Separate a broken device from a broken path, then document the workaround and the request.

Why a VPN or network filter can break voice input comes down to one fact: dictation is a live round trip, not a file upload. While you speak, the page or app has to keep an exchange open with a transcription service, and that exchange continues for as long as you keep talking. Anything that delays, blocks, or interrupts the exchange will look like a microphone fault. The device can be perfect and the level meter can dance while the round trip fails behind the scenes. That is why the symptom is often no text rather than bad text, and why the fix usually lives in a setting you never touched.
Four mechanisms account for most failures. A filter can block the request outright, in which case the control starts and nothing arrives. A tunnel can interrupt an established stream, in which case text appears for a few seconds and then stops. A proxy can rewrite or buffer the connection, which produces delayed or partial text. A name resolution failure can prevent the round trip from starting at all. Each has a slightly different symptom, and the symptom is the first clue to which mechanism you are dealing with. Write the symptom down before you change a single setting. Write the mechanism next to the symptom before you write a theory.
A tunnel is not inherently hostile to dictation; it changes the path and the exit point. If the exit point is far away, the round trip becomes longer, and a service that expects a prompt response may time out. If the exit point is in a region the service treats differently, requests may be refused. If the tunnel reconnects mid-session, the stream breaks even though the page never noticed. These are path properties, not device properties, which is why the same headset works before the tunnel connects and fails immediately after it does. The region matters as much as the distance when a service decides what to accept.
Network filters exist for good reasons, and the goal is not to defeat them. The goal is to identify which rule is interfering and give the owner the evidence to adjust it. Start from the assumption that the filter is doing its job and your dictation is hitting a rule that was never designed with live speech in mind. That framing changes the conversation from a complaint about a broken tool into a request for a specific exception, which is far more likely to be granted quickly and correctly. A specific exception is also easier to review when someone asks why it exists, and easier to remove when the rule is no longer needed.
The hotspot test is the single most useful symptom check. Connect the same device to a phone hotspot, run the fixed sentence, and compare with the usual network. If the hotspot passes and the usual network fails, the device and the page are innocent. If both fail, the network is not the differentiator and you should look elsewhere. Do the test in both directions, because a hotspot that also passes through a tunnel is not a clean control. Note the time of day as well, since some filters behave differently under load. A hotspot that shares the same tunnel is a weak control, so note what it bypasses.
Timing splits the two main network failures. An immediate stop, where the control reacts and nothing ever arrives, usually means a request was refused before any stream began. A delayed stop after twenty or thirty seconds usually means a stream started and was then interrupted or timed out. Record the seconds with a stopwatch rather than describing the delay as a moment. The number is what an owner needs to tell a block from a timeout, and it is the difference between two very different fixes. Write both numbers, because a stop at three seconds and a stop at thirty are different reports.
Look for the pattern across surfaces. If the same network fails for both a page and an installed app, the network is the common factor. If it fails only on the page, the page's request pattern may be the trigger, and the network may be treating one kind of connection differently. If it fails only after you sign in, the session may be routed differently. Three observations, written as a small table, usually tell you whether the network is a cause or merely a coincidence. A table of three rows is enough to move from opinion to evidence that someone else can check.
Partial text is a network symptom too, even though it feels like a quality problem. Words that arrive late, sentences that end early, or punctuation that appears in a burst all point to buffering or a stream that is being interrupted and resumed. Do not respond by speaking more slowly; that changes the input, not the path. Instead, repeat the same sentence twice and compare where the text stops. If both runs stop at the same delay, you have a path limit, and the note should say so with the timing attached. Two runs are enough; a third only adds patience to the cost.
Run the toggle test first because it is reversible in seconds. Disconnect the tunnel, wait for the network to settle, and repeat the fixed sentence. Then reconnect and repeat again. If the failure appears and disappears with the tunnel, you have a clean correlation. Do this twice so a single coincidence does not become a conclusion. Record the time of each toggle, because a filter can cache a refusal for a short period and make the second test look different from the first one. A cached refusal is the most common reason the second toggle looks wrong when it should not.
If your client supports split tunneling, route only the page through the tunnel or only around it, and test both directions. This is the most precise test available because it changes one route rather than the whole connection. If dictation works when the page bypasses the tunnel, the tunnel is the interfering element; if it works only when the page uses the tunnel, the direct path is the problem. Write which direction you tested, since the two directions answer different questions and are easy to confuse later. One direction is not enough, because the two directions answer different questions about the same path.
Check name resolution and proxy settings next. A page that cannot resolve the transcription host will fail immediately with no visible error, which looks exactly like a blocked permission. Test resolution with a simple lookup tool outside the page, and compare the answer with the one you get on a different network. Then check whether a proxy is configured and whether it applies to the page's traffic. One change at a time, and write the before and after values rather than a general impression. Record the resolver you used, since two tools can disagree about the same name on the same day.
