Lokutor
Summary
Most voice AI pipelines quietly require a GPU cluster to hit acceptable latency — and that requirement quietly kills your on-prem deal, your edge device project, or your cost model. Lokutor is built around the premise that none of those stages need an accelerator.
The vendor describes a five-stage pipeline — noise suppression, turn-taking, speech-to-text, LLM, and speech synthesis — where every stage except the LLM runs on Lokutor's own CPU models. The stated first-audio latency is approximately 120 ms in streaming mode and roughly 0.9 seconds to first reply on a 4-vCPU node. Turn-taking is handled by Turno, which the docs describe as semantic rather than silence-timer-based, so a filler 'mm-hmm' does not cut the agent off. Self-hosting is confirmed via a Go-based orchestrator with install instructions on GitHub. The LLM slot is yours to fill — Lokutor does not supply the language model, which means you control that cost and that compliance boundary, but you also wire it yourself.
Bottom line: Lokutor earns its place in a regulated or edge deployment where audio cannot leave the device and GPUs are not available — but if you need the LLM managed for you, or your language is outside the 33 supported, you are doing integration work the platform does not cover.
Pricing Plans
Usage-Based- Free Tier
- 30 agent minutes per month, 1 concurrent call
Free
30 agent minutes/mo, 1 concurrent call
- Speech synthesis 33 languages
- Transcription
- Orchestration
- Tool calling
- Cross-call memory
Starter
200 agent minutes/mo, 3 concurrent calls
- All Free features
- $0.08/min overage
Growth
1,500 agent minutes/mo, 8 concurrent calls
- All Starter features
- LLM
- Voice cloning
- $0.06/min overage
Business
5,000 agent minutes/mo, 20 concurrent calls
- All Growth features
- $0.04/min overage
View full pricing on lokutor.com →
Pricing may have changed since last verified. Check the official site for current plans.
Community Performance Report Card
No community ratings yet. Be the first to rate this tool!
Pros
Sign in to edit- CPU-only inference across noise suppression, turn-taking, transcription, and synthesis — so you avoid GPU provisioning costs entirely and can deploy inside environments where accelerator hardware is not available or permitted.
- Self-host path is confirmed via a Go-based GitHub orchestrator with ARM64 and x86 support, which means audio never leaves your perimeter — removing the compliance blocker that kills most cloud voice vendors in regulated verticals.
- Semantic turn-taking via Turno classifies actual conversational intent rather than silence gaps, so the agent does not interrupt on backchannels and does not hang waiting for a silence threshold that a noisy caller never produces.
- Provider-agnostic LLM slot, so switching language models — for cost, capability, or compliance reasons — does not require rearchitecting the audio pipeline.
- Visemes are included with the 10 voices, which means lip-sync for avatar or video applications works without a separate synthesis pass.
Cons
Sign in to edit- The LLM is never included — you supply and manage it yourself. Teams that arrive expecting a fully managed, single-vendor voice stack will discover they are operating two separate systems from day one, and the integration complexity lands entirely on them.
- Language support is fixed at 33 languages; anything outside that list requires submitting a vendor request with no stated delivery commitment. A team building for a market in that gap cannot ship until the vendor responds — and the page offers no workaround.
- The free tier is designed for experimentation; production traffic on the vendor's cloud moves to per-minute billing that bundles synthesis, transcription, and orchestration together. Teams that want to separate and optimize costs by component cannot — they pay the bundle rate or they self-host the full stack.
- Voice variety is limited to 10 named voices. Teams that need custom brand voices or a wider casting pool will hit this ceiling before long and either accept the constraint or move to a platform with voice cloning — at which point Lokutor is no longer the synthesis layer.
About
- Platforms
- Cloud, on-prem, on-device, GitHub (Go), Python SDK
- API Available
- Yes
- Self-Hosted
- Yes
- Last Updated
- 2026-09-16T10:36:36.122Z
Best For
Who it's for
- Developers building CPU-efficient voice agents
- Applications requiring low-latency turn-taking
- Teams needing self-host or on-prem options
- Multi-language voice synthesis projects
What it does well
- Real-time voice agents for customer support
- Conversational AI applications with tool integration
- Low-latency speech synthesis in multiple languages
- On-device or on-prem voice orchestration
Integrations
Add notes, reviews, and benchmarks so the next visitor gets a clearer picture.
Compare Lokutor
Spotted incorrect or missing data? Join our community of contributors.
Sign Up to ContributeFrequently Asked Questions
- Is Lokutor free?
- Lokutor has a permanent free tier alongside paid upgrades. You can keep using a baseline version indefinitely without paying.
- Is Lokutor open source?
- No — Lokutor is a closed-source tool. Source code is not publicly available.
- Does Lokutor have an API?
- Yes. Lokutor exposes a developer API. See the official documentation at https://lokutor.com for details.
- Can I self-host Lokutor?
- Yes. Lokutor supports self-hosting on your own infrastructure.
- What platforms does Lokutor support?
- Lokutor is available on: Cloud, on-prem, on-device, GitHub (Go), Python SDK.
Curated lists that include this category
Most voice AI pipelines quietly require a GPU cluster to hit acceptable latency — and that requirement quietly kills your on-prem deal, your edge device project, or your cost model. Lokutor runs noise suppression, turn-taking, speech-to-text, and speech synthesis on its own CPU models, leaving only the LLM for you to supply.
Pipeline and latency
The vendor describes a five-stage pipeline with first-audio latency of approximately 120 ms in streaming mode and roughly 0.9 seconds to first reply on a 4-vCPU node. Turn-taking uses Turno, which the docs describe as semantic rather than silence-timer-based, so a filler ‘mm-hmm’ does not cut the agent off. Self-hosting works through a Go-based orchestrator on GitHub with ARM64 and x86 support. Synthesis covers 33 languages.
Pricing and limits
Pricing is usage-based with a free tier of 30 agent minutes per month and 1 concurrent call. The LLM slot, platform choice, and any tool-calling logic stay under your control via the Python SDK or listed integrations.
Who it is for / who should skip it
Best for developers building CPU-efficient voice agents, teams that need self-host or on-prem options, and projects requiring low-latency turn-taking. Teams that expect a fully managed stack including the LLM or support for languages outside the fixed list of 33 will hit integration work and waiting periods from day one.