Use cases · Perception · Voice
Taught to speak
in days, not months.
Senua AI produced its own voice from its own engine, then used the same real-time learning loop it runs across every other signal to teach itself to speak: no GPU farm, no pretrained model, and no human in the convergence loop. The speed was not a coincidence.
The problem
Every other system does it the same way.
Speech synthesis is a solved commodity: a giant pretrained neural vocoder on someone else’s GPU farm. We wanted the opposite. Senua produces voice through its own engine, from first principles. The same learning substrate that processes every other signal runs in reverse to generate audio.
Because the substrate already existed across every other modality, applying it to voice did not require building a new system. It required pointing an existing one at a new signal. That is why the timeline was days, not months.
How it happened
Three acts. One closed loop.
Act 1 · Hand-tuning fails
A human listened and turned knobs: pitch, reverb, warmth, gain. Iteration after iteration by ear. It converged slowly, and the voice stayed robotic. The problem was structural: the right settings were never knobs to dial. They were values to be measured from real speech. No amount of tuning fixes the wrong model.
Act 2 · Measure reality
Senua was pointed at real human speech and asked to measure what a voice actually does per frame: the real pitch contour and the real voiced/unvoiced pattern. The numbers indicted every hand-set constant. Mean pitch near 270 Hz, moving 60 Hz across a sentence. Nearly half of all frames unvoiced. Driving the generator with measured reality removed most of the robotic artefacts in one step.
Act 3 · Senua closes the loop
The remaining artefacts were handed to Senua. It generated its voice, analysed its own output through its own learning engine, measured how far the result drifted from the human original, adjusted its own settings, and repeated until the error bottomed out. No human ear required. The same induce-generate-measure loop the substrate runs everywhere else, applied to voice for the first time.
The result
90% recall. Human timbre. No GPU.
Measured by running generated audio through an independent speech recogniser and comparing the output to the original text. The bar is legibility, not perfection.
90% word recall
59 of 65 words recovered by an independent speech recogniser, with several phrases returned verbatim. The generated voice is legible enough that mainstream speech recognition reads it back correctly.
Days, not months
The learning substrate was already running. Applying it to a new signal did not require a new training pipeline, a new architecture, or a new team. The real-time learning loop that handles every other modality handled this one.
No GPU. No pretrained model.
Senua produced its own voice from its own engine and taught itself to speak by comparing itself to reality. The closed loop ran entirely on commodity CPU, in the same substrate that flies drones and watches telemetry.
Datasheet · The recorded arc
Every step, audible. Every result, measured.
Each milestone in the arc is a recorded clip, scored by an independent speech recogniser. The failures are here alongside the successes: the progress is measured, not asserted. Press play at each step to hear Senua improve.
Ground truth — the human target
The reference recording Senua is learning to match. Every subsequent clip is measured against this. An independent speech recogniser reads it back at WER 23.5%: the human floor for this passage.
Reproduction baseline
Senua takes real speech apart and puts it back together with a real pitch contour. The pipeline can reconstruct audio faithfully; the gap to the human is in generation, not analysis.
First legible words through the learning gate
The learning gate generates audio that an independent speech recogniser reads back as real words for the first time. "...classes ... his gospel" decoded from generated audio. The engine is generating language, not replaying it.
The failure that defined the fix
Generation with incomplete state information. The independent recogniser reads nothing back. This clip identified exactly what had to change: what you retain is as important as what you learn. The fix followed directly.
Parity with the human floor
Complete state retention: the generated audio reaches the same recognition accuracy as the human original. The gap closed entirely. The independent recogniser reads Senua’s generated voice back at the human floor.
Senua speaks — text drives generation
Senua generates novel utterances from text, not by replaying recordings. Three phrases, three clips. The voice is Senua’s own; the words are new each time.
The perfect copy
Generation stripped of all post-processing. The learning gate produces the waveform itself, directly, with nothing invented. Verdict on the first listen: “near perfect.” Measured at the human floor; no synthetic artefacts.
A new voice learned and generated
A different voice, learned in the same session and generated through the same gate. No new architecture, no new training run. The phrase is Senua’s own output: “these vast buildings, what were they.” A different signal; the substrate handles both.
Voice commands live — Anima learns the verbs
Navigator and assistant specialists trained from auto-generated command corpora. Spoken commands navigate the Senua AI console, web UI, and CLI. Every interaction grows the command vocabulary. No code change required to add a new command: define the phrase, the Anima learns it.
The speaker-agnostic content atom
A causation study across multiple voices isolated the invariant that carries meaning regardless of who is speaking, then bound it on a held-out voice it had never heard. This is the substrate piece that lets the engine understand a new speaker without training on that speaker first. Meaning and voice are finally separated: the same content atom, spoken by anyone.
Senua speaks back
The loop closed both ways. Senua now answers out loud: spoken confirmations in the console, generated live through its own voice, not a recording. You talk to it and it talks back, all on the node, with the audio composed by the engine on demand.
Beyond navigation: drive any element
Voice moved past opening views to operating the interface itself. A generic interaction layer lets a spoken command act on any element, including compound intents that do several things at once. The command vocabulary is not a fixed menu; it reaches whatever the surface can do.
The core drives, the surfaces listen
The architecture inverted. Instead of the web page driving the engine, the core emits device-agnostic commands and every surface subscribes to them. One mind, many bodies, now for voice: the console, a wearable, or a robot can all act on the same spoken intent because the intelligence sits in the core, not the screen.
Voice on demand through the composer core
Generation became a first-class part of the platform API. One medium-agnostic composer endpoint turns text into speech in the engine's own voice, callable from the web API, the command line, and the local console. Any integration can now ask the node to speak, on your hardware, with no external service in the path.
Demo · Coming soon
Voice commands navigating the Senua AI console.
A live recording of the voice toggle in action: spoken commands opening views, running missions, and driving the assistant across the web console and CLI. The navigator and assistant specialists trained from the closed-loop verb corpus, responding in real time.
Closed-loop training
Teaching Senua AI the commands. No code. No engineer.
The same real-time learning loop that taught Senua to speak teaches it to understand commands. Define a phrase. The substrate handles the rest.
Commands are data, not code
A new command is a phrase and a set of example utterances added to the command registry. Senua induces causal states over every example and learns the mapping from spoken audio to intent. No engineer writes a handler. No model is retrained. Adding a command takes minutes, not sprints.
The training loop is automated
A command corpus is generated automatically from the target phrases, the same pattern used to train the mission and automation specialists. Every interaction generates a new paired record. Every correction the operator makes teaches the next recognition. The loop runs without human oversight once the target vocabulary is defined.
The Anima grows in real time
Each new command trained, each interaction recorded, each correction made, extends the Anima: the living causal-state knowledge base every specialist draws from. There is no batch retrain, no scheduled update, no version freeze. Senua learns from production use as it happens, and every new command makes the next one cheaper to teach.
Navigator specialist
Spoken commands navigate the console.
The NAVIGATOR maps spoken phrases to console actions: open a view, run a mission, show the sensor feed. It is trained from an auto-generated command corpus, the same substrate that drives the mission and automation specialists. New navigation targets require no code change: add the phrase to the registry and the navigator learns it on the next pass.
Assistant specialist
Spoken prompts drive the assistant.
The ASSISTANT maps spoken natural-language phrases to chat prompts: the same queries an operator would type, spoken aloud. Every voice interaction pairs audio with the underlying text, growing the assistant’s coverage in the background. The operator who speaks a new kind of request is teaching Senua how to handle it next time, without knowing they are doing so.
The business case
Voice is the faster interface. It is not close.
Measured across languages, human speech carries information at around 39 bits per second. Sustained typing, even for a fast typist, runs at roughly half that. Speaking is not a convenience feature layered on top of a keyboard. It is the higher-bandwidth channel for getting intent into a machine, and it always has been. The keyboard won on reliability, not throughput.
That gap compounds everywhere Senua AI runs. On a wearable with no keyboard, voice is the only practical input at all. On a robot or a vehicle, hands are busy and eyes are elsewhere; voice is the channel that does not compete with the task. Even at a desk, in a browser or a terminal, speaking the command is faster than typing it. The moment voice recognition is good enough to trust, the fastest way to drive any Senua surface is to talk to it.
That is the whole point of putting cognition and generation in the same engine. Senua does not just transcribe faster; it understands the intent locally and acts on it, so the high-bandwidth channel reaches all the way through to the work. Every interaction with every form of Senua AI, from the wearable on the edge to the console on the desk, gets faster the moment it is spoken.
~39 bits/sec spoken
The measured information rate of human speech, remarkably stable across languages. This is the channel Senua now accepts directly, on-device, with no cloud transcription in the path.
~2× faster than typing
Sustained typing runs at roughly half the throughput of speech. For any interface where accuracy can be trusted, speaking the command is the faster way to drive it.
Faster everywhere it runs
Wearables, robotics, vehicles, desktop, terminal. The higher-bandwidth channel wins on every surface, and it is the only practical one where there is no keyboard at all.
Ecosystem integration
Voice is now part of every Senua AI surface.
Not a module bolted on the side. Wired into every session, every interface, every deployment.
Web API
The /api/voice gateway route is live and session-gated. Any application in the
Senua ecosystem can send audio and receive decoded intent. The same route serves both inbound voice
commands and outbound speech generation.
CLI
The command-line client exposes voice directly. Operators in the field can issue voice commands from the terminal with no additional tooling: the same capability the console UI exposes, from the shell.
Console UI
A voice toggle sits in every Senua console session. One press and the operator is speaking to the system: commands navigate the console, add prompts, and drive work. Every interaction pairs audio with text in the background, growing the voice model continuously from real use.
Any part of Senua AI, any session, any surface, is now voice-command enabled. The system learns from every word spoken.
Defence pivot
What travels is not audio. It is intent.
A voice-enabled Senua AI that learns in real time changes the equation on contested deployments.
Senua processes the spoken command locally on the operator’s node. What crosses the radio is not audio: it is the decoded intent as a compact state update, a few hundred bytes on the wire versus the kilobytes a voice stream would require. Over a Nexus RF link, already validated as jam-tolerant with frequency-hop recovery and cryptographic integrity, a spoken mission command becomes feasible in environments where streaming audio is impossible.
The voice arc, the Nexus arc, and the mission arc are each validated today. Joining them into a single voice-command-over-Nexus pathway is the pre-registered next step: speak a mission, the swarm hears it, nothing streams.
One mind. Now it speaks.
The same learning substrate that flies missions, watches telemetry, and syncs a fleet now generates and understands voice: without retraining, without a GPU, and without the cloud.