Ito: Streaming Neural Speech on a $5 Chip
A 4.9 MB text-to-speech model for the ESP32-S3. It streams, runs without a cloud and is verified in emulation. Board timings to come.
Blog
Insights into the frontier of ultra-low latency voice AI.
Oído: Open-Source Speech Recognition on a $5 Chip
English and Spanish speech recognition on a $5 ESP32-S3 at 3.7% word error. Accurate, but about twice slower than real time. Open source.
Our September 2026 Pricing, Since Replaced
The plans announced on 23 September 2026 at about 2¢ a minute, replaced on 9 October by a prepaid balance at about 5¢ a minute, model included.
Edge Voice AI: Why Voice Interfaces Are Moving Off the Cloud
Robots, wearables and connected devices can't carry a GPU cloud bill for every unit sold. Why voice AI is following compute back to the edge.
What Is CPU-Native Voice AI? A Plain-English Explainer
CPU-native voice AI runs recognition, understanding and speech synthesis on ordinary processors instead of GPUs. What it means and why it matters.
From Technical Demo to Enterprise Traction: Our Experience at 4YFN
Lokutor exhibited at 4YFN, co-located with MWC Barcelona, and came away with enterprise pilot discussions and validated demand.
Joining NVIDIA Inception
Lokutor has been accepted into NVIDIA Inception — a program that supports startups revolutionizing industries with AI and accelerated computing.
Lokutor TTS Joins Pipecat as a Community Integration
Lokutor TTS is now a community integration for Pipecat, the open-source voice AI framework. Install with pip install pipecat-lokutor.
Selected for the BSC AI Factory Program
Lokutor was selected for the Barcelona Supercomputing Center's AI Factory program, which backs AI startups with European supercomputing.
Crystal Clear Conversations: Introducing Lokutor's AI Noise Cancellation
Discover how our new noise suppression technology transforms voice AI in real-world environments, from busy cafes to windy streets.
Vela, Now Turno: Our Turn-Taking Model for Natural Conversation
Vela, now called Turno, predicts when a speaker is about to finish their turn, enabling sub-200ms response gaps without clipping.
Psst: Our Noise Cancellation Model for Real-World Voice AI
Psst is a lightweight real-time noise suppression model that makes voice AI work in busy cafes, streets and open offices.
Lokutor Selected as a Finalist at South Summit 2025
South Summit selected Lokutor as a finalist among thousands of startups from over 100 countries. We will show our voice AI platform in Madrid.
The Illusion of End-to-End: Orchestrating the Perfect Voice Agent
Why Lokutor isn't just a text-to-speech engine, but a seamless conversational platform enabled by ultra-fast orchestration.
The Future of Sound: From Interaction to Symbiosis
We are moving from 'Command and Control' to 'Conversational Symbiosis'. Why the best interface of the future is no interface at all.
Building for the Community: Introducing Lokutor Orchestrator for Go
We're open-sourcing the Go orchestrator behind our voice agents: full-duplex audio with VAD, barge-in and pluggable providers.
Beyond Transformers: Why We Use Generative Flow Matching
The architectural secret behind Versa's speed. Learn how Flow Matching revolutionizes voice synthesis by moving past slow, sequential processing.
The Physics of Trust: Watermarking Generative Audio at the Source
Detecting AI voice fraud takes more than classifiers: invisible, mathematical watermarks embedded in the sound wave itself.
Technical Benchmark: Lokutor Versa vs. Auto-Regressive Transformers
Why audio is moving beyond Transformers: auto-regressive models against flow matching, on latency, stability and variable cost.
The Architecture of Interruption: A Developer's Guide to Real-Time Voice Agents
A reference architecture for low-latency voice agents: managing state, handling interruptions and optimizing the turn-taking loop.
The 200ms Threshold: The Psychology of Latency
Why does 500ms of lag feel like an eternity? A deep dive into the psycholinguistics of turn-taking and why 'silence' is the loudest sound in a conversation.