Telepath
Technical Case Study: Signal-to-Noise Ratio in Voice AI
Analyzing the impact of packet-level audio hygiene on End-to-End Latency.
Executive Summary
In a controlled A/B test conducted in us-east-1 (Ashburn, VA), the Telepath Gateway demonstrated a 2.9x reduction in Call Setup Time and a 6.6x reduction in First Turn Latency compared to a direct carrier integration.
By implementing an adaptive Clean Signal Protocol at the edge, Telepath eliminates the "VAD Hesitation" caused by standard carrier background noise, effectively decoupling AI performance from network variability.
Methodology & Control Environment
To isolate the impact of our signal processing, we enforced total geographic parity.
- Region: All components (Telnyx Media, Telepath Gateway, ElevenLabs Inference) were co-located in Ashburn, VA.
- Variable: The only variable was the Audio Transport Protocol (Direct SIP vs. Telepath Optimized).
- Significance: Network latency (ping) was neutralized (<2ms). All measured differences are purely due to software efficiency.
The "First Impression" Gap
The most critical moment in a Voice AI interaction is the "Cold Start"—the time between the user answering and the AI knowing to speak.
Session Establishment (SIP Handshake)
- Direct Path (Telnyx → ElevenLabs):
1,812 ms - Telepath Path (Telnyx → Telepath → ElevenLabs):
622 ms - Result: Telepath establishes the media session 2.91x Faster.
Telepath maintains warm, persistent WebSocket connections to the Inference Cluster, avoiding the overhead of "Cold Booting" a new SIP dialog for every call.
First Turn Latency (The "Hello" Moment)
- Direct Path:
ASR 829 ms— The Cloud VAD hesitated for nearly a full second, struggling to distinguish the user's initial breath/noise from speech due to the "dirty" carrier stream. - Telepath Path:
ASR 124 ms— Telepath's Clean Signal Protocol fixes this by adaptively handling breaks in the user's speech. The VAD triggered instantly. - Result: Telepath is 6.68x Faster at the first interaction.
The difference: Standard carriers pass background noise that confuses AI VADs. Telepath implements a sophisticated clean signal protocol that gives the VAD perfect timing when the moment speech begins.
Steady State & Stability
Once the conversation stabilized, we measured performance under stress.
ASR Latency (Steady State)
| Turn | Direct Carrier | Telepath | Delta |
|---|---|---|---|
| Turn 1 (Cold) | 829 ms | 124 ms | -705 ms |
| Turn 2 | 139 ms | 130 ms | -9 ms |
| Turn 3 | 142 ms | 124 ms | -18 ms |
| Turn 4 | 130 ms | 132 ms | +2 ms |
Insight: While Direct integration eventually "warms up" (Turn 2+), it fails the critical First Impression. Telepath is instant from millisecond zero.
The Jitter Resilience Stress Test
During testing, the Telepath leg experienced a severe Residential Cable Jitter event.
- Measured Jitter:
77.10 ms(2.5x higher than standard VoIP tolerance). - Measured Packet Loss:
0.00% - Audio Artifacts: None.
Telepath's predictive clean signal protocols absorbed the jitter variability without introducing robotic artifacts, proving the system is robust against public internet instability.
Conclusion
The data proves that Standard SIP Trunks are inefficient for Voice AI.
They introduce:
- 1.2 Seconds of wasted Setup Time.
- 700ms of wasted VAD Hesitation on the first turn.
Telepath reclaims this lost time. We do not make the AI think faster; we clear the path so it can start thinking 1.9 seconds sooner.