Tyllow TTS + Evermind AI

The problem

Voice AI and LLM companions normally require sending audio and conversation to a third-party API. For a private companion product, that is the wrong trade — the most sensitive data in the system leaves the device.

The approach

Built emotion-aware text-to-speech with Coqui, Kokoro, OpenMOSS and an ISTFT vocoder, paired with an offline 4-bit quantized LLM companion running Qwen3 through llama.cpp. Everything runs locally, with inference tuned to the hardware it lands on.

Results

• Fully offline, privacy-first inference — no audio or conversation leaves the device.
• Emotion-aware speech synthesis rather than flat TTS.
• 4-bit quantization brings the companion model within consumer hardware budgets.

Engineering notes

Hardware-aware inference is the constraint that shapes everything: quantization level and vocoder choice are dictated by the target device, not by benchmark scores.