Performance Benchmarks

Real production measurements, not marketing rounding. Lokutor runs the entire voice pipeline on commodity CPUs — here's exactly where the time goes, with the methodology to reproduce it.

The Full Pipeline Latency

In real-world voice AI, TTS latency alone doesn't tell the story. We measure time to first audio: from the moment the user stops speaking to the first sound of the reply. The agent streams token by token, so the caller never waits for a whole sentence — only for the first short chunk.

Where the time goes

Measured on live production traffic, not staging — AWS c8g.2xlarge (Graviton4, 8 vCPU). Hover a segment.

Published guidance treats under 800ms as conversationally natural; an independent survey places production voice-AI platforms at a 1,400–1,700ms median. Speech recognition costs nothing on the critical path — it completes inside the window the turn-detector is already waiting out. The language model is now the largest remaining term, and the only stage not running on our own hardware.

Plans

About 2¢ a minute, everything included: TTS, STT, the LLM and orchestration are in the minute on every plan; voice cloning and phone numbers on Growth and above.

Phone numbers $5/month each; a minute of a call the agent receives counts as three, of one it makes as six. Dedicated and private deployments are quoted separately.

Noise Suppression — Psst

Psst is our proprietary strictly-causal noise suppression model, running as low-latency middleware ahead of speech recognition. We score it on downstream word-error rate — the only metric that reaches the caller — not on perceptual audio quality, which turned out to point the wrong way entirely.

300 LibriSpeech test-clean utterances mixed with held-out noise and interfering speakers; no evaluation audio appears anywhere in training. Psst leads on downstream word error by 36% over DeepFilterNet3, the strongest published real-time CPU denoiser, at comparable cost and inside a 32ms strictly-causal budget DeepFilterNet3 doesn't meet. Against competing speech — other people talking nearby, the most common real-world complaint — Psst recovers 14–28% of word errors. It runs 18 concurrent call legs per CPU core at 2.7MB, so noise suppression adds no meaningful capacity cost.

The word-error and perceptual rankings invert completely. DeepFilterNet3 leads PESQ by 0.70 and SI-SDR by 8dB and is the worst system here for transcription — worse than no denoising at all. Psst is the only system tested that improves on doing nothing. Optimizing for the industry-standard perceptual metric would have degraded the product.

Turn-Taking — Turno

Turno, our turn-taking and barge-in model, is 18,651 parameters. On a neutral real-world test set it beats Silero VAD on Average Precision (0.931 vs 0.876) and holds that precision flat from clean audio down to 0dB signal-to-noise, at roughly half Silero's per-inference cost (0.083ms per frame). Retraining on real echo data cut the real-device echo false-confirm rate from 93.0% to 11.4%.

Speech Recognition — Conv

Conv 1.0 serves production today: a fine-tuned Parakeet 0.6B running entirely on local CPU with no cloud fallback, at 86–137ms per utterance on live traffic and a real-time factor of 0.057–0.13.

Conv 2.0, our from-scratch streaming recogniser now in development, benchmarked against Deepgram Flux — the current leader in streaming latency — on our own harness, same audio, same protocol:

Conv 2.0 leads on the metric that bounds live conversation — roughly 2× Flux's published number and ~10× what we measure from Flux ourselves — with no GPU at inference. On accuracy it reaches 7.53% WER on LibriSpeech dev-clean; the remaining gap on conversational speech is bound by training-data composition, not model size or architecture, and a distillation pass from an open Flux-class teacher is already in training.

Why CPU-Native Matters

State-of-the-art voice AI today means giant models on giant GPU fleets. That model breaks down the moment voice AI needs to live outside a data center.

• Lifetime cost: sell a voice-enabled device once, and GPU-dependent inference means paying cloud bills for that device's entire lifetime. CPU-native inference removes that tax.
• Latency & connectivity: round-trips to a GPU cloud fight against real-time conversation. Running on the CPU already in the device (or a commodity server) removes a hop.
• Privacy & compliance: for regulated buyers, audio that never has to leave the customer's environment shortens procurement under GDPR and the EU AI Act — compliance as a distribution advantage, not a cost center.

Verify Data

Want to run these benchmarks on your own data? Contact us for a technical consultation or access to our evaluation scripts.