Voice AI developers are obsessed with optimizing speech-to-text, LLM, and text-to-speech pipelines, but they are often ignoring the biggest source of latency: the phone call itself. Dialtrace, a new open-source tool created by a Telnyx forward-deployed engineer, puts carrier events and custom pipeline stages on a single timeline. This allows teams to see exactly where the time goes, distinguishing between network signaling, human pick-up delays, and actual AI processing time. The tool is particularly relevant for teams using Telnyx, as it leverages existing webhook data to generate detailed latency reports without requiring packet captures or deep SIP expertise.

The Hidden Cost of Signaling and Ring Time

Standard voice AI benchmarks measure the model pipeline, but Dialtrace highlights that a significant portion of caller wait time occurs before the AI even starts thinking. In a real-world example from a US number to a Dutch mobile, the caller waited 640ms after picking up before hearing any audio, with nearly half of that time spent on setup and ring delays. The tool explicitly excludes ring time from its latency percentages because it represents a human reaching for a phone, not a technical bottleneck. By separating signaling delays from application logic, Dialtrace prevents teams from optimizing the wrong part of the stack.

Zero-Dependency Integration for Real-Time Debugging

Dialtrace is designed for minimal friction, requiring zero dependencies and a single command to run. It integrates with Telnyx Call Control APIs, using webhooks to timestamp events like call.initiated, call.answered, and stt.first_partial. Developers can inject custom markers into their agent code with a simple three-line helper function that POSTs to a local receiver. The tool handles clock skew between the carrier and the developer's machine by estimating the offset based on API round-trip times, providing error bars when the precision is insufficient for critical decisions.

Key Takeaways

  • Dialtrace visualizes the entire voice call timeline, including carrier signaling, ring time, and AI pipeline stages.
  • It is built for Telnyx users, leveraging existing webhooks and Call Control APIs without requiring packet capture tools.
  • The tool distinguishes between technical latency (pipeline, network) and human latency (ring time), preventing misdirected optimization efforts.
  • Custom pipeline stages can be added via simple HTTP POST requests to a local receiver, making it easy to instrument existing agents.
  • Clock skew between carrier timestamps and local events is estimated and reported with error bars to maintain data integrity.

The Bottom Line

If your voice AI feels slow despite a sub-second pipeline, you are likely ignoring the phone call. Dialtrace provides the visibility needed to stop guessing and start fixing the actual bottlenecks.

Technical Limitations and Future Extensibility

While Dialtrace is powerful for Telnyx users, it does not analyze RTP streams, meaning it cannot detect one-way audio or provide second-by-second quality metrics. The tool relies on the Call Control API, which does not emit a distinct ringing event, forcing the tool to report setup and ring time as a single block. However, the architecture is carrier-agnostic, with only about 80 lines of code specific to Telnyx. This design invites contributions for other providers that emit call progress events with timestamps, potentially expanding its utility across the voice AI ecosystem.