access and session context must still match the recording or response. Tenant membership alone is insufficient. Voice does not bypass the additional permissions required by a proposed action.
Check that speech is available for the task. Recognizing a recording and generating a spoken response are separate capabilities. Available languages and voices come from the server; a voice saved previously can become unavailable after the server's assets change.
Use the text workflow when speech is unavailable. Repeated recording attempts are not a remedy for missing server assets, changed access or a service outage.
Prepare a spoken request¶
- Decide whether you are starting a new instruction or answering a specific request for clarification or approval. A response to an existing request must remain attached to that exact pending question or proposal.
- Keep the recording brief, within two minutes, and state names, dates and intended changes clearly. Avoid unnecessary personal or confidential information.
- Finish the recording, review the current disclosed maximum recognition cost, and proceed only if you accept it. Wait for recognition and its usage result. Upload completion alone does not mean the words are ready to send.
- Read the text carefully. Check organization names, people, dates, numbers, negation and the action you intended.
- Correct any error before submitting. The public submission accepts corrected text up to 16,000 characters. If the recording expires or the request changes, return to the current request and start again deliberately.
- Submit only the text you have reviewed. Then follow the saved assistant request and any required clarification or approval.
Only one recording can be actively capturing in the same staff session context at a time. Finish or cancel that capture before starting another. Other processing states are not a guarantee that the entire assistant workflow has finished.
Targeted work needs particular care. The current new-recording submission does not carry an explicit target organization into the new assistant request. Do not assume that naming an organization aloud binds the request to that organization. Establish and verify the target through the staff assistant's supported workflow before approving any tenant-specific change.
When recognition needs correction¶
Automatic language detection evaluates the recording itself. Your selected playback voice is not an instruction to interpret the recording in that language.
| Result | What it means | What to do |
|---|---|---|
| Text ready | Recognition produced usable text and its measured usage was settled | Review every material detail before sending |
| Correction needed | Speech was uncertain, unsupported or not sufficiently detected | Enter the intended words accurately, or use a fresh text request |
| Processing | Recognition or accounting is still being resolved | Check the existing recording's status |
| Failed | No usable text is available from this attempt | Check its usage outcome; decide whether a fresh recording or text request is appropriate |
| Outcome unknown | The system cannot yet confirm the job's usage or result | Seek status recovery; do not assume it was free or resend an action blindly |
Mixed languages, very short speech, silence and low-confidence recognition can require correction. This is not a guarantee that every accent or language switch will be detected correctly. Human review remains necessary.
A cancelled recognition request can race with a result already produced. Inspect the returned state before deciding whether text is available or another attempt is needed.
Answer a clarification or proposal¶
A spoken clarification currently supports selecting an offered option by its number. Arbitrary form answers and protected steps still need the supported text or review workflow.
For a proposal decision, review the current change, acting staff member, affected organization, source details and deadline first. The supported spoken decisions are the explicit words “approve” and “reject”; do not depend on an indirect phrase being treated as agreement.
A spoken decision has a single-use confirmation associated with the exact pending proposal. Changed details, an expired proposal or an expired recording require a fresh review. Recognizing “approve” does not, by itself, execute a change: the reviewed transcript must be submitted and the action must pass its normal authority checks.
The stored confirmation timestamp does not establish that you actually heard or read the proposal. Take the review step yourself. Staff speech does not provide voice identity verification or replace a protected approval step.
Listen to a response¶
The backend can generate speech for an eligible current proposal preview or a completed provider answer. It does not offer arbitrary text-to-speech for everything in the conversation. A completed action receipt is not necessarily an eligible spoken answer.
- Review the current text and confirm that it is the response you want to hear.
- Select a currently available language and server voice. The supported staff speaking speed is 0.75 to 1.5 times normal speed.
- Review the current disclosed maximum cost before starting generation.
- Wait for generation, storage and usage settlement to complete.
- Listen while the audio is valid. Keep the written response available for exact names, figures and decisions.
A changed answer or proposal can make older audio unavailable even before its time window ends. Refresh the current text before requesting another spoken response. A saved voice preference is not a guarantee that the same voice will remain available. The spoken voice is generated by AI; it is not a recording of a staff member.
Listening to a proposal does not confirm it. Listening to an answer does not prove that a separate business workflow has finished.
Stop recording, speech or work¶
These are different actions:
| Intent | Effect and limit |
|---|---|
| Cancel a capture | Stops that recording workflow; saved temporary audio becomes eligible for cleanup |
| Cancel recognition | Requests cancellation; processing already dispatched may still have a measured or uncertain usage result |
| Stop a spoken response | Requests cancellation of generation; an already ready response is not erased by this request |
| Pause sound already playing | Requires the client player's own control; server cancellation does not prove local playback stopped |
| Cancel assistant work | Uses the separate staff assistant workflow and its cancellation limits |
After submission, cancelling the recording does not withdraw the assistant instruction. Stopping a voice response does not undo an approved action. See staff action cancellation and recovery.
Recover an interrupted attempt¶
Check the original recording, response or saved assistant request before starting another. A connection failure can happen after work was accepted.
Status checks may advance recovery and settle a known job. They do not automatically rerun recognition or generation, and a recovered usage receipt does not reconstruct lost words or audio.
If submission appears stuck after answering a proposal, inspect the assistant's current request and the source operation. The backend can accept a continuation before saving the link back to the recording; replay against an already changed proposal may then fail. Obtain help with that original attempt rather than repeating approval.
If progress remains open after failure, use the saved status or Support. The current input event stream has a known failure-termination gap, so an open progress connection is not proof that recognition is still running.
Understand costs¶
Recognition measures audio duration and includes language detection. Voice generation measures the approved response text. They use platform internal accounting, with separate reservations and measured settlement. A later assistant response or action can involve additional work under the staff assistant's rules.
A confirmed job that never started can release its reservation without usage. A job that ran can be charged even when its text or audio could not be delivered. An unknown amount must remain unknown until reconciliation establishes the result.
Do not quote a fixed price from this guide. The current server rate and disclosed maximum must be used for each job.
Audio expiry and saved records¶
The ten-minute window starts when the recording or output record is created, not when recognition finishes or playback begins. Slow processing uses part of that window.
Expired audio is unavailable through the governed read path. The background cleanup works in bounded batches on an hourly signal. Running work, storage failures, backlog or failed cleanup can delay physical deletion beyond expiry.
Audio cleanup does not remove saved original, corrected or submitted text, voice preferences, assistant requests or accounting evidence. Treat those records according to their applicable retention controls. This guide does not establish a universal deletion deadline for them.
Guidance for each reader¶
Staff and Support should record the original request identity, approximate time, visible status and the stage that failed. Avoid sending raw recordings or confidential transcripts through an unrelated support channel.
Managers and executives should evaluate whether voice improves access while retaining deliberate review, accurate outcomes and accountable internal costs. No productivity or financial saving has been measured here.
Tenant administrators, employees and external participants should use their organization's authorized assistant. A public explanation of staff voice does not grant access to Platform staff tools.
Designers, engineers and operators must finish the missing staff screen flow, qualify the delivery and reconcile the documented implementation limits before representing the experience as complete.