Swahili Speech to text

Finetuning whisper-small for swahili speech to text.

photo credits: Devonyu

Description:

This project uses the small version of Whisper, a general-purpose speech recognition model created by OpenAI, to convert Swahili audio to text. Whisper is pretrained for ASR (Automatic speech Recognition) and speech translation on 680k hours of labelled data

Author

• Github @JM_Rono
• Linked_in @John Michael Rono

Table of Contents

A Data
B Machine learning
C Deploying

Design

A. Data

The data consist of about 82K instances of swahili audio form Mozilla common voice. I got the dataset from participating in a Zindi competition. Join the completed competiotion to get acces to the data.

B. Machine Learning

• Check-out notebook: @notebook

Evaluation

The model had a WER score of 8.365 wandb

C. Deploying

• Deployed at: https://swahilispeechtotext-htxcv6sjkjcazseovpfrkk.streamlit.app/
• Upload a 30 second audio or less

• Hit the text button.