Finish with an alternate path that does not share the suspect elements: a different connection type, a different exit, or a different device on the same connection. If the alternate path works, you have narrowed the fault to a specific element rather than to the network as a whole. Keep a short log with four columns: test, path, result, seconds. Four columns are enough to show an owner exactly where the behavior changes, and they prevent the conversation from restarting every time a different person joins the thread. The log is the artifact; the conversation is only the delivery mechanism for it.
Build a packet for the network owner with the test table, the times, the exact hostnames involved, and the tool versions. Add the sentence you spoke and whether text arrived. Ask a specific question: can this path be allowed for live dictation, or is there an approved alternative exit? A packet that names the behavior and the requested change is a work order; a message that says the tunnel breaks my microphone is a conversation. The difference in response time is usually measured in days rather than minutes. Days, not minutes, is the honest unit for a path change request in most organizations.
Do not disable a filter permanently to make dictation work. A permanent bypass solves your afternoon and creates a policy problem that someone else has to find later. If you need a temporary measure while a request is reviewed, document it, set an end date, and tell the person who owns the policy. The goal is a supported path, not a quiet exception that disappears when you change desks. Write the end date in the same note as the workaround so it cannot be forgotten. An exception nobody can find is an exception that will be removed by accident six months later.
While the path is being fixed, keep drafting. Voiskey is an AI voice typing app for macOS and Windows, and it inserts at the insertion point in any app or text field after you press a shortcut and speak. Because it does not go through the page at all, testing it is a fast way to see whether the block is specific to the page route or applies to dictation in general. It removes filler words, stutters, and false starts and adapts punctuation, formatting, and tone to the task. That comparison is also the cleanest evidence that the fault is not the microphone itself.
Use the comparison to decide what to report. If the page route fails and the system route works, the report should say the page route is blocked on this path, not that dictation is broken everywhere. You can download Voiskey for macOS and Windows and keep the system route as the documented fallback for filtered networks, with a note that it is a workaround until the path is allowed. That distinction keeps the ticket narrow and the work moving at the same time. The fallback is not an excuse to skip the report; it is what lets the report stay accurate.
Request the change in the terms the network team uses: destination, protocol, direction, and business reason. Attach the test table and the exact times. Avoid asking for a blanket exception; ask for the narrowest rule that lets live dictation work. A narrow request is easier to approve and easier to audit later. If the answer is no, ask for the approved alternative: a different exit, a different port, or a scheduled window. A documented no is more useful than an undocumented workaround that nobody can explain. Ask for a review date in the same message so the request does not age silently.
Expect the path to change when you travel. Hotel networks, conference networks, and mobile hotspots all apply different rules, and a path that works at your desk may fail in a meeting room across town. Keep a short travel checklist: test the sentence on arrival, note the network name, and switch to the fallback if the first attempt fails. This takes two minutes and prevents the worst version of the problem, which is discovering it five minutes before you need to speak. A travel kit is small: one sentence, one hotspot, one note file, and the name of the fallback you trust.
Share what you learn in one place. A single table of network names and results is enough: network, date, result, action. After a few months, the table tells new colleagues which networks to avoid for dictation and which fallback to use. It also gives the network team a pattern instead of a series of one-off complaints. Knowledge about paths is perishable, so date every row and review the table when the office network changes or a new tunnel is introduced. A dated table is the difference between institutional memory and a rumor that costs an afternoon every month. Two minutes at the end of an incident is cheaper than an hour of rediscovery later.
Close the loop with a short review. If the path was allowed, confirm that the failure is gone for a week and then retire the workaround. If it was not, keep the fallback documented and revisit the request when the policy is next reviewed. Voiskey supports more than one hundred languages and translation, turns spoken requests into AI-ready prompts with file names and paths, and follows a speak, polish, type, edit rhythm. You can download Voiskey for Mac and Windows and make the fallback a deliberate part of the plan. A retired workaround is a small win worth recording in the same table.
The meter only shows that the device is capturing. Live dictation also needs an uninterrupted exchange with a transcription service, and a filter or tunnel can block, delay, or cut that exchange. The device looks fine while the round trip fails. Testing the same sentence through Voiskey shows whether text can still reach the field another way.
If the same device and sentence pass on a hotspot but fail on the usual path, the device and the page are fine and the path is the suspect. If both fail, the path is not the differentiator. Either result narrows the search before you spend time on a system-level route such as Voiskey.
Run the toggle test first: disconnect the tunnel, repeat the sentence, reconnect, and repeat again. Do it twice so one coincidence does not become a conclusion. If the failure follows the tunnel, record the times and then test a split route. If the failure does not follow the tunnel, switch to Voiskey and keep drafting while you look elsewhere.
Voiskey works at the system level and inserts at the insertion point in any app or text field after a shortcut and speech, so it does not depend on the page route. If it keeps working while the page fails, the block is specific to that route. If both fail, the path is affecting dictation more broadly.

Download Voiskey for Mac and Windows and type by speaking in the app you already use.
Use it across macOS, Windows, iOS, and Android