This model is part of the research presented in the paper How much speech data is necessary for ASR in African languages? An evaluation of data scaling in Kinyarwanda and Kikuyu.

The akera/whisper-large-v3-kin-full model is a fine-tuned version of OpenAI's Whisper-large-v3, specifically adapted for Kinyarwanda Automatic Speech Recognition (ASR). It has been trained on approximately 1400 hours of Kinyarwanda speech data.

The paper investigates the development of ASR systems for low-resource African languages, addressing the challenges posed by limited transcribed speech data. It evaluates Whisper's performance through comprehensive experiments on Kinyarwanda and Kikuyu, analyzing data scaling and characterizing primary failure modes. The results demonstrate that practical ASR performance (WER < 13%) can be achieved with as little as 50 hours of training data, with significant improvements up to 200 hours (WER < 10%). The research also highlights that data quality issues, such as noisy transcriptions, are critical for robust system performance.

This model is primarily focused on:

For more details, including training and evaluation code, please refer to the official GitHub repository: SunbirdAI/kinyarwanda-whisper-eval.

To get started with the evaluation environment and run inference with this model, follow the steps below:

To run evaluation on this specific model:

Evaluation on dev_test[:300] subset:

If you find this work helpful or inspiring, please consider citing it: