A voice agent on a phone line and the same agent in a browser can share the same speech model while giving the model very different audio to process. The difference begins before the agent answers: the phone path may pass through carrier, SIP, and media-gateway stages, while the browser can establish a real-time session directly from the user’s device using WebRTC.

That makes voice agent SIP vs WebRTC an architecture decision. Use browser WebRTC when you control the calling surface and want richer audio interaction. Use SIP or PSTN when being reachable on an ordinary phone matters more than preserving the browser path. In either case, measure the complete route you will deploy instead of assigning a universal latency penalty to one channel.

Voice agent SIP vs WebRTC starts with the access path

A browser voice agent normally begins with local microphone capture, browser permission, signaling, and a WebRTC session. The W3C WebRTC specification describes the API as enabling media to be sent to and received from another browser or device implementing the required real-time protocols. That gives an application a direct starting point for the user’s microphone and speaker, although the actual session can still involve signaling services, ICE negotiation, relays, an SFU, or an application server.

A phone-based agent begins somewhere else. The caller enters through the public telephone network or a SIP trunk, and the application typically uses a telephony provider or gateway to expose that call to the agent. Twilio’s documented browser-to-phone pattern, for example, combines Twilio Client/WebRTC in the browser with Programmable Voice commands such as <Dial><Number> for PSTN or <Dial><Sip> for SIP. That is one provider’s architecture rather than a description of every telephony deployment.

The practical distinction is the number and location of media-interworking stages. A browser session can remain within a WebRTC media architecture. A phone call may need to cross between PSTN or SIP signaling, a carrier media network, a gateway, and the agent-facing audio interface. The resulting delay depends on carrier routing, gateway placement, buffering, codec conversion, and the rest of the application path.

That path difference leads to the first trade-off: browser access gives you more control over the media session, while phone access gives the caller a familiar endpoint. The audio format then determines what the agent receives.

Voice agent SIP vs WebRTC changes the available audio

Conventional telephone audio is narrowband. ITU-T Recommendation G.711 specifies a sampling rate of 8,000 samples per second. G.711 carries that audio as a 64 kbit/s bitstream, and the conventional telephone band runs roughly 300-3,400 Hz. These figures describe codec and nominal telephone-band characteristics. They do not describe every modern mobile or HD-voice call.

Twilio gives a concrete example of how this can appear at the application boundary. Its Media Streams documentation describes phone-call audio delivered as audio/x-mulaw at 8,000 Hz with one channel. Its SIP documentation lists G.711 μ-law and A-law as supported media codecs. Other carriers, trunks, and gateways may expose different formats or require transcoding.

WebRTC endpoints are required to implement Opus, PCMA, and PCMU under RFC 7874. That requirement establishes interoperability support. Actual codec use follows offer-and-answer negotiation, endpoint preferences, device capabilities, and network conditions.

Opus also supports several bandwidth modes. The 2015 RFC 7587 describes a wideband mode covering 0–8,000 Hz at a 16,000 Hz sampling rate and a fullband mode covering 0–20,000 Hz at a 48,000 Hz sampling rate. A browser session may negotiate a narrower mode, so browser audio can range from narrower speech modes to wideband or fullband operation depending on the session.

The advantage of a browser path is conditional: it can make more audio bandwidth available, but only when the session negotiates and preserves it end to end. The next question is what that difference means for an agent’s speech pipeline.

What the audio difference means for an agent

Speech recognition and turn detection work on the acoustic signal itself. Changing the channel bandwidth changes the information available to those downstream systems. A 2015 NIST/NTIA study found that speech intelligibility depends strongly on the noise environment, data rate, and audio bandwidth. The result supports treating the channel as an input condition for the agent rather than assuming the speech model sees the same signal on every route.

NIST’s 2015 technical report uses nominal categories that illustrate the distinction: narrowband is approximately 300 Hz to 3.5 kHz, wideband approximately 50 Hz to 7 kHz, and fullband approximately 20 Hz to 20 kHz. Those categories describe nominal bandwidths. Microphones, speakers, processing, negotiated codec mode, and network conditions can reduce the end-to-end result.

Research on automatic speech recognition also treats acoustic bandwidth as a material variable. A 2014 study on wideband telephone speech reported that ASR performance strongly depends on the speech bandwidth available to the system. That finding does not provide a universal word-error or turn-detection penalty for every voice agent. It does justify testing the chosen agent with the actual phone codec, browser devices, background noise, and conversational overlaps it will encounter.

For a phone agent, test whether names, numbers, addresses, accents, and interruptions remain reliable through the deployed carrier path. For a browser agent, test low-cost microphones, laptop speakers, mobile browsers, Bluetooth headsets, and permission failures. The channel decision should be made with those tests, because a nominal codec label does not describe the whole user experience.

Where SIP and PSTN add operational stages

A browser call can still have complex infrastructure, but the application team usually controls the session model more directly. It can choose how the browser joins the real-time room, how media is routed, how reconnects are handled, and where the agent subscribes to the audio track.

