“Hello Afrika!”:Speech Commands in Kinyarwanda
INTRODUCTION
Speech commands are concise voice recordings, typically a single word or a short phrase used for voice control and activating AI systems through keywords spotting. They are critical for voice interfaces and smart devices controls. Their models are designed to be small, involve less compute, energy-efficient and accurate enough to avoid false triggers.
These commands generally consist of directives {“Start”, “Stop”, “Left”, “Right”, “Play”, “Record”, …}, confirmations {“Yes”, “No”}, numbers {0 – 9} and in some cases brand specific wake words {“Hey Alexa”, “Hey Siri”, “Ok Google”, …}. When a device hibernates, the keyword spotting model built wakes the device with the wake word.
> 2618 speakers
> 105829 utterances
> 3.8 GB audio
> 16-bit Mono-Channel
> WAV file format
> 16kHz Sampling Rate
"Hello Afrika!" is particularly beneficial for people with disabilities, speech command systems aid mobility, visual, and cognitive impairments, enhancing life quality through voice-enabled personal assistants and focuses on their native language (Kinyarwanda).
Kinyarwanda is a considered a High Resource Language but the available data on Mozilla Common Voice was sentence-based recordings that needed further processing which was handled by the Multilingual Spoken Word Corpus.
METHODOLOGY
A. Corpus Design and Data Collection
To achieve our objectives, Houssain et al gave recommendations on Corpus Design for Speech Commands that we applied in this work. Our corpus included 23 positive keyword classes and a wake word ‘Muraho Afrika’. For our data collection process, we adopted three approaches, the first was using the MSWC Kinyarwanda dataset to find our keywords as shown below:
However, the MSWC dataset omitted some of the keywords required in our corpus, to supplement this, we collected our own audio samples from 140+ native speakers in Kigali through the Microsoft Power Platform and used Google Speech Commands for the Zero/Zeru class.
We segmented the cleaning to 5 classes: Wrong_word, Empty, Overlap, To_crop and Incomplete. Due to encoding issues files were converted to OPUS and back to WAV to re-encode with RIFF tag for model processing.
B. Data Cleaning
C. Model Building
Shift, Zero-Padding, PlaySpeed, Amplify
Data Augmentation
Waveform Generation (Librosa)
Spectrogram Generation (Librosa)
Resize of Spectrogram
Sawa
Model Output
Feeding Input to Model for Training
Cleaned Data
A. Model Performance - Pretrained Model
Special thanks to R. Ihabwikuzo and N. Hailemariam for their guidance on this project.
ACKNOWLEDGEMENTS
CONCLUSION
We deployed the model on a LSTM-based model on a Linux PC. The model's performance was evaluated by inputting locally saved WAV files. Additionally, we investigated the utilization of Edge Impulse for deploying the model on both a local machine and an Edge Device, with the Wio Terminal chosen as the preferred Edge Device.
We have successfully curated the inaugural speech command dataset in Kinyarwanda and developed models to showcase its practical application. Our future endeavors involve refining the data, optimizing models, and promoting cultural inclusivity. We aspire to expand this initiative to encompass various African languages, adapting to linguistic diversity and contributing to the broader accessibility of speech command technology across diverse landscapes.
Speech command datasets of popular languages such as English, French and Spanish are available and have been used to finetune models for their native speakers. However, based on our search we found no paper on Speech Commands for African Languages.
RESEARCH PROBLEM
EXPERIMENTAL RESULTS
REFERENCES
Pretrained model
MSWC
Collected
Combined
91.3% 71.8%
61.8% 36.8%
93.1% 78.1%
75.2% 36.7%
George Igwegbe, Martins Awojide, Mboh Bless, Nirel Kadzo
Supervisor: Prof. Ahmed Biyabani