Field note
ChatGPT Voice Not Working: Firewall Checks
Text chat can work while ChatGPT Voice fails, so separate microphone permission, device audio, UDP port 3478 and company firewall policy before blaming the app.
Text chat can work while ChatGPT Voice fails, so separate microphone permission, device audio, UDP port 3478 and company firewall policy before blaming the app.
Symptoms
Typing works normally, but the voice button does nothing, the voice screen opens and immediately closes, ChatGPT says it cannot hear you, or the conversation connects and then drops after a few seconds. Sometimes it works on a phone on mobile data and fails on the same phone on office wifi. That last pattern is the one that tells you the network is involved.
Why voice is different from text
Text chat is ordinary HTTPS traffic on port 443 and passes through most business firewalls without anyone thinking about it. Real-time voice uses WebRTC. That means the app needs to reach STUN and TURN servers, which by convention listen on UDP port 3478, to work out a path for the audio stream. If UDP is blocked, WebRTC can fall back to TCP or TLS, but the fallback is slower, sometimes unavailable, and often the thing that fails silently on a locked-down network.
At the time of writing OpenAI's network recommendations in its Help Centre include allowing UDP traffic on port 3478 for voice, alongside the usual list of OpenAI domains. Check the current article for the exact domains, because they change.
Likely causes, most common first
- Microphone permission. The browser or operating system has denied ChatGPT access to the microphone, or the permission was granted to a different browser profile.
- Wrong or busy input device. A headset that is unplugged, a docking station that presents a second microphone, or Teams and Zoom holding the device.
- Firewall blocking UDP 3478. The most common cause when text works and voice does not, and the one that shows up as "works on hotspot".
- Web filtering or TLS inspection. A filter that allows chatgpt.com but blocks the media relay domains, or inspects the fallback connection and breaks it.
- VPN with UDP blocked. Many always-on VPN configurations pass only TCP. Voice then relies entirely on the fallback path.
- Plan, region or app version. Voice features vary by plan and roll out in stages. An old app version can lose voice after a server change.
Step-by-step fix
- Test the microphone outside ChatGPT. Record a voice note or use the operating system's sound settings. If nothing registers, the problem is the device, not ChatGPT.
- Check the site permission. In Chrome or Edge, click the icon by the address bar and confirm the microphone is allowed for chatgpt.com. On Windows and macOS, confirm the browser or app itself is allowed to use the microphone in privacy settings.
- Close other apps that may hold the microphone, then reload ChatGPT.
- Try a phone hotspot with the VPN off. If voice works there, the office network or VPN is the cause. Stop here and hand the note below to IT.
- If it fails on the hotspot too, update the app or browser, sign out and back in, and try a different browser.
- Confirm your plan includes the voice feature you are trying to use. Voice options differ between Free, Plus, Business and Enterprise.
What IT should check
- Outbound UDP 3478 from user subnets to OpenAI's voice endpoints, as listed in the current OpenAI network recommendations. Allow it as a specific rule, not a general UDP open.
- That the TCP and TLS fallback on port 443 to the same endpoints is not being broken by TLS inspection. Exempt the OpenAI domains from inspection if voice matters to the business.
- Web content filtering categories. A recently added block for "AI" or "streaming media" can catch the relay domains while leaving the main site reachable.
- VPN split-tunnel or UDP policy for the affected users.
- That the change is documented. If you hold Cyber Essentials, the firewall rule needs a business reason, an owner and a review date, the same as any other exception.
If the web filter is Defender for Business, the Defender web content filtering guide shows where those category blocks live. Whether voice should be enabled at all for a given team is a policy question, and an AI governance review is the place to settle it rather than a firewall ticket.
When to escalate to OpenAI
Contact OpenAI support when voice fails on a phone hotspot, with the microphone confirmed working, on an up-to-date app, on a plan that includes the feature, and with no incident on the status page. Include the device, operating system and app version, the plan, the exact behaviour, the time in UTC and whether text chat works at the same moment. Do not send firewall configurations or internal network diagrams; describe the test results instead.