simoneVU/Voice-Recognition-Mali-Seeds

# Voice Recognition System for the Mali market system The Voice Recognition system is trained to specifically recognize the seed types used in Mali: corn, cowpea, fonio, millet, okra, peanute, rice, sesame, sorghum, and soy. ## Data Collection The initial and crucial step in deploying a model is data collection, which involves considering two main factors: data quantity and data quality. In terms of data quantity, for fine- tuning a 10-class detector, a total of 190 samples are collected, with 10 samples per participant and 19 participants in total. Similarly, for testing, a total of 110 samples are collected, with 10 samples per participant and 11 participants in total. When it comes to data quality, participants were instructed to collect audio recordingsin quiet environments with minimal background noise, while the 10 audio recordings of the seed pronouncing the words as naturally as possible. The length of the audios varies from 1 sec to 2 sec maximum and the training participants are evenly split 10-9 male-female. ## Data Preprocessing During the data pre-processing stage, the audio data is saved in a stan- dardized 16 kHz 16-bit single channel PCM wave file format, ensuring consistency in the format for subsequent analysis. Additionally, the amplitude of the audio signals is normalized to a standard level, promoting uniformity across different recordings. As part of the data augmentation process, five supplementary samples are generated for each audio recording. These samples introduce varia- tions to the original audio, expanding the dataset’s diversity. One such variation is the creation of a noise sample, achieved by adding Gaussian noise to the original recorded audio. Furthermore, lower and higher pitch samples are derived from both the original audio and the noisy version, enabling the model to learn representations across different pitch ranges and noise levels. Consequently, the train- ing dataset grows to include a total of 1140 samples, while th …