iTantra
(2026)
A walkie-talkie that carries speech where there is no network. Indic speech recognition runs on the phone, the sentence is compressed to about 45 bytes, it crosses any link you have, Wi-Fi, Bluetooth or a LoRa radio, and neural text-to-speech says it out loud at the other end. Built for the ISRO problem statement at Smart India Hackathon 2026.
Measured on real speech from the FLEURS benchmark rather than estimated: character error rate 2.9% for Hindi, 4.0% Bengali, 7.8% Telugu, 9.2% Malayalam. The quantised int8 models run at a real-time factor of 0.055 to 0.119 for recognition and 0.049 for speech synthesis, on CPU alone. The compressor holds 4.45 to 5.36 bits per character across ten languages, against 7.3 to 8.7 for the standard alternatives.
Native Android in Kotlin over a single ONNX runtime for recognition, synthesis and voice activity detection, with an ESP32 LoRa bridge in firmware, AES-GCM encrypted frames, an audio modem that pushes a message through an analog radio at 6 dB SNR, and error correction on that path. One speech-enhancement model was benchmarked at every signal-to-noise ratio and rejected for making recognition worse, the negative result is in the repository. 413 Python assertions, 127 firmware parity cases, 44 Android tests.
