Ito: speech synthesis in 4.89 MB per voice on ESP32-S3

Lokutor released the Ito inference engine, ESP32-S3 firmware and two English voice models: 4.05M parameters and 4.89 MB of weights per voice, producing 24 kHz speech with integer arithmetic. The firmware is verified in Espressif's QEMU emulator, but nothing has run on a physical board yet.

• 4.05M parameters and 4.89 MB of weights per English voice
• 24 kHz synthesis with integer math on a 240 MHz chip without an NPU
• QEMU output is bit-identical to the host engine's PCM
• Estimated RTF 0.53–1.27; first audio 137–318 ms