Acoustic Intelligence

Acoustic Intelligence

AI that discovers speech from sound, not text.

An evolutionary system that discovers speech-like acoustic structure through search and selection — without training data, without phoneme models, without being told what speech is.

First product in development.

What This Enables

Acoustic Intelligence enables a new class of AI systems that generate speech directly from physics rather than text or training data.

Speech generation without datasets

No phoneme models, no transcribed audio, no language-specific training data required.

AI-native communication

Systems that develop their own acoustic communication protocols from first principles.

Ultra-efficient voice synthesis

Evolutionary discovery on cloud infrastructure — scalable acoustic generation without massive neural networks.

Beyond language-first AI

New human-machine interaction paradigms where sound comes before text.

A New Category of AI

Not text-to-speech

Traditional TTS converts text through phoneme pipelines. We don't start from text. We start from physics.

Not voice cloning

We don't imitate human recordings. The system discovers acoustic structure from scratch through evolutionary search.

Emergent discovery

Given only a physically realistic instrument and a speech detector, the system independently discovers how to produce speech-like sound.

Compute-native intelligence

Sound generation learned as continuous control intelligence — evolutionary search over millions of acoustic trajectories on cloud compute.

Breakthrough Results

Early-stage deep-tech startup building the first generation of acoustic intelligence systems.

12 seeds converged

12 independent random starting points all evolved toward the same speech-like acoustic structure. Not random chance — deterministic convergence.

0.97 speech score

Composed acoustic primitives score 0.97 on speech detection — higher than any individual evolved genome. Composition improves on its building blocks.

Universal onset primitive

The system discovered a single 50ms acoustic gesture that acts as a universal speech initiator — causally responsible for 82% of the speech signal. Found independently by every seed.

$79 total compute

From blueprint to discovery in 36 hours on AWS. Parameter sweeps on 192-core Graviton instances. Efficient evolutionary search, not brute-force training.

Listen

Early-stage research outputs. Distinct speech-like acoustic structures discovered through evolution.

Evolution Progression

Random noise → evolved structure → discovered speech primitive → composed utterance

Discovered Speech Onset (Segment 106)

A 50ms acoustic gesture the system discovered to initiate speech-like sound — found independently by every seed

Evolved Speech-Like Utterance (0.97 score)

Four discovered primitives composed into the highest-scoring speech-like output

How It Works

🔊

Explore

A physically realistic acoustic instrument generates millions of sound trajectories

📊

Evaluate

Multiple critics score each sound for structure, diversity, and speech-likeness

🧬

Evolve

MAP-Elites preserves diverse discoveries while improving quality through mutation

🎵

Compose

Discovered primitives are sequenced into higher-quality acoustic structures

Built on AWS

Acoustic Intelligence runs large-scale evolutionary search across cloud compute, evaluating millions of acoustic trajectories. The system is designed to scale natively on AWS — from distributed exploration on 192-core Graviton instances to real-time composition and deployment. Currently running large-scale experiments on AWS EC2. Our research has already leveraged compute-optimised Graviton instances for parameter sweeps, seed replication, and multi-seed validation campaigns.

About

Acoustic Intelligence is an Australian deep-tech AI research startup exploring emergent acoustic intelligence — systems that discover the building blocks of speech through evolutionary search rather than supervised learning.

Founded by Richard James. Based in Brisbane, Australia. Compute-intensive AI/ML research on AWS.

This work demonstrates that speech-like structure can emerge purely from physical constraints and selection, without supervision.

Get in Touch

Partnerships, research collaborations, and early access inquiries welcome.

richard@acousticintelligence.ai