A phone call introduces boundaries owned by other systems. The call may be signaled through a carrier, mapped to a SIP endpoint, converted by a gateway, buffered for jitter, and exposed to the agent through a media interface. Each stage can affect setup time, mouth-to-ear delay, audio format, interruption behavior, and the point at which failures become visible.

That does not make PSTN unsuitable for agents. Phone access remains valuable when callers may not have a supported browser, when the service must work from an existing telephone number, or when an organization already operates a contact-centre and SIP environment. It does mean the phone route needs its own observability.

  • Call setup time: measure from the caller’s initiation to the agent’s first usable audio.
  • Mouth-to-ear delay: measure the spoken audio reaching the agent and the response returning to the caller.
  • Media-stream delay: measure the provider’s handoff from the phone call to the agent-facing application.
  • Agent response time: separate model and application processing from carrier and gateway delay.
  • Audio conversion: record the codec, sample rate, channel count, and any transcoding applied at each boundary.

The same measurements are useful in the browser, especially when an SFU, relay, recording service, or server-side agent sits between the user and the model. The difference is that the phone path adds telephony interworking as a first-class part of the test plan.

Choosing the channel for a voice agent

Choose browser WebRTC when the user can reach a web or application interface and the conversation benefits from a controllable real-time media path. It gives the product team access to browser media capture and can negotiate Opus modes that preserve more speech information than a conventional narrowband phone path. Its limits are browser permissions, device variability, network access, and the fact that Opus wideband or fullband is not guaranteed by the existence of WebRTC alone.

Choose SIP or PSTN when reachability through an ordinary phone is the primary requirement. It works with familiar numbers and telephony operations, but the design must accommodate carrier and gateway stages, possible codec conversion, and the narrower audio commonly exposed by telephone interfaces. Its limit is that the application cannot treat the phone network as a transparent extension of a browser media session.

For a service that needs both, keep the agent’s conversation logic above a channel adapter. Normalize events such as speech-start, speech-stop, audio frames, interruption, transfer, and hang-up into one internal contract. Keep channel-specific details—SIP headers, PSTN call identifiers, browser permissions, WebRTC track state, codec metadata, and provider callbacks—at the edge.

Then test the same conversation through both routes. Compare caller effort, interruption handling, recognition of critical entities, setup time, mouth-to-ear delay, and failure recovery. The recommendation is straightforward: use browser WebRTC for controlled, richer interaction; add SIP or PSTN when telephone reachability justifies the extra interworking stages and narrower audio conditions.

Bottom Line

Browser WebRTC and phone access can connect to the same voice agent, but they deliver different media paths and potentially different acoustic signals. Browser WebRTC is the stronger default when you control the interface; SIP or PSTN is the practical choice when ordinary telephone access is essential.

Do not attach an assumed millisecond penalty to PSTN or an assumed fullband mode to WebRTC. Measure the deployed route, codec, device, gateway, and agent pipeline together.

What's Next

Channel choice is only one part of the risk in a real-time platform. The companion analysis of Deepgram acquiring Daily looks at what teams should verify when the provider behind a video or voice capability changes ownership.

FAQs

Is SIP or WebRTC better for a voice agent?

WebRTC is usually the better starting point when users can access a browser or application, because the product controls more of the real-time media session and may negotiate higher-bandwidth audio. SIP or PSTN is better when callers need ordinary telephone access, provided the team measures the added telephony stages and tests the narrower audio conditions.

Does WebRTC always use wideband or fullband audio?

No. WebRTC endpoints are required to implement Opus, PCMA, and PCMU, but the actual codec and Opus bandwidth mode are selected through session negotiation. Device capability, endpoint preferences, and network conditions determine whether a session uses a narrower, wideband, or fullband mode.

Does every PSTN call use G.711 at 8 kHz?

No. Conventional telephone paths commonly expose G.711-like narrowband audio, and some provider interfaces document 8 kHz μ-law audio, but mobile and HD-voice paths can use other codecs and bandwidths. Confirm the format at the application boundary of the telephony provider you plan to use.

Does PSTN add a fixed amount of latency to a voice agent?

No fixed amount applies to every deployment. Carrier routing, gateway placement, buffering, codec conversion, network conditions, and agent processing all contribute, so measure call setup, mouth-to-ear delay, provider media-stream delay, and agent response time separately.

Why does audio bandwidth matter to an AI voice agent?

Bandwidth affects the acoustic information available to speech recognition and turn-detection systems. Studies link speech intelligibility and ASR performance to bandwidth among other conditions, but the effect for a particular agent must be tested with its actual codec, microphone, noise environment, and model pipeline.

Should a product support both browser and phone access?

Support both when the product needs browser interaction for some users and telephone reachability for others. Keep the conversation events and agent logic channel-neutral, while isolating SIP, PSTN, browser, codec, and provider-specific behavior behind adapters.