Your agent hears only your customer.
One model keeps the voice on the call and removes everything else — the second person in the room, the TV, the bullpen, the bed of noise. No enrollment, no voiceprint, 16 ms of delay.
The same model your SDK loads, on your own voice.
This one is not a stand-in: one take, processed on our server by the 16 kHz engine we sell, played back raw and enhanced without the playhead moving.
Record a few seconds, and hear the take twice — untouched, and through the engine — switching without the playhead moving. Your take is uploaded once, processed on our server, and held only in its memory — the opposite privacy model from a widget that never sends audio anywhere, and the copy on this page will not blur the two.
Talk to an agent that hears only you.
A voice agent with the engine on its input path — flip the filter mid-conversation and hear what the model takes away.
Talk to an agent that hears only you.
The agent hears your microphone through the model — or raw, on a switch you flip mid-conversation. Its transcript tags every line with which one it was hearing. Your voice goes to our server and on to Google’s Gemini Live API for the conversation; nothing is stored on our side.
One caller. Two rooms. The same engine.
Same take, same interferer level (6 dB under the caller) — only the source changes. Every level printed here was measured from the audio you are about to hear.
Broadband noise is the case a suppressor owns. In the caller's pauses the bed falls 29.2 dB, while the caller's own level moves 1.3 dB — the room goes and the voice stays.
LibriSpeech test-clean (Panayotov et al., 2015) — CC-BY-4.0 speech · checkpoint fastenhancer-t MIT · 16 kHz. The rooms are synthetic; the results are not — the model actually ran.
Half the latency of the engine everyone benchmarks against.
15 ms of algorithmic delay against their measured 29.94, in 5.3 MB and 1.3 M parameters. Streaming and offline give the same samples, so what you hear in a file is what the call gets.
See the measurementsFour lines into the stack you already run.
npm `@anecho-official/sdk` and PyPI `anecho` — the model runs inside your process, on CPU. Instant key, no sales call. The SDK will not silently resample your audio.
import { Model, Processor, UsageReporter } from "@anecho-official/sdk";
const reporter = new UsageReporter(process.env.ANECHO_API_KEY!);
const model = Model.fromFile("anecho.ai_focus_model_16khz_v4_1.anecho"); // from your dashboard
const proc = new Processor(model, await reporter.ensureLicense());
reporter.attach(proc);
reporter.start();
// float32 at 16 kHz in, the same number of samples out, 15 ms behind.
const enhanced = proc.process(block);The first `model.audioDelay` samples are the algorithmic delay and come back as silence — exactly as the delay says they will. Runs on CPU, no network in the audio path.
Per minute, from zero, without a call.
Full pricingThe nearest comparable vendor starts at $135 a month committed annually. Two others put a contact form where the price should be. We put a signup form where the contact form should be.
Hear it on your own voice, in one command.
npx anecho mic records ten seconds from your microphone, runs the model, and plays both back. Then run it on your own call recordings — measure before you buy anything, from us or from anyone else.
Measurements on this page: Null Test run 2026-08-14 · faster-whisper:base.en · 19 conditions · prepared clips 2026-08-16 · fastenhancer-t