Can I re-train an already deployed custom voice model with newly added data without undergoing the entire training time again (approximately 24 hours)?

Question

Here’s the context: We set up a voice talent, added training data, trained the model, and deployed it. We've now updated the dataset with more audios and transcripts, increasing the number of utterances from 1300 to 1500. When I try to train this voice model again, the only option is to "Train a New Model," which will take another 24 hours and re-train the initial 1300 utterances along with the newly added 200.

Is there a way to train just the additional 200 utterances to improve the existing voice?

Answer 1

Hello PAVAGEAU Perrine ,

Thanks for your question.

Currently, Azure AI Speech does not support incremental training, which means you cannot simply add new data to an existing model without retraining the entire model from scratch. When you update your dataset and retrain, the system will include all previous data along with the newly added data, resulting in the same approximate 24-hour training time. To manage this, consider accumulating larger batches of new data before retraining, ensuring the new data is of high quality, and automating the data preparation and training initiation process using tools like Azure DevOps. After each training iteration, evaluate the model's performance with a validation dataset to understand the impact of the new data.

If you find this helpful, please accept this answer to close the thread. Thanks!