ETTS Built Haven's Voice as Infrastructure, Not Decoration
EMPHOS Group · April 23, 2026 · 6 min read
Most AI assistants borrow their voice.
Sometimes that voice comes from a cloud provider. Sometimes from a generic model licensed into a product that otherwise wants to feel distinctive. The result is usually serviceable. It is rarely owned. The product may have its own interface, philosophy, and architecture, but the moment it speaks, it sounds like infrastructure rented from somewhere else.
EMPHOS did not want that for Haven.
ETTS exists because voice is not a cosmetic layer. In an assistant, voice is part of the identity surface.
What ETTS actually is
ETTS stands for EMPHOS Text To Speech. At the implementation level, it is a full build pipeline: generate corpus, train models, run joint fine-tuning, and export deployable artifacts.
The current system was built around a 3,572-sentence corpus spanning five emotional modes: calm, warm, focused, empathetic, and enthusiastic. That corpus feeds HavenAM, the acoustic model, and HavenVoc, the vocoder, before a joint pass sharpens the two together for deployment.
This is what makes ETTS different from a feature-level voice integration. It is not a single model glued into a product. It is a stack with its own internal logic and its own engineering discipline.
One voice, five emotional modes
The emotional structure matters more than it may appear to.
Haven is not meant to sound like the same person in every context. A morning briefing, a focused work check-in, a supportive response, and a higher-energy prompt do not all deserve the same cadence, emphasis, or pause shape. ETTS encodes that distinction directly into the system with five mode-specific presets.
That is how a product voice starts becoming a real product voice. Not by becoming theatrical, but by becoming context-aware in a way that still feels coherent.
The goal is not range for its own sake. The goal is emotional precision.
The engineering standard is unusually high
One of the strongest signs that ETTS is a serious system is that EMPHOS did not stop at getting a plausible output.
The project includes an architecture review against multiple reference implementations, and the work that followed was not superficial cleanup. It found a real KL divergence problem in the HavenAM training path, corrected it to match canonical VITS byte-for-byte, removed the stabilizing hacks that had been compensating for the bug, and tightened the pipeline around verifiable behavior instead of hopeful behavior.
That kind of engineering honesty matters. Voice systems are easy to make look fine in a demo and much harder to make correct as a training and deployment pipeline. ETTS chose the harder path.
Deployment matters as much as training
A voice stack that only works in the lab is not finished.
ETTS exports real deployment artifacts, including haven_am_fp16.onnx and haven_voc_fp16.onnx, with paths designed for GPU deployment and optional TensorRT acceleration where it is available. That matters because EMPHOS is not building a voice demo. It is building a voice layer that has to survive product reality.
Once a voice system reaches export, the question changes. It is no longer only about whether the model can speak. It becomes about whether the voice can travel across devices, runtimes, and future products without breaking its identity.
ETTS was built with that question in mind from the start.
Why this matters for the EMPHOS stack
Haven is the obvious home for ETTS, but it is not the only one. A properly built voice layer compounds across products. It gives future systems a way to sound recognizably EMPHOS without starting from zero each time.
That is one of the deeper patterns across the company: internal systems are rarely built for a single use case. They are built so the work done once can keep paying off later.
ETTS is a perfect example of that mentality. What looks like a voice project is really infrastructure for presence.
What comes next
ETTS is already past the point of concept. The corpus exists. The models exist. The exports exist. The engineering work now is about refinement, quality, and the same pressure-testing every serious subsystem has to survive before it deserves to become invisible inside a finished product.
If ETTS succeeds fully, most users will never think about it by name. They will simply hear Haven speak and feel that the voice belongs there. That is the real success condition. A voice system disappears into the product because it fits the product so well that it no longer feels borrowed.
Engineered for Presence.
Stay in the loop
EMPHOS publishes twice a week — product updates, research, and the thinking behind the build.
Explore Haven · HEINRICH Intelligence · The EMPHOS Vision · All Posts
EMPHOS Group · Chilliwack, BC, Canada · info@emphosgroup.com