voxracer

VoxRacer is an open-source, provider-neutral local profiler designed to measure and break down response time latencies in conversational voice agents. Instead of relying on provider self-reporting or guessing missing data, it provides an objective, turn-by-turn audit of where time is spent during a voice interaction.

1. For the Non-Technical Reader

Imagine a voice agent turn as a relay race: your speech stops, the agent listens, transcribes, thinks, queries external databases, generates speech, and streams audio back to your phone. If a call feels slow, where did the handoff stall? VoxRacer acts like a precision camera at every leg of the race. It allows businesses using voice AI to see exactly why a conversation felt laggy, pinpointing whether the bottleneck was caused by speech recognition, internal AI reasoning, or text-to-speech audio generation.

2. For the Technical Reader

VoxRacer defines a provider-neutral session model and timing math to decompose voice-agent turn latencies without assuming zero for unmeasured values. It measures up to eight canonical timing components:

• Caller-perceived time to first audio
• Endpointing
• Speech-to-text (STT)
• Language-model time to first token (LLM TTFT)
• Tool time
• Text-to-speech time to first audio (TTS TTFA)
• Playback
• Unattributed time

The initial release includes adapters for providers such as ElevenLabs and Vapi, reading OTLP and transcript timing without confusing provider self-reports with true caller-perceived audio timing. Developed in Python with strict mypy type checking and distributed under the Apache-2.0 license, VoxRacer handles missing data explicitly as nulls in canonical JSON output.

3. Why It Matters

In real-time Voice AI, latency is the difference between a natural conversation and an awkward interaction. However, vendor-reported latencies often ignore client playback, carrier overhead, and endpointing detection times. VoxRacer introduces much-needed observability and vendor-neutral benchmarking to the Voice AI stack, helping engineering teams audit SLA claims, optimize multi-vendor pipelines, and isolate ghost delays.

4. The Voice AI Space Lab Idea

The Automated Voice SLA Referee: Integrate VoxRacer into your continuous integration (CI) test suite. Every night, run automated test calls across different provider combinations (e.g., Deepgram + OpenAI + ElevenLabs vs. unified voice platforms). Use VoxRacer to parse the call artifacts and automatically flag regressions if unattributed time spikes or LLM TTFT breaches conversational budgets.

Check out the project on GitHub.