Speech recognition is in every single place nowadays, but some languages, equivalent to Shakhizat Nurgaliyev and Askat Kuzdeuov’s native Kazakh, lack sufficiently giant public datasets for coaching key phrase recognizing fashions. To make up for this disparity, the duo explored producing artificial datasets utilizing a neural text-to-speech system referred to as Piper, after which extracting speech instructions from the audio with the Vosk Speech Recognition Toolkit.

Past merely constructing a mannequin to acknowledge key phrases from audio samples, Nurgaliyev and Kuzdeuov’s main purpose was to additionally deploy it onto an embedded goal, equivalent to a single-board pc or microcontroller. In the end, they went with the Arduino Nicla Voice improvement board because it incorporates not simply an nRF52832 SoC, a microphone, and an IMU, however an NDP120 from Syntiant as effectively. This specialised Neural Choice Processor helps to tremendously velocity up inferencing instances due to devoted {hardware} accelerators whereas concurrently decreasing energy consumption. 

With the {hardware} chosen, the group started to coach their mannequin with a complete of 20.25 hours of generated speech knowledge spanning 28 distinct output lessons. After 100 studying epochs, it achieved an accuracy of 95.5% and solely consumed about 540KB of reminiscence on the NDP120, thus making it fairly environment friendly.

To learn extra about Nurgaliyev and Kuzdeuov’s venture and the way they deployed an embedded ML mannequin that was skilled solely on generated speech knowledge, try their write-up right here on Hackster.io.