Voice AI is only as good as its ability to listen. In a laboratory setting, speech recognition is nearly perfect. But the real world isn’t a lab. It’s a symphony of chaos: the hiss of an espresso machine, the roar of a passing bus, or the persistent hum of an office air conditioner.
Today, we are excited to unveil Lokutor’s Intelligent Noise Suppression (INS)—a breakthrough in real-time audio processing designed to make voice interaction reliable anywhere.
The Problem: The “Cocktail Party” Effect for AI
Humans are remarkably good at filtering out background noise. We can focus on a single speaker in a crowded room—a phenomenon known as the cocktail party effect. Computers, however, traditionally struggle. Background noise “pollutes” the audio signal, causing speech-to-text models to hallucinate or fail entirely.
For developers building voice agents, this is the #1 cause of user friction. If your agent depends on a clean signal, it’s tethered to quiet rooms.
Our Approach: Adaptive Deep Cleaning
Most noise cancellation filters are static; they remove specific frequencies (like low-end hums) but struggle with dynamic noises like dishes clattering or people talking in the background.
Lokutor INS uses a specialized Real-Time Neural Filter that runs concurrently with our STT engine. It doesn’t just block frequencies; it recognizes the unique patterns of human speech and reconstructs the voice while discarding everything else.
Better Accuracy, Lower Latency
By cleaning the audio before it hits the STT engine, we’ve seen a 25% increase in transcription accuracy in high-noise environments. Because our model is optimized for CPU, this adds less than 5ms of overhead to the total pipeline.
Beyond Just “Clear Audio”
Noise cancellation isn’t just about making things sound better for humans; it’s about making voice AI functional in the wild.
- Confidence: Users aren’t afraid to speak to their devices in public.
- Efficiency: Cleaner signals mean the STT engine doesn’t have to work as hard (less CPU cycles spent on “fuzzy” tokens).
- Accessibility: For users with speech impairments, INS can help isolate primary vocal cords from secondary environmental sounds.
Hear The Difference
We’ve put Vex to the test. Listen to how it transforms a real noisy environment:
| Audio Sample | Listen |
|---|---|
| Clean Reference | |
| Original Noisy | |
| After Vex |
The “Noisy” sample was captured with strong background interference. Vex removes the noise while preserving the speaker’s natural tone and rhythm.
Benchmark Results (Vex)
Our model, codenamed Vex, has been tested against standard objective metrics on a held-out evaluation set:
| Metric | Score | Description |
|---|---|---|
| PESQ (WB) | 1.74 | Perceptual quality, higher is better (max 4.5) |
| STOI | 77.3% | Speech intelligibility, higher is better |
| SI-SDR | -14.2 dB | Signal to distortion ratio, higher is better |
Benchmark performed on our demo evaluation set with 16kHz audio. The low SI-SDR reflects the challenging nature of the noise profile—our real-time filter prioritizes preserving voice naturalness over aggressive noise removal, which is critical for natural-sounding conversations.
Deployment
Lokutor INS is now available as a standard middleware in our Go Orchestrator and as a configuration flag in our Python/JS SDKs.
const client = new VoiceAgentClient({
apiKey: '...',
noiseCancellation: true, // Enable INS with one line
});Try it yourself in our developer dashboard and start building voice agents that work wherever your users go.
Available now for all Enterprise and Pro tier